DEV Community: Thurmon Demich

Best GPU for AI Agents in 2026 (5 Picks Ranked)

Thurmon Demich — Sun, 07 Jun 2026 01:14:13 +0000

From the Best GPU for LLM archive. The canonical version has interactive calculators, an up-to-date GPU comparison table, and live pricing.

You're building an AI agent that needs to think fast — maybe it's browsing the web, writing code, or orchestrating multi-step workflows. Every tool call waits on your GPU. Slow inference means slow agents.

Quick answer: The RTX 4090 is the best GPU for local AI agents. Agents need fast inference with moderate VRAM — 24GB handles 13B-34B models at speeds that keep multi-step reasoning under 30 seconds per chain.

Who this is for

You're running autonomous AI agents locally — frameworks like AutoGPT, CrewAI, LangChain agents, or custom tool-calling pipelines. You need a GPU that delivers fast inference because agents make dozens of LLM calls per task.

Why agents need different GPU specs

Unlike single-turn chat, agents make multiple sequential LLM calls per task. A web research agent might:

Plan the search (1 LLM call)
Generate queries (1 call)
Summarize each result (5-10 calls)
Synthesize a final answer (1 call)

That's 8-13 calls per task. If each call takes 5 seconds, the whole thing takes over a minute. With a fast GPU, you cut that to 15-20 seconds.

Factor	Importance for agents
Tokens/sec	Critical — multiplied across many calls
VRAM	Important — 13B+ models reason better
Batch support	Nice — some frameworks parallelize calls

Best GPUs for AI agents

GPU	VRAM	Speed (13B Q4)	Agent chain (10 calls)	Price
RTX 5090	32GB	~55 tok/s	~15 sec	~$2,000
RTX 4090	24GB	~40 tok/s	~20 sec	~$1,600
RTX 5080	16GB	~30 tok/s	~28 sec	~$1,000
RTX 4060 Ti 16GB	16GB	~20 tok/s	~40 sec	~$400

For agent work, model quality matters more than for simple chat. A 13B model reasons better than 7B, and a 34B model handles complex tool-calling more reliably. That pushes you toward 24GB+ VRAM. Check our Ollama guide for model-specific benchmarks and our RAG guide if your agent uses retrieval.

GPU tier list available at the original article

Which GPU should you buy?

Simple 7B agent on a budget? → RTX 4060 Ti 16GB ($400). Works but agent quality suffers with smaller models.
Serious agent development? → RTX 4090 ($1,600). 24GB runs 34B models that reason well.
Production agent system? → RTX 5090 ($2,000). 32GB + fastest inference = shortest agent chains.
Just prototyping? → Whatever you have. Test the framework first, optimize hardware after.

Common mistakes to avoid

Using a 7B model for complex agent tasks. Smaller models fail at multi-step reasoning and tool calling. Agents need at least 13B, preferably 34B.
Optimizing for single-call latency instead of chain latency. A 10% speed improvement multiplied across 10 calls saves meaningful time per task.
Forgetting that agents need context for history. Each step adds to the conversation context. Budget VRAM for 8K+ context, not just the model.

Final verdict

Need	Best pick	Price
Best overall	RTX 4090	~$1,600
Best performance	RTX 5090	~$2,000
Best budget	RTX 4060 Ti 16GB	~$400

Agents multiply your GPU's speed advantage. Every token-per-second improvement compounds across dozens of LLM calls per task.

Related guides on Best GPU for LLM

Read the full guide on Best GPU for LLM — includes our VRAM calculator, GPU comparison table, and live pricing.

How Much VRAM for AI Video Generation in 2026? (Guide)

Thurmon Demich — Sat, 06 Jun 2026 01:14:17 +0000

Cross-posted from Best GPU for AI — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

AI video generation has a VRAM problem. Still-image models like SDXL or Flux can be squeezed into 12-16 GB with quantization tricks. Video models cannot — they must hold multiple frames in VRAM simultaneously, and the memory requirements scale aggressively with resolution and clip length. Here is exactly what each major tool needs.

Quick answer: Stable Video Diffusion works at 8 GB minimum. HunyuanVideo needs 24 GB or more. If you want to run serious local AI video generation, plan for 24 GB minimum.

Why AI video needs so much VRAM

Static image diffusion generates one frame. Video models generate 16-121 frames at once. Each frame is a full image tensor, and the temporal attention layers need to attend across all frames simultaneously. A 5-second clip at 24fps means 120 frames in memory at once — approximately 8-15x the VRAM of a single image at the same resolution.

Additionally, video models use larger base architectures than image models. HunyuanVideo's transformer checkpoint alone is 40+ GB unquantized. Even aggressively quantized, the working VRAM requirement rarely drops below 18-20 GB for full operation.

VRAM requirements by tool

Tool / Model	Min VRAM	Comfortable	Optimal	Notes
Stable Video Diffusion (SVD)	8 GB	12 GB	16 GB	Short clips, 576x1024 max
SVD-XT (25 frames)	10 GB	16 GB	16 GB	Extended clip length
CogVideoX-2B	12 GB	16 GB	16 GB	Open-source, solid quality
CogVideoX-5B	16 GB	24 GB	24 GB	Better quality, needs more VRAM
AnimateDiff + SDXL	12 GB	16 GB	16 GB	Needs optimized workflow
Wan2.1 (14B)	16 GB	24 GB	24+ GB	Strong open-source option
HunyuanVideo	24 GB	40 GB	80 GB	SOTA quality, needs quantization under 40GB
Runway / Kling (local)	Not available locally	—	—	Cloud-only

Stable Video Diffusion — the 8 GB baseline

Stable Video Diffusion generates short clips (14-25 frames) from a single input image. The original SVD model runs at 8 GB with careful settings — you need --medvram mode and frame generation at 576x1024 or lower. The SVD-XT extension pushes to 25 frames and needs 10-12 GB to avoid constant swapping.

SVD is dated by 2026 standards. The output is limited to 3-4 second clips, resolution is capped, and there is no text prompt input. It remains useful as an animation tool (bring a still image to life) but does not produce the kind of AI video that newer models do.

VRAM chart available at the original article

CogVideoX — the 16 GB sweet spot

CogVideoX-5B is a practical open-source video model that runs in 16 GB with INT8 quantization. It generates 6-second clips from text prompts at 720p, with quality that is genuinely useful. The 2B variant runs in 12 GB with better headroom.

For users with a 16 GB card (RTX 4060 Ti 16GB, RTX 5060 Ti, RTX 5070 Ti, RTX 5080), CogVideoX-5B with quantization is the best locally runnable option today. Expect generation times of 10-20 minutes per clip on a 16 GB card — this is slow, but it runs without cloud costs.

Wan2.1 — emerging open-source option

Wan2.1 is a strong contender from 2025-2026. The 14B model produces high-quality video output and runs on 16-24 GB with quantization. At 16 GB (with aggressive quantization), clips are short and generation is slow. At 24 GB, it runs more comfortably.

This is the model most recommended for users with an RTX 4090 who want the best locally-runnable video quality without paying for HunyuanVideo's full requirements.

HunyuanVideo — 24 GB minimum

HunyuanVideo is the state-of-the-art open-source video generation model as of 2026. It produces cinematic-quality 720p video at 3-10 second lengths. The requirements are brutal:

Full fp16: ~40+ GB. RTX 4090 cannot run it.
INT8 quantized: ~24 GB. RTX 4090 can run it, slowly.
INT4 quantized: ~18-20 GB. Fits in RTX 4060 Ti 16GB with aggressive tuning, but quality degrades.

For HunyuanVideo at acceptable quality, the RTX 4090 (24 GB) is the minimum practical card. Expect 20-60 minutes per clip depending on length and settings. The RTX 5090 (32 GB) runs INT8 HunyuanVideo with more headroom and better speed.

GPU recommendations by budget

Budget	GPU	What AI Video runs
$400	RTX 4060 Ti 16GB	SVD, CogVideoX-2B, AnimateDiff
$450-500	RTX 5060 Ti	SVD, CogVideoX-5B (INT8), Wan2.1 (quantized)
$750	RTX 5070 Ti	CogVideoX-5B, Wan2.1, HunyuanVideo (INT4, slow)
$1,000	RTX 5080	Same as 5070 Ti, faster
$1,600+	RTX 4090	HunyuanVideo (INT8), full Wan2.1
$2,000+	RTX 5090	HunyuanVideo at quality settings

Which GPU should YOU buy?

You want to run SVD or CogVideoX-2B: 12-16 GB is enough. The RTX 4060 Ti 16GB at $400 or RTX 5060 Ti at $450 both work.

You want CogVideoX-5B or Wan2.1 at good quality: 16 GB with quantization works, but 24 GB is comfortable. The RTX 4090 hits the sweet spot here.

HunyuanVideo is your target: Do not buy anything with less than 24 GB. The RTX 4090 is the entry point. A used RTX 3090 (24 GB) at lower cost is viable but slower.

You want the absolute best local AI video: RTX 5090 (32 GB). Nothing else comes close for HunyuanVideo at quality settings with reasonable generation times.

Common mistakes to avoid

Buying 16 GB specifically for HunyuanVideo. It technically runs with INT4 quantization, but the quality loss is significant and generation is extremely slow. You will be disappointed.
Ignoring generation time. AI video is slow even on good hardware. A 5-second clip on a 24 GB card can take 20-40 minutes. Budget your expectations accordingly.
Treating AI video like AI images. The same tricks that reduce image model VRAM (tiled decoding, attention slicing) often do not work well for video models, which need the temporal context of the full sequence in memory.

Final verdict

AI video generation is the most demanding local AI workload in 2026. If your goal is running HunyuanVideo locally, 24 GB is the minimum — full stop. For lighter tools like SVD or CogVideoX, 16 GB works with quantization. See our Best GPU for AI Video for full recommendations, Best GPU for HunyuanVideo for that specific model, and How Much VRAM for AI for a broader breakdown across all AI workloads.

Related guides on Best GPU for AI

The full version lives on Best GPU for AI — VRAM calculator, GPU comparison table, and live Amazon pricing.

RTX 5070 Ti vs RTX 3090 for LLM: New $750 vs Used $600

Thurmon Demich — Fri, 05 Jun 2026 01:14:18 +0000

This article was originally published on Best GPU for LLM. The full version with interactive tools, FAQ, and live pricing is on the original site.

Two GPUs, almost the same price, completely different strengths. The RTX 5070 Ti brings 16GB of fast GDDR7 and 5th-gen tensor cores for $750. A used RTX 3090 gives you 24GB of GDDR6X — 50% more VRAM — for around $600. I've tested both for local LLM inference, and the right choice depends entirely on what models you plan to run.

Raw specs comparison

Spec	RTX 5070 Ti	RTX 3090 (used)
VRAM	16GB GDDR7	24GB GDDR6X
Memory bandwidth	~896 GB/s	~936 GB/s
Tensor cores	5th gen	3rd gen
TDP	300W	350W
Price	~$750 new	~$600 used
Warranty	Full manufacturer	None (used)
7B Q4 tok/s	~45	~55
13B Q4 tok/s	~27	~35

The 3090 is actually faster in raw tok/s on these models because its wider 384-bit memory bus and higher effective bandwidth feed tokens quickly. But the 5070 Ti's newer architecture narrows that gap more than the numbers suggest — its tensor cores handle quantized inference more efficiently per watt.

VRAM chart available at the original article

Where the 5070 Ti wins

For 7B and 13B parameter models — which covers Llama 3 8B, Mistral 7B, Phi-4, Qwen 2.5 14B (at Q4), and most coding assistants — 16GB is plenty. You won't bump into VRAM limits, and the 5070 Ti runs cool, draws less power, and comes with a warranty.

The 5070 Ti is the better choice if you:

Run 7B-13B models as your daily driver
Want a new card with manufacturer warranty
Plan to use the GPU for gaming or creative work too
Don't want to deal with used market risks

At 45 tok/s on 7B Q4, the 5070 Ti delivers fast, interactive responses. That's well above the ~30 tok/s threshold where output feels instantaneous for chat use.

Where the 3090 wins

The 3090's 24GB advantage becomes decisive the moment you try to load a 34B model. CodeLlama 34B at Q4_K_M needs ~20GB of VRAM. Qwen 2.5 32B at Q4 needs ~19GB. The 5070 Ti simply cannot fit these models. The 3090 loads them with room to spare.

The 3090 is the better choice if you:

Want to run 30B-34B parameter models locally
Plan to add a second GPU later for 70B inference
Need headroom for larger context windows
Are comfortable buying used hardware

At ~35 tok/s on 13B and ~12-18 tok/s on 34B models, the 3090 handles heavier workloads that the 5070 Ti physically cannot attempt. For a full guide on buying one safely, see Used RTX 3090 Buying Guide.

The model size decision tree

This is how I frame it:

Only running 7B models? Either card works. Save money with an RTX 3060 12GB at $150 used.
Running 7B-13B regularly? 5070 Ti. Newer, faster per watt, and 16GB is sufficient.
Running 34B models? 3090. No alternative at this price. The next 24GB+ option is the RTX 4090 at $1,600. Wondering whether the cheaper non-Ti RTX 5070 might squeeze 34B in at all? See can the RTX 5070 run 34B? for the bad news at 12GB.
Planning multi-GPU later? 3090. Two 3090s give you 48GB combined for ~$1,200, enough for 70B models.

Value per dollar

Metric	RTX 5070 Ti	RTX 3090
Price	$750	$600
VRAM per dollar	21.3 MB/$	40.0 MB/$
7B tok/s per $100	6.0	9.2
13B tok/s per $100	3.6	5.8
Max model size (Q4)	~13B comfortably	~34B comfortably

The 3090 wins on pure value metrics. But value isn't everything — warranty, power efficiency, and noise matter for a daily-use workstation.

My recommendation

If your budget is under $1,000 and you want maximum model flexibility, buy the used RTX 3090. The 24GB VRAM ceiling is simply more future-proof for LLM work. Models keep getting bigger, and VRAM is the one spec you can't work around.

If you want a clean, new-card experience and only run 7B-13B models, the RTX 5070 Ti is the smarter pick. You get warranty coverage, lower power draw, and enough VRAM for the most popular open-weight models.

For more options in this price range, see the full best GPU for LLM under $1,000 roundup.

Related guides on Best GPU for LLM

The full version lives on Best GPU for LLM — VRAM calculator, GPU comparison table, and live Amazon pricing.

Best GPU for ControlNet in 2026: 5 Cards (16GB Sweet Spot)

Thurmon Demich — Thu, 04 Jun 2026 01:13:59 +0000

Cross-posted from Best GPU for AI — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

Quick answer: The RTX 4070 Ti Super (16GB) is the best GPU for ControlNet in 2026. It absorbs SDXL plus a 3-stack of preprocessors (Canny + Depth + OpenPose) and an IP-Adapter without spilling into system RAM, and it costs roughly half what a 4090 does.

Who this is for

This guide is for anyone running ControlNet on top of Stable Diffusion locally — whether you're posing characters with OpenPose, fixing hands with depth maps, or chaining IP-Adapter for style transfer. If you're already on a 12GB card and watching ComfyUI swap to disk every other render, you're the target reader. We assume SDXL or SD 1.5 as the base model; Flux + ControlNet is a different (heavier) beast that we flag where it matters.

How ControlNet VRAM actually adds up

ControlNet doesn't replace your base model — it sits next to it. Every preprocessor (Canny, Depth, OpenPose, Soft Edge, etc.) loads its own conditioning model into VRAM alongside the SDXL checkpoint. Stack two or three and the math gets ugly fast.

Here's the realistic accounting we see in production workflows at 1024×1024:

Workload	Base SDXL	+ ControlNets	+ IP-Adapter	Total VRAM
SDXL alone	~10GB	—	—	~10GB
SDXL + 1 ControlNet (Canny)	~10GB	+1.5GB	—	~11.5GB
SDXL + 2 ControlNets (Canny + Depth)	~10GB	+3GB	—	~13GB
SDXL + 3 ControlNets (Canny + Depth + OpenPose)	~10GB	+4.5GB	—	~14.5GB
SDXL + 3 ControlNets + IP-Adapter	~10GB	+4.5GB	+2GB	~16.5GB
Same stack + 4× upscaler in same workflow	~10GB	+4.5GB	+2GB	~19–20GB

The line where 12GB cards die sits between the second and third row. A 3-stack with IP-Adapter clears 16GB easily — which is why 12GB and even some 16GB cards start swapping mid-generation. Activations during the actual denoise add another ~1–2GB of headroom you don't get to spend on models.

VRAM chart available at the original article

GPU ranking for ControlNet workloads

Generation times below are for 1024×1024 SDXL at 25 steps, measured in our test workflows. The "3-stack" column is Canny + Depth + OpenPose running simultaneously with an IP-Adapter active — the realistic creative-control case, not a synthetic best-case.

GPU	VRAM	SDXL solo	+ 1 ControlNet	+ 3-stack + IP-Adapter	Price
RTX 5090	32GB	~2s	~2.5s	~3.5s	~$2,000
RTX 4090	24GB	~3s	~3.5s	~5s	~$1,600
RTX 5080	16GB	~3.5s	~4s	~5.5s	~$1,000
RTX 5070 Ti	16GB	~4.5s	~5s	~6.5s	~$750
RTX 4070 Ti Super	16GB	~6s	~7s	~8.5s	~$700
RTX 4060 Ti 16GB	16GB	~9s	~11s	~14s	~$400
RTX 3090 (used)	24GB	~7s	~8s	~10s	~$700
RTX 3060 12GB	12GB	~14s	~18s	OOM / swap	~$200

Two things jump out. First, the 3060 12GB simply doesn't finish the 3-stack workflow without offloading to system RAM, which pushes per-image time into the minutes. Second, the 4060 Ti 16GB clears the same workflow that breaks the 3060 — but at less than half the speed of the 4070 Ti Super, because its 288 GB/s memory bandwidth chokes when ControlNet conditioning models hammer VRAM each step.

The used 3090 is a real sleeper here. It's slower than the 5070 Ti per image but its 24GB headroom means you can keep IP-Adapter, multiple ControlNets, and an upscaler all hot in VRAM without juggling node unloads.

Which GPU should YOU buy?

You only run SD 1.5 with one ControlNet at a time: A used RTX 3060 12GB at ~$200 is enough. Don't overspend.
You run SDXL with single-ControlNet workflows (just pose, or just depth): The RTX 4060 Ti 16GB at ~$400 is the cheapest survivable option. Slow, but it won't crash.
You stack 2–3 ControlNets with IP-Adapter and want speed: The RTX 4070 Ti Super at ~$700 is our pick. This is the sweet spot of the entire 2026 lineup for ControlNet.
You routinely chain ControlNet + IP-Adapter + AnimateDiff or train LoRA adapters with Kohya_ss: Go to 24GB. RTX 4090 new or RTX 3090 used.
You're doing AI research at multi-model scale — custom ControlNet training, large-batch ablations, or experimental architectures: 24GB is the floor and 32GB is comfortable. The RTX 5090 makes sense if you're iterating on novel pipelines. Our GPU picks for AI research covers the workstation-class context.

If you're new to the broader image-gen stack and unsure where ControlNet even fits, our best GPU for Stable Diffusion guide covers the base-model VRAM picture. And for the full node-graph workflow that ControlNet typically lives inside, see our best GPU for ComfyUI breakdown — ComfyUI is where most serious ControlNet work happens in 2026.

Contrarian take: we recommend against 12GB cards for ControlNet

The 3060 12GB is famously the budget AI darling, and it deserves that reputation for plain SD 1.5 and SDXL solo. For ControlNet in 2026, though, we think 12GB is a trap. The whole point of ControlNet is composability — one preprocessor solves pose, another solves depth, IP-Adapter handles style. The moment you start stacking (which you will, within a week of installing it), 12GB triggers offloading and your iteration loop dies.

We've watched users hold onto 12GB cards "until it really hurts" and the answer is always the same: it already hurts, they just normalized it. Spend the extra $200 on a 4060 Ti 16GB if the budget is tight.

Common mistakes with ControlNet hardware

Running 8GB or 10GB cards with a 3-stack. Anything below 12GB doesn't even start an SDXL + 3-ControlNet workflow without aggressive CPU offloading. You'll see "out of memory" before the first step finishes, or per-image times measured in minutes.
Assuming all 16GB cards are equal. The 4060 Ti 16GB and the 4070 Ti Super both have 16GB, but the 4070 Ti Super has roughly 2.3× the memory bandwidth. ControlNet preprocessors are bandwidth-hungry because they're sampled every denoising step. In our experience, the 4060 Ti runs the same workflow but takes ~60% longer per image.
Forgetting IP-Adapter overhead. IP-Adapter quietly eats ~2GB on top of your ControlNet stack. People plan their VRAM budget for the ControlNets, then add IP-Adapter and wonder why the workflow OOMs. Always count IP-Adapter as if it were a fourth ControlNet.
Leaving preprocessors loaded after generating the control map. This is a free 1.5–2GB win on 16GB cards. In ComfyUI, drop in an "unload model" node after Canny/Depth/Pose generates its conditioning image. You only need the preprocessor once per generation, not every step.

Final verdict

Budget	GPU	Best for in ControlNet
~$200 used	RTX 3060 12GB	SD 1.5 + single ControlNet only
~$400	RTX 4060 Ti 16GB	Full SDXL stacks, slowly
~$700	RTX 4070 Ti Super	3-stack + IP-Adapter, sweet spot
~$700 used	RTX 3090 24GB	Heavy stacks with VRAM headroom
~$1,600	RTX 4090	Multi-stack + training, no compromises
~$2,000	RTX 5090	32GB, research-scale workflows

The best GPU for ControlNet is the one that keeps every preprocessor, conditioning model, and IP-Adapter resident in VRAM at the same time — the moment you spill, your iteration loop is dead.

Related guides on Best GPU for AI

The full version lives on Best GPU for AI — VRAM calculator, GPU comparison table, and live Amazon pricing.

Mac vs NVIDIA GPU for Local LLM: Which Platform Wins?

Thurmon Demich — Wed, 03 Jun 2026 01:14:09 +0000

Cross-posted from Best GPU for LLM — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

Neither platform dominates completely. Apple Silicon Macs win when you need large model inference with unified memory and near-zero noise. NVIDIA wins when you need raw speed, CUDA tooling, or fine-tuning. The right answer depends on what you are actually doing. For a specific answer on whether the most affordable Apple device is enough, see our can the Mac Mini run LLMs guide.

The fundamental difference: unified memory vs dedicated VRAM

This is the core tradeoff:

Apple Silicon uses unified memory — the same pool serves both CPU and GPU. An M4 Max with 128GB of unified memory can address all 128GB for model weights. This means running a 70B model quantized to Q4_K_M (~38GB) on a machine with 64GB RAM is feasible.

NVIDIA uses dedicated VRAM on the GPU die, physically separate from system RAM. An RTX 4090 has 24GB, period. Loading anything larger requires multi-GPU setups or offloading — which tanks inference speed.

Attribute	Apple Silicon	NVIDIA GPU
Max single-device memory	128GB (M4 Max)	32GB (RTX 5090)
Memory bandwidth	~500-800 GB/s (M4 Max)	1,008-1,792 GB/s
Architecture	Unified (CPU+GPU share)	Dedicated VRAM
Best for	Large model inference	Fast inference, training
OS support	macOS only	Linux, Windows, macOS
CUDA support	No	Yes

Inference speed comparison

Running Llama 3 8B at Q4_K_M:

Device	~Tok/s	Memory	Price
RTX 5090 (32GB)	~105 tok/s	32GB dedicated	~$2,000 (GPU only)
RTX 4090 (24GB)	~65 tok/s	24GB dedicated	~$1,600 (GPU only)
M4 Max (128GB)	~38 tok/s	128GB unified	~$4,000 (full system)
M4 Pro (48GB)	~28 tok/s	48GB unified	~$2,500 (full system)
RTX 4060 Ti 16GB	~28 tok/s	16GB dedicated	~$400 (GPU only)
M4 (24GB)	~22 tok/s	24GB unified	~$1,600 (full system)

NVIDIA is faster per token at the same memory size. The RTX 4090 generates ~65 tok/s versus the M4 Max's ~38 tok/s. However, comparing a $1,600 GPU to a $4,000 Mac system is not the same trade.

Large model inference: where Mac genuinely wins

Load Llama 3 70B at Q4_K_M (~38GB). Your options:

M4 Max 64GB Mac: runs it at ~8-12 tok/s. Slow but functional, fully on-device.
M4 Max 128GB Mac: runs it comfortably at ~12-15 tok/s with full context headroom.
RTX 4090 alone: cannot fit it. 38GB model, 24GB card.
RTX 5090 alone: ~32GB card, barely fits at Q3_K_M (degraded quality), no headroom.
2x RTX 4090: fits at Q4_K_M, ~25 tok/s, costs $3,200 in GPUs plus a compatible motherboard.

For 70B+ model inference on a budget you control, the M4 Max Mac is genuinely competitive. You pay more upfront, but it is a complete system that just works.

Software ecosystem: NVIDIA's real advantage

CUDA is the bedrock of the LLM software stack:

Tool	NVIDIA (CUDA)	Apple (Metal/MPS)
Ollama	Native, fast	Supported
llama.cpp	cuBLAS backend	Metal backend
vLLM	Full support	Not supported
ExLlamaV2	Full support	Not supported
Fine-tuning (LoRA)	Full support	Limited/slow
PyTorch training	First-class	MPS backend, gaps
GPTQ / AWQ quants	Full support	Limited

Mac runs Ollama and llama.cpp well. Anything beyond basic inference — production serving with vLLM, fine-tuning with LoRA, or advanced quantization formats — requires NVIDIA.

Which use case fits which platform?

Mac wins for:

Running 30B-70B models on a single device
Quiet, integrated, always-on personal assistant setups
Privacy-first inference with no separate GPU box
Users who already work in macOS and want zero friction

NVIDIA wins for:

Fastest token throughput on 7B-14B models
Fine-tuning and LoRA training workflows
Production LLM serving with vLLM
Advanced quantization formats (GPTQ, AWQ, EXL2)
Linux-first or Windows-first environments

Which platform should YOU choose?

You want to run 7B-14B models fast and cheap? NVIDIA RTX 4060 Ti 16GB ($400). Faster than a Mac Mini for inference, far cheaper as a GPU add-on to an existing machine.
You want to run 34B-70B models without multi-GPU complexity? M4 Max Mac (64GB or 128GB). The unified memory advantage is decisive at this model tier.
You do fine-tuning or LoRA training? NVIDIA, full stop. Mac's MPS backend for training is functional but significantly slower and missing key optimizations.
You want an all-in-one quiet personal AI machine? Mac. The integrated experience with no extra boxes or power draw is unmatched.
You want maximum inference speed per dollar? NVIDIA. A $400 RTX 4060 Ti outperforms most Macs on 7B-14B inference.

Common mistakes to avoid

Comparing GPU price to Mac system price. An RTX 4090 at $1,600 needs a full PC to run in. A Mac at $2,500 is a complete computer. Factor in the total system cost, not just the GPU.
Assuming Apple Silicon is slow for LLMs. Modern M4 chips have excellent memory bandwidth. They are slower than NVIDIA for small models but competitive for large-model inference where VRAM limits NVIDIA.
Buying a Mac expecting CUDA compatibility. Rosetta does not translate CUDA. vLLM, ExLlamaV2, and many training frameworks simply will not run on macOS. Check your toolchain before buying.
Ignoring Ollama on Mac. Ollama's Metal backend on Apple Silicon is polished and reliable. For casual local inference, the Mac experience is genuinely good.

Final verdict

Goal	Platform	Estimated cost
Fast 7B-14B inference	NVIDIA RTX 5060 Ti	~$450 (GPU only)
Best all-round inference	NVIDIA RTX 4090	~$1,600 (GPU only)
34B-70B on one device	M4 Max Mac (64GB)	~$3,500 (full system)
Fine-tuning / training	NVIDIA RTX 4090	~$1,600 (GPU only)
Quiet all-in-one LLM box	Mac (any M4)	~$1,600+

For Ollama-specific GPU advice on NVIDIA, see best GPU for Ollama. Need a VRAM reference for your target model size? See how much VRAM for local LLM. Comparing NVIDIA to AMD instead? See NVIDIA vs AMD for LLM. If you prefer LM Studio's graphical interface over Ollama, see our best GPU for LM Studio guide for hardware picks tuned to that tool.

Pick Mac if unified memory solves a size problem you cannot solve with affordable NVIDIA hardware. Pick NVIDIA if speed and the CUDA ecosystem matter more than model size headroom.

Related guides on Best GPU for LLM

The full version lives on Best GPU for LLM — VRAM calculator, GPU comparison table, and live Amazon pricing.

Best GPU for Wan 2.2 in 2026: 5 Picks Ranked (14B Ready)

Thurmon Demich — Tue, 02 Jun 2026 01:13:58 +0000

Cross-posted from Best GPU for AI — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

Alibaba dropped Wan 2.2 14B under Apache 2.0 this May, and ComfyUI integration landed within days. The early numbers are interesting — Wan 2.2 is lighter than HunyuanVideo at comparable resolutions, which finally puts open-source video generation in reach of 16GB cards (with caveats). The Apache 2.0 license also means commercial use is fine, so this isn't just a tech demo. It's the model people are actually shipping social content with right now.

Quick answer: The RTX 4090 24GB is the best local pick for Wan 2.2 14B at FP16, while the RTX 4070 Ti Super 16GB handles int8/Q4 quants comfortably. If you want headroom for 5-second 720p clips at full precision without offload tricks, go RTX 5090.

Who this is for

You want to run Wan 2.2 locally — image-to-video, text-to-video, or short loops for social content — and you're tired of cloud queues and per-minute billing. This guide assumes you have ComfyUI installed (or are willing to), and want hard VRAM numbers instead of marketing-speak. If you're brand new to AI video, start with our best GPU for AI video primer first, then come back here once you've narrowed it to Wan.

VRAM requirements for Wan 2.2

Wan 2.2 ships in two flavors right now: the headline 14B model and a leaner 5B variant. The 5B is dramatically more forgiving on VRAM, but quality drops are noticeable — texture stability and prompt adherence both take a hit.

Variant + Quant	Minimum VRAM	Comfortable VRAM	Resolution Tested
Wan 2.2 14B FP16	24GB	28-32GB	720p, 5-sec
Wan 2.2 14B int8	16GB	20GB	720p, 5-sec
Wan 2.2 14B Q4	12GB	16GB	480p, 3-sec
Wan 2.2 5B FP16	12GB	16GB	720p, 5-sec
Wan 2.2 5B int8	8GB	12GB	480p, 5-sec

In our experience running both, the 14B at int8 on a 16GB card is the practical floor for results you'd actually post. Q4 on 12GB works, but artifacts creep in fast on motion-heavy scenes — faces deform mid-motion, text in backgrounds smears, and complex camera movements fall apart. A second factor most VRAM tables ignore: context length. Wan 2.2 scales VRAM roughly linearly with clip duration. A 5-second clip is the tested baseline; pushing to 8-10 seconds adds 30-50% more VRAM at the same resolution.

VRAM chart available at the original article

Generation times per GPU (Wan 2.2 14B)

Numbers below are community-aggregated estimates from ComfyUI Wan 2.2 workflows in early May 2026. Treat them as ballparks — sampler choice, step count, and motion complexity move these meaningfully.

GPU	VRAM	5-sec 480p (int8)	5-sec 720p (int8)	5-sec 720p (FP16)
RTX 5090	32GB	~2 min	~5 min	~7 min
RTX 4090	24GB	~3 min	~8 min	~12 min (tight)
RTX 3090	24GB	~5 min	~13 min	~19 min
RTX 5080	16GB	~4 min	~10 min	offload only
RTX 5070 Ti	16GB	~5 min	~12 min	offload only
RTX 4070 Ti Super	16GB	~6 min	~14 min	offload only
RTX 4060 Ti 16GB	16GB	~10 min	~24 min	not recommended
RTX 3060 12GB	12GB	~14 min (Q4)	not recommended	no

The 4060 Ti 16GB row is the most surprising — it technically fits Wan 2.2 14B int8, but the 128-bit memory bus chokes throughput. Bandwidth matters more than raw VRAM once you clear the minimum.

Best GPU picks for Wan 2.2

RTX 5090 — fastest local option

32GB of GDDR7 and ~1.8TB/s of memory bandwidth is overkill for Wan 2.2 5B and exactly right for 14B at FP16. If you're doing batch generation or planning to train LoRAs on Wan checkpoints, the headroom matters. At ~$2,000 it's not cheap, but it's the only consumer card that runs the 14B comfortably at FP16 without quantization compromises. We recommend against this card for anyone generating fewer than 20 clips a week — the marginal speed gain over a 4090 isn't worth $400 unless you're rendering daily.

RTX 4090 — best value for serious local use

24GB hits the Wan 2.2 14B FP16 minimum and breezes through int8. Generation times are roughly 1.7x slower than the 5090, which is the right trade for a $400 price gap. This is our default recommendation for anyone generating video weekly. We also flag it across our ComfyUI buying guide for the same reasons — it's the card that ages best across video, image, and LoRA training.

RTX 4070 Ti Super — the 16GB sweet spot

At ~$700, this is the cheapest card we'd buy specifically for Wan 2.2. 16GB handles 14B int8 cleanly at 720p, and the memory bandwidth (672 GB/s) keeps generation times sane. The RTX 5070 Ti is roughly equivalent at ~$750 — pick whichever is in stock.

RTX 3090 — used market value play

If you can find one at ~$700 used and trust the seller, 24GB GDDR6X runs Wan 2.2 14B at FP16. Generation is meaningfully slower than the 4090 (memory bandwidth gap is real — 936 GB/s vs 1008 GB/s plus the architectural gap), but the dollar-per-VRAM math is hard to beat for hobby use. Watch out for ex-mining cards: high hours don't kill 3090s outright, but thermal pads on the backside VRAM degrade and need replacing. Budget another $30 and an afternoon for that if you go used.

Don't bother: RTX 4060 Ti 16GB and below

This is the contrarian take. On paper, the 4060 Ti 16GB looks like a Wan 2.2 bargain. In practice, the 128-bit bus turns generation into a slideshow — a 720p int8 clip that takes 8 minutes on a 4090 takes 24 minutes here. Spend the extra $300 on a 4070 Ti Super and stop suffering. The RTX 3060 12GB is fine for Wan 2.2 5B at low resolution, but the 14B model isn't really its job.

Which GPU should YOU buy?

Generating Wan 2.2 daily, want max headroom? RTX 5090 32GB — FP16, no quantization, room for batches.
Weekly Wan 2.2 user, want best dollar-per-frame? RTX 4090 24GB — runs everything Wan ships at FP16 or int8.
First serious AI video build, $700 budget? RTX 4070 Ti Super 16GB — int8 14B is your sweet spot.
Already have a 3090? You're fine. Don't upgrade unless you're doing this professionally.
Stuck with 12GB? Use Wan 2.2 5B or Q4 14B at 480p — and consider cloud for anything you'd publish.
Training Wan LoRAs or fine-tuning? Skip consumer cards. See our AI research GPU guide — for research workloads on diffusion video models you want 48GB+ (RTX 6000 Ada or rented H100s), full stop.

Cloud is still reasonable for Wan 2.2

Wan 2.2 14B at FP16 on an A100 or H100 takes 3-5 minutes per 5-second 720p clip. At RunPod's spot pricing, that's roughly $0.20-0.40 per finished clip. If you generate fewer than ~50 clips a month, you'll never break even on a 4090 purchase. Power draw also matters — a 4090 pulls 350-450W during video generation, which adds up on a metered electric bill. Cloud GPUs externalize that cost. The other underrated cloud benefit: H100s with NVLink let you run Wan 2.2 14B at FP16 with 10-second clips, which no consumer card can match yet.

For high-volume daily use, local pays back inside 6 months. For experimentation, rent.

Common mistakes to avoid

Treating Wan 2.2 like HunyuanVideo. Wan is lighter — you don't always need 24GB. See our HunyuanVideo GPU breakdown for the contrast. Picking your GPU by the wrong model's requirements costs you ~$400-600.
Skipping int8 quantization on a 16GB card. ComfyUI's GGUF nodes for Wan 2.2 are stable as of May 2026. Use them. FP16 on 16GB will OOM on anything past 3 seconds.
Buying a 4060 Ti 16GB for the VRAM number. Memory bandwidth is the real bottleneck for video diffusion. The 4070 Ti Super is night-and-day faster despite the same VRAM.
Ignoring storage. Wan 2.2 14B FP16 weights are ~28GB on disk. Add ComfyUI, LoRAs, and output clips and you'll fill a 1TB SSD inside a month of active use.

Final verdict

Use case	Recommendation	Why
Maximum local performance	RTX 5090 32GB	FP16 14B with batch headroom
Best value for serious users	RTX 4090 24GB	FP16 14B at ~$400 less than 5090
Budget local entry	RTX 4070 Ti Super 16GB	int8 14B sweet spot at ~$700
Used market play	RTX 3090 24GB	FP16 14B if priced right
Hobby / 5B only	RTX 3060 12GB	Wan 2.2 5B at 480p, that's the ceiling
Occasional use	Cloud (RunPod / Vast.ai)	Cheaper than hardware under ~50 clips/month

For Wan 2.2 specifically, the RTX 4090 at 24GB is the GPU we'd buy with our own money today — fast enough at FP16, comfortable at int8, and priced where the math still works for serious local users without forcing a $2,000 commitment to the 5090 tier.

Related guides on Best GPU for AI

Read the full guide on Best GPU for AI — includes our VRAM calculator, GPU comparison table, and live pricing.

Best GPU for Microsoft Phi-4 in 2026 (5 Picks Ranked)

Thurmon Demich — Mon, 01 Jun 2026 01:14:26 +0000

Cross-posted from Best GPU for LLM — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

Phi-4 is one of the most hardware-friendly capable models available. Microsoft's 14B parameter design punches well above its weight in reasoning benchmarks while staying lean enough to run on budget hardware. This is genuinely one of the few situations where a $250 used GPU gets you excellent local inference.

Quick answer

Phi-4 14B at Q4_K_M is ~8.5GB. Any GPU with 8GB+ VRAM can run it — but 12GB+ is the comfortable sweet spot. Budget picks shine here more than almost any other capable model.

Phi-4 VRAM requirements

Quantization	Model Size	Min VRAM	Notes
FP16	~28GB	32GB	Only RTX 5090
Q8	~14GB	16GB	RTX 4060 Ti 16GB
Q6_K	~11GB	12GB	RTX 3060 12GB (tight)
Q5_K_M	~9.7GB	11GB	12GB card ideal
Q4_K_M	~8.5GB	9.5GB	8GB possible, 12GB comfortable
Q3_K_M	~6.5GB	8GB	8GB card fits

This is unusually accessible for a 14B model. Phi-4's architectural efficiency means you get near-13B quality at memory requirements closer to 7B.

The budget case for Phi-4

Phi-4's slim VRAM footprint means affordable GPUs that would struggle with Llama 2 13B can handle Phi-4 14B without issue:

RTX 3060 12GB (~$250 used): Runs Q4_K_M at ~22 tok/s. Comfortable for daily use.
RTX 4060 8GB (~$280): Runs Q4_K_M at ~28 tok/s with tight memory. No context headroom.
RTX 4060 Ti 16GB (~$400): Q4_K_M at ~35 tok/s with plenty of headroom. The smart buy.

If Phi-4 is your primary model and you want the best value for running it, a used RTX 3060 12GB is hard to argue against.

GPU tier list available at the original article

Performance benchmarks

Tested with Ollama at Q4_K_M:

GPU	Phi-4 14B tok/s	Price	Value score
RTX 5090 (32GB)	~85 tok/s	~$2,000	Poor for this model
RTX 4090 (24GB)	~55 tok/s	~$1,600	Overkill
RTX 4060 Ti 16GB	~35 tok/s	~$400	Excellent
RTX 4060 (8GB)	~28 tok/s	~$280	Good
RTX 3060 12GB (used)	~22 tok/s	~$250	Best value
Arc B580 (12GB)	~18 tok/s	~$250	Decent (Intel)

No reason to buy a flagship card for Phi-4 unless you also want to run larger models. The 3060 12GB and 4060 deliver perfectly usable performance at a fraction of the cost.

Which GPU should YOU buy?

RTX 3060 12GB (used) (~$250) — The best pure value pick for Phi-4. Runs Q4–Q5 comfortably. If Phi-4 is your only target model and budget is tight, this is the answer.

RTX 4060 (~$280) — New card, slightly less VRAM than the 3060 12GB but faster bandwidth. Better if you want a new card with a warranty and primarily run Phi-4 or smaller models.

RTX 4060 Ti 16GB (~$400) — The smart future-proof buy. Phi-4 is effortless on this card, and the extra VRAM means you can later move up to 13B models from Llama, Mistral, or Gemma 3 12B without needing a new GPU.

RTX 4090 or above — Only worth it if Phi-4 is one of several models you plan to run, including 34B variants. Purely for Phi-4, it is significant overkill.

Why Phi-4 is special for budget builds

Most capable AI models in the 13–14B range need at least 12–16GB VRAM for comfortable inference. Phi-4 is the exception. Microsoft's training approach compresses reasoning capability into a leaner architecture, which means:

The 8GB RTX 4060 can load it (other 14B models cannot)
You get sub-$300 access to GPT-3.5 class reasoning
Even slow inference (~20 tok/s) is tolerable for non-real-time tasks

For home labs, privacy-focused deployments, and low-power inference servers, Phi-4 combined with a budget GPU is one of the most compelling setups available in 2026.

Common mistakes to avoid

Buying 8GB expecting long context. Phi-4 fits at Q4 in 8GB, but there is essentially zero room for KV cache. You will hit memory errors with longer inputs. 12GB is the minimum for comfortable context lengths.
Spending $1,600 on a 4090 just for Phi-4. Unless you plan to run multiple larger models, a 4090 delivers perhaps 2x the tok/s for 6x the cost. The efficiency math does not work.
Dismissing Phi-4 as "too small." Phi-4 14B matches or beats some 30B+ models on specific reasoning and math benchmarks. Small parameter count does not mean weak performance.
Running Q3 when Q4 fits. On a 12GB card, Q4_K_M fits with room. No reason to run Q3 and accept worse output quality.

Final verdict

Your goal	Best GPU	Price
Phi-4 on a tight budget	RTX 3060 12GB (used)	~$250
New card for Phi-4	RTX 4060	~$280
Future-proof + Phi-4	RTX 4060 Ti 16GB	~$400
Phi-4 + larger models	RTX 4090	~$1,600

Phi-4 democratizes capable local inference. The 3060 12GB and the 4060 are the right picks if Phi-4 is your target — save the premium GPU budget for when you genuinely need 34B models.

For the older Phi generation, see our best GPU for Phi-3 guide. Our best budget GPU for local LLM covers the full sub-$400 market. If you are running 7B models, see our dedicated best GPU for 7B models picks.

Related guides on Best GPU for LLM

Read the full guide on Best GPU for LLM — includes our VRAM calculator, GPU comparison table, and live pricing.

RTX 5090 for AI in 2026: 6-Month Honest Retrospective

Thurmon Demich — Sun, 31 May 2026 01:14:12 +0000

This article was originally published on Best GPU for AI. The full version with interactive tools, FAQ, and live pricing is on the original site.

Quick answer: The RTX 5090 earned its keep for VRAM-bound work — Llama 70B at Q4, Flux.2 in FP16, beefier LoRA batches. But for image generation and most hobbyist workflows, the $400 premium over a 4090 wasn't worth it. Honestly, six months in, we recommend the 5090 only if you're VRAM-limited.

Who this is for

This is for the person sitting on a perfectly good RTX 4090 wondering if the 5090 upgrade is worth $2,000. Or the buyer choosing between them for a fresh AI rig. We've run both daily since January 2026 — SDXL, Flux, Llama 70B inference, LoRA training, some PyTorch research code — and the answer is more nuanced than the launch-week reviews implied.

What actually improved (the wins)

Let's start with what the 5090 genuinely changed.

32GB GDDR7 is the headline, and it earned it. The 4090's 24GB is a real ceiling. We hit it constantly: Llama 70B at Q4_K_M needs ~40GB but barely squeezes onto a 5090 with offloading at usable speeds — on a 4090 you're stuck at Q3 or splitting layers to CPU and watching tok/s collapse. Flux.2 at FP16 (full precision, not the gimped FP8 version) wants ~28GB. SDXL LoRA training with batch size 4 and full text encoder unfrozen? 4090 OOMs, 5090 fits with headroom. This is real, not marketing.

FP8 training is finally usable on consumer hardware. Blackwell's native FP8 tensor cores aren't just a checkbox — we've trained LoRAs in FP8 with measurable VRAM savings and ~1.7x throughput vs BF16 on the same card. The 4090 can do FP8 via software (Transformer Engine emulation) but it's clunky and the speedup evaporates. If you're serious about training, this matters. We've covered this trade-off in more depth in our best GPU for PyTorch guide, where FP8 native support genuinely shifts the recommendation.

Memory bandwidth at 1,792 GB/s vs 1,008 GB/s is a real LLM inference boost. Llama 70B Q4 went from "barely usable" (~12-15 tok/s with offloading on a 4090) to "actually fine" (~35-40 tok/s on a 5090, no offloading). For a 13B model, 5090 hits ~140 tok/s vs 95 on the 4090. That 35-46% spread on LLM tok/s is consistent across our RTX 4090 vs 5090 head-to-head testing and matches the Knightli benchmarks the community has been citing.

For AI research workloads, the 32GB unlocks experiments you couldn't run before. Bigger context windows, full-fp16 attention on longer sequences, MoE models with more experts loaded — the kind of work we cover in our best GPU for AI research deep-dive simply wasn't possible on a 4090 without dropping to quantization or splitting across cards. If you're doing actual research workloads (not just running pretrained models), the 5090's VRAM is a structural advantage.

What didn't change as much as advertised

Now the contrarian half — and this is where the launch-week articles oversold the card.

Image generation only improved 20-25%. SDXL went from 6.5s/img to ~4.0s/img. Flux dev from 18s to ~14s. That's nice, but not life-changing. If you're cranking out images, you'll notice. If you generate a few per session, you genuinely will not feel $400. Look — if you only do image gen, save your $400 and buy a 4090.

Gaming-grade BF16 isn't 2x faster. Despite the spec sheet showing ~130 TFLOPS FP16 vs 82.6 on the 4090 (a ~57% theoretical jump), real-world BF16 training throughput sits closer to +35-45% in our runs. The Blackwell scheduler and driver stack are still maturing — we've seen kernels regress between driver versions. Honestly, the 5090 disappointed us here. We expected 1.7-2x, got 1.4x.

The $400+ premium hurts when you account for PSU and case. 575W TGP vs 450W means a lot of 850W PSUs that ran a 4090 fine are now marginal. Add ~$150 for a quality 1000W unit, factor in transient spikes that have crashed some 1000W units (real reports, not theory), and you're looking at $500-600 total upgrade cost over a 4090. For 25% faster image gen.

For the majority of AI hobbyists, the 4090 is still the right answer. It's not slower in any way that breaks workflows — it's slower in ways that add seconds, not minutes. We still recommend it as the default in our best GPU for AI cluster guide, and after 6 months with both cards we're not walking that back.

Workload-by-workload verdict

Here's how we'd actually advise per workflow, based on six months of daily use:

Workload	Recommendation	Why
Image gen (SDXL, Flux dev)	RTX 4090	20-25% speedup doesn't justify $400. Both fit the models comfortably.
Image gen (Flux.2 FP16, video models)	RTX 5090	24GB OOMs on Flux.2 full precision. 32GB is the fix.
LLM inference (≤13B)	RTX 4090	Both run flat-out. 95 vs 140 tok/s is a "nice to have."
LLM inference (34B-70B)	RTX 5090	This is where 32GB earns the upgrade. Q4 70B is usable, not painful.
LoRA fine-tuning (small models)	RTX 4090	Workflows fit. Speedup exists but isn't transformative.
Full fine-tuning / FP8 training	RTX 5090	Native FP8 + bandwidth + VRAM all compound here. Real win.
AI research (long context, MoE, novel architectures)	RTX 5090	The 8GB extra unlocks experiments. Not optional.
Mixed hobbyist (some of everything)	RTX 4090	Honest answer. Most workflows aren't VRAM-limited.

Common mistakes we've watched people make

After six months of watching the 4090 → 5090 upgrade discourse, four mistakes keep coming up.

1. Buying the 5090 just for image generation. If your SDXL/Flux workflow runs fine on a 4090, you are buying 25% speed for $400. Save the money or put it toward a better monitor.

2. Underestimating the PSU and thermal upgrade. That 850W gold-rated PSU you bought for your 4090 build is marginal at 575W card TGP plus a modern CPU. Plan for $150-200 of platform upgrades, not just the GPU swap.

3. Assuming "newer architecture" means "always faster." Blackwell drivers were rough through Q1 2026. We saw PyTorch training kernels actually slower than Ada Lovelace on certain ops until the April driver. If you need rock-stable today, the 4090's mature software stack is genuinely an asset.

4. Buying the 5090 because the 5080 is bad. The RTX 5080 at 16GB is a real letdown for AI — too little VRAM for the price. Don't let that push you up to the 5090 unconsidered. The 4090 is the actual sweet spot in that lineup, not the 5090 by default.

Final verdict

	RTX 4090	RTX 5090
VRAM	24GB GDDR6X	32GB GDDR7
Bandwidth	1,008 GB/s	1,792 GB/s
Image gen	6.5s/img SDXL	4.0s/img SDXL
70B inference	Painful	Actually usable
FP8 training	Software emulation	Native, ~1.7x faster
TGP	450W	575W
Street price	~$1,600	~$2,000
Our verdict	Default pick for most	Worth it only if VRAM-bound

One-sentence verdict: The RTX 5090 is a real upgrade for VRAM-bound and FP8-training workflows — but for the average AI hobbyist running SDXL and 13B models, the 4090 still wins on value, and we'd buy it again.

Related guides on Best GPU for AI

The full version lives on Best GPU for AI — VRAM calculator, GPU comparison table, and live Amazon pricing.

Used RTX 3090 Buying Guide for Local LLM in 2026

Thurmon Demich — Sat, 30 May 2026 01:13:58 +0000

Cross-posted from Best GPU for LLM — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

The RTX 3090 is three generations old, costs under $900 used, and still fits 34B models that a new $500 GPU cannot touch. For LLM inference, VRAM is the hard constraint — and 24GB at ~$850 is the best value on the market in 2026. But the used GPU market has landmines: mining-worn cards, dead VRAM chips, and sellers who know you can't easily tell the difference before you buy.

This guide gives you the tools to buy one safely.

Why 24GB still matters for LLM in 2026

VRAM is a filter, not a preference. A model either fits in VRAM or it doesn't — and the boundary between "fits" and "doesn't fit" falls squarely at the 24GB mark for 30B+ models.

7B models (Q4_K_M): need ~4.5GB — runs on almost anything
13B models (Q4_K_M): need ~8GB — fits an RTX 4060 Ti 8GB, barely
34B models (Q4_K_M): need ~20-22GB — requires 24GB VRAM
70B models (Q4_K_M): need ~40GB+ — requires dual 24GB cards or a 48GB workstation GPU

For anyone running CodeLlama 34B, Qwen 2.5 32B, or Yi-34B locally, the RTX 3090 is the cheapest single GPU that actually fits the model. A new RTX 5070 Ti (16GB) cannot do it. A new RTX 5080 (16GB) cannot do it. The 3090's 24GB is the threshold card at the lowest price.

On typical 34B Q4_K_M inference, community benchmarks show a 3090 producing roughly 12-18 tok/s — slow compared to the RTX 4090's 20-25 tok/s, but well above the ~8 tok/s threshold most people consider interactive. For a full speed breakdown, see RTX 4090 vs 3090 for LLM.

VRAM chart available at the original article

Price tiers: what to pay and what to avoid

Price	Signal	Verdict
Under $600	Too low — suspect dead VRAM, damaged card, or scam	🔴 Red flag
$600–$699	Possible mining-heavy card or cosmetic damage — needs heavy scrutiny	🟡 Caution
$700–$900	Healthy range for a used 3090 with normal wear	🟢 Target zone
$900–$999	Still reasonable if from a reputable seller with receipts	🟡 Borderline
$1,000+	Overpaying — at this price, a used RTX 4090 is ~$1,100-1,200	🔴 Walk away

Set a ceiling of $900. If you're patient, quality 3090s appear regularly in the $750-850 range. Cards under $650 almost always have a reason for the discount.

Mining wear vs gamer wear — how to tell the difference

Both mining and gaming cards can be fine to buy. The concern is not the use case — it's the intensity and conditions of that use.

Mining wear signals (in photos):

Thermal pads look freshly replaced (miners often replace pads — this is actually good)
Heatsink fins have no dust (miners keep their rigs clean to manage temps)
PCB has a slight yellow tinge near VRMs from sustained heat
Mounting bracket shows no scratches (mining cards rarely get swapped between systems)
Backplate shows slight bowing — a sign of sustained thermal expansion/contraction cycles

Gamer wear signals (in photos):

Heavy dust accumulation in heatsink fins
Scratches on bracket from repeated install/removal
Original thermal paste (never replaced) — look for grey dried paste at the edges

Neither type is inherently bad, but a mining card that ran 24/7 for 18+ months at 250W+ has more total operating hours than most gaming cards. A gamer card with original paste at 5 years may actually have worse thermal compound degradation.

Questions to ask sellers before buying:

"What was the primary use — gaming, mining, or professional work?"
"How many hours of total runtime, roughly? Any way to check?"
"Has the thermal paste or pads been replaced? When?"
"Does the card throttle under load? Any driver crashes?"
"Will you accept a return within 14 days if I find a hardware defect after testing?"

A seller who answers these questions confidently and offers a return window is a better signal than any photo.

Inspection checklist before you commit

Run these checks within the first 48 hours — before your return window closes.

Step 1 — Install and verify with GPU-Z

Open GPU-Z immediately after driver install
Check VRAM reported: must show exactly 24384 MB
Check GPU clock speed: should boost to ~1695 MHz under load
Any VRAM showing as less than 24GB indicates chip failure

Step 2 — VRAM stress test

Run CUDA-Z or use python -c "import torch; t = torch.zeros(24000, 1024*1024//4).cuda(); print('VRAM OK')" in a Python env with CUDA
Alternatively, load a large model in Ollama: ollama run llama3:70b — this will try to allocate ~40GB (will fail gracefully but exercises VRAM access patterns)
Better: run memtest_vulkan or OCCT GPU Memory Test to fully exercise all VRAM cells

Step 3 — Temperature and throttle check

Run a 30-minute Ollama inference session on a 34B model
Monitor with nvidia-smi dmon -s pct — watch for thermal throttling (clock dropping while temp is above 83°C)
Expected idle temp: 30-45°C. Under load: 70-83°C is normal, above 85°C sustained is a concern

Step 4 — Fan and coil noise check

Under load, listen for coil whine (high-pitched electrical buzz — varies from imperceptible to irritating)
Fan noise: one fan bearing rattling is common and cheap to fix; all three fans rattling means the card was run hard without maintenance
A brief fan stop at low load is normal (zero-RPM mode)

Step 5 — Backplate inspection

Remove the card and inspect the backplate for bowing (slight curve away from PCB at center)
Mild bowing (1-2mm) is common and harmless
Severe bowing suggests the card was run without proper support — check PCB traces under the backplate if possible

Return-window strategy: Buy from sellers offering at least 14 days returns. Ship to a work address or a friend's address if you're buying a second card and your package history makes you a target for "item not as described" scams. Complete all testing within the first 72 hours.

Used RTX 3090 vs alternatives

GPU	VRAM	Tok/s (13B Q4)	Tok/s (34B Q4)	Price	Notes
RTX 3090 (used)	24GB	~40 tok/s	~14 tok/s	~$850	Best VRAM-per-dollar, no warranty
RTX 4090 (new)	24GB	~55 tok/s	~22 tok/s	~$1,600	57% faster on 34B, warranty
RTX 5090 (new)	32GB	~90 tok/s	~40 tok/s	~$2,000	Runs 34B and some 70B at Q4, best new card

The 3090 is the only option in this table under $1,000. It fits every model the 4090 fits, at roughly 60% of the speed, for roughly 55% of the price. If you're debating between a used 3090 and a new 4090, see our RTX 4090 vs 3090 for LLM comparison.

Common mistakes when buying a used 3090

Skipping the VRAM stress test. Dead VRAM cells are the most common failure mode on used 3090s. They often don't show up in normal use — only under sustained 24GB load. Run the test before the return window closes, not after.

Buying without a return option. Facebook Marketplace deals with no returns are high-risk. If a seller refuses any return policy, you're betting $800+ on their honesty. The risk-adjusted price of a return-eligible purchase from eBay or Craigslist with a verifiable seller is almost always worth any premium.

Paying $1,000+ because the listing says "lightly used." Every listing says "lightly used." Price discipline matters more than seller claims. If you see a 3090 over $950, compare it against used 4090 listings before committing.

Ignoring coil whine. Coil whine doesn't affect performance, but if it bothers you, there is no fix short of card replacement. Test under load before finalizing, especially if you work in a quiet environment.

Which 3090 buyer are you?

Buy a used 3090 at $800-900 if:

You need 24GB VRAM for 34B models and can't spend $1,600 on a new 4090
You're comfortable doing a basic hardware inspection and have a return window
You're building a multi-GPU LLM setup and need two 24GB cards affordably — our dual RTX 3090 guide walks through the full build

Buy a new RTX 4090 instead if:

You need warranty coverage and don't want to test hardware
You run 34B models interactively all day and want noticeably faster speed
Your budget is $1,600 and you plan to keep the card for 3+ years

Buy a new RTX 5090 instead if:

You want 32GB VRAM to run 70B models on a single card
Budget isn't the constraint and you want the best available hardware

Verdict

The used RTX 3090 is the best VRAM-per-dollar GPU available for local LLM in 2026, and that isn't close to changing. The 24GB threshold matters more than generational speed gains at this price point, and nothing new under $1,600 matches it for model capacity.

Buy one in the $750-900 range, run the inspection checklist within 48 hours, and you'll have a card that handles 34B models for years. The how much VRAM guide explains exactly why 24GB is the sweet spot if you want to understand the math behind the recommendation. With the 2026 GPU shortage tightening new-card stock, used 3090s have become even more attractive — but also more competitively priced.

Related guides on Best GPU for LLM

Read the full guide on Best GPU for LLM — includes our VRAM calculator, GPU comparison table, and live pricing.

Best GPU for LTX-Video in 2026: 5 Picks (Real-Time)

Thurmon Demich — Fri, 29 May 2026 01:14:20 +0000

From the Best GPU for AI archive. The canonical version has interactive calculators, an up-to-date GPU comparison table, and live pricing.

LTX-Video is the first open-source video model that genuinely feels interactive. Lightricks shipped it late 2025, and by May 2026 it has become the go-to local pick for creators who want to iterate on prompts without waiting 20 minutes for each clip. On a 4090, a 5-second 768x512 generation finishes before the clip would even finish playing — that is the hook, and it is real.

Quick answer: The RTX 4090 (24GB) is the best GPU for LTX-Video in 2026. It generates faster than real time on standard clips, has the VRAM for longer durations and i2v with reference images, and prices have settled at ~$1,600. The RTX 5090 is faster but pricier; a 16GB card like the RTX 5070 Ti is the budget floor.

Who this is for

This guide is for people generating short AI video locally with LTX-Video — social creators iterating on hooks, indie animators prototyping shots, and developers building on top of the model. If you batch-render long Sora-style cinematics overnight, this is not your guide. LTX-Video earns its keep when you want to try ten variations of a prompt in the next ten minutes.

What LTX-Video actually needs

LTX-Video is unusually efficient for a video model. The 2B-parameter DiT architecture and Lightricks' aggressive optimization mean it fits where Hunyuan and CogVideoX-5B do not. Realistic VRAM at common settings:

Workflow	Resolution	Duration	Min VRAM	Recommended
Text-to-video, standard	768x512	5 sec	10GB	12GB
Image-to-video, standard	768x512	5 sec	12GB	16GB
Longer clips	768x512	10 sec	14GB	16GB
Higher resolution	1216x704	5 sec	16GB	24GB
Long + high-res	1216x704	10 sec	20GB	24GB
Reference + ControlNet	varies	varies	18GB	24GB

In practice 12GB cards run the model but you fight constant memory pressure. 16GB is the comfortable floor. 24GB is where the workflow opens up — multiple LoRAs, longer durations, queued generations.

LTX-Video generation speed ranked

These are wall-clock seconds for a single 5-second 768x512 clip at default settings (30 steps, fp8). Numbers come from our own ComfyUI runs and community benchmarks; expect ±10% variance depending on your sampler and node graph.

GPU	VRAM	5-sec 768x512	10-sec 768x512	Price
RTX 5090	32GB	~3 sec	~7 sec	~$2,000
RTX 4090	24GB	~4 sec	~9 sec	~$1,600
RTX 3090	24GB	~7 sec	~15 sec	~$700 used
RTX 5080	16GB	~6 sec	~13 sec	~$1,000
RTX 5070 Ti	16GB	~8 sec	~17 sec	~$750
RTX 4070 Ti Super	16GB	~9 sec	~19 sec	~$700
RTX 4060 Ti 16GB	16GB	~16 sec	~34 sec	~$400

A 5-second clip plays for 5 seconds. The 4090 and 5090 generate it faster than that — that is what Lightricks means by "faster than real time." Everything from the 5080 down is still fast enough to iterate, just not literally instantaneous.

RTX 4090 — the right answer for most people

The 4090 is the GPU LTX-Video was tuned around in the community. 24GB of GDDR6X means you stop thinking about memory and start thinking about prompts. In our experience, the 4090's 24GB lets you stage 3-4 generations in queue while the next one renders, run a refiner pass, and keep ComfyUI's preview pipeline live without OOMs. That workflow alone is worth the price gap over 16GB cards.

The other quiet win: i2v (image-to-video) with a high-resolution reference frame fits comfortably. On 16GB you have to downscale references or trim frame counts. On 24GB you do not.

RTX 5090 — only if you batch

The 5090 generates roughly 25-30% faster than the 4090 and gives you 32GB. For interactive single-shot work that gap barely matters — both finish before you can read the seed. Where the 5090 pays back is batch: queue up 50 prompts overnight and it will finish meaningfully sooner. If you are not batching, the $400 premium over a 4090 buys you bragging rights more than throughput.

Mid-range: RTX 5080 and 5070 Ti

Both are 16GB, both run LTX-Video well at standard settings. The 5080 is roughly 30% faster than the 5070 Ti and worth the gap if you can stretch the budget. The 5070 Ti at ~$750 is the value sweet spot in the mid-range — fast enough to feel interactive, with enough VRAM to handle i2v and standard clip lengths.

What you give up at 16GB: long clips at higher resolution start hitting memory ceilings, and you cannot stack ControlNet + multiple LoRAs the way 24GB cards can.

Budget pick: RTX 4060 Ti 16GB

At ~$400 the 4060 Ti 16GB is the cheapest card we recommend for LTX-Video. It is meaningfully slower than the 5070 Ti — about 2x — but 16 seconds per clip is still usable. If you are learning the tool, prototyping concepts before committing to longer renders, or your video gen is a side project rather than a full workflow, this card does the job.

Used RTX 3090 — the wildcard

A used 3090 at ~$700 gives you 24GB of VRAM for less than half what a 4090 costs. LTX-Video runs noticeably slower than on the 4090 (the memory bandwidth difference shows) but you get the same VRAM ceiling. If you are price-sensitive but want 24GB for longer clips and reference frames, a tested used 3090 is a defensible pick. Just buy from a seller with returns.

Should you just use cloud?

Image-to-video at scale often pencils out better on cloud than on your own hardware. RunPod's A6000 and L40S instances run LTX-Video at full quality, and if you are spinning up generation jobs 3-4 hours a week the rental cost stays under what GPU depreciation alone would be. For experimentation, learning the model, or one-off campaigns, cloud is the honest answer.

Where local wins: daily iteration, working without an internet dependency, and the privacy of keeping reference imagery off third-party servers.

Which GPU should YOU buy?

Real-time interactive work — your main creative tool? RTX 4090 (24GB). Faster-than-real-time generation, headroom for i2v with reference frames, room for LoRA stacks.
Batch generation, dozens of clips per session? RTX 5090 (32GB). The 25-30% speed advantage and extra VRAM matter when you are queueing 50+ jobs.
Hobbyist, learning the model, occasional clips? RTX 4060 Ti 16GB (~$400) or used RTX 3090 (~$700). Both are honest entry points.
Mid-range new build? RTX 5070 Ti (16GB) at $750 is the value sweet spot — fast enough to feel interactive without flagship pricing.
Generate fewer than 5-10 clips a week? Skip the hardware entirely. RunPod is cheaper than a depreciation curve.

Common mistakes to avoid

Expecting Sora-quality output. LTX-Video is not Sora. Set expectations — it produces good 5-10 second clips with sometimes-shaky temporal coherence, not 60-second cinematic perfection. The trade-off for speed is fidelity, and that trade-off is the point of the model.
Underestimating temporal coherence at low VRAM. Pushing duration on a 12GB card forces the sampler into corners and you get more flicker, more identity drift on subjects, more "morph" artifacts. If coherence matters, do not undershoot VRAM.
Running fp16 when fp8 will do. The fp8 build is roughly 2x faster with quality differences most viewers cannot see at 768x512. Default to fp8 unless you have a specific reason not to.
Buying a 4090 for occasional use. If you generate 10 clips a month, you have just bought a $1,600 ornament. Use RunPod and put the money elsewhere.

Final verdict

Use case	GPU	Why
Best overall	RTX 4090 (24GB)	Faster-than-real-time, 24GB headroom, settled pricing
Maximum speed / batch	RTX 5090 (32GB)	25-30% faster, 32GB for the heaviest workflows
Mid-range value	RTX 5070 Ti (16GB)	Interactive speed at $750
Budget local	RTX 4060 Ti 16GB	Slowest of the picks, but $400 gets you in the door
24GB on a budget	Used RTX 3090	Same VRAM ceiling as a 4090, half the price
Occasional use	Cloud (RunPod)	Pays back vs. ownership under ~10 clips/week

LTX-Video changes the local AI video math. It is the first open model where "buy a 4090 and iterate fast" beats "rent an H100 and wait." For broader video-gen context see our best GPU for AI video overview, our AI animation GPU picks for AnimateDiff and SVD workflows, or the Hunyuan-Video hardware breakdown if you want the higher-fidelity (and much slower) alternative. Browse our full Guides library for related video-gen comparisons.

LTX-Video is not Sora — and that is the feature, not the bug. Buy the 4090, iterate ten times in the time Hunyuan renders once.

Related guides on Best GPU for AI

Read the full guide on Best GPU for AI — includes our VRAM calculator, GPU comparison table, and live pricing.

Best GPU for LLM Fine-Tuning in 2026 (Ranked Picks)

Thurmon Demich — Thu, 28 May 2026 01:14:00 +0000

Cross-posted from Best GPU for LLM — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

The RTX 4090 is the best consumer GPU for LLM fine-tuning in 2026. Its 24GB VRAM handles QLoRA on models up to 34B and full LoRA on 7B-13B. For anything larger, you need multi-GPU setups or cloud.

Who this is for

You want to fine-tune an open-source LLM on your own data — customer support responses, domain-specific documents, coding style, or creative writing. You need to know which GPU handles your training workload without running out of memory.

VRAM requirements by method

Method	7B Model	13B Model	34B Model	70B Model
Full fine-tuning	~30GB	~55GB	~140GB	~280GB
LoRA (r=16)	~18GB	~32GB	~72GB	~150GB
QLoRA (4-bit)	~8GB	~14GB	~24GB	~48GB

QLoRA is the game-changer for consumer GPUs. By quantizing the base model to 4-bit and training only the adapter layers, you reduce VRAM by 60-75% with minimal quality loss.

VRAM chart available at the original article

Best GPUs for fine-tuning

GPU	VRAM	Best Method	Max Model Size	Price
RTX 5090	32GB	QLoRA 70B / LoRA 13B	70B QLoRA	~$2,000
RTX 4090	24GB	QLoRA 34B / LoRA 7B	34B QLoRA	~$1,600
RTX 3090 (used)	24GB	QLoRA 34B / LoRA 7B	34B QLoRA	~$800
RTX 4060 Ti 16GB	16GB	QLoRA 13B	13B QLoRA	~$400
RTX 3060 12GB	12GB	QLoRA 7B	7B QLoRA	~$250

The used RTX 3090 at $800 is exceptional value for fine-tuning — same 24GB as the 4090 at half the price. Training is less bandwidth-sensitive than inference, so the older architecture barely matters. See our VRAM planning guide for more detail.

Which GPU should you buy?

Fine-tuning 7B models (QLoRA)? → RTX 4060 Ti 16GB ($400). Handles it with room to spare.
Fine-tuning 13B-34B (QLoRA)? → RTX 4090 ($1,600) or used RTX 3090 ($800). 24GB is the sweet spot.
Fine-tuning 70B? → RTX 5090 ($2,000) for QLoRA. Full LoRA on 70B requires multi-GPU.
Just experimenting? → Whatever GPU you already have. QLoRA on 7B works on 8GB cards.

Common mistakes to avoid

Attempting full fine-tuning on consumer GPUs. Full fine-tuning a 7B model needs ~30GB. Use QLoRA or LoRA instead — quality is nearly identical for most use cases.
Buying by TFLOPS instead of VRAM. Training needs VRAM first, compute second. A 24GB RTX 3090 beats a 16GB RTX 5080 for fine-tuning.
Forgetting gradient checkpointing. Enabling gradient checkpointing in your training config reduces VRAM by 30-50% at the cost of ~20% slower training.
Training without validation data. This isn't a GPU mistake, but overfitting on your dataset is the #1 reason fine-tunes fail. Always split your data.

Final verdict

Need	Best pick	Price
Best overall	RTX 4090	~$1,600
Best value	RTX 3090 (used)	~$800
Best budget	RTX 4060 Ti 16GB	~$400

QLoRA changed the game for consumer GPU fine-tuning. A $400 card can fine-tune 13B models that would have required $10,000 hardware two years ago.

Related guides on Best GPU for LLM

The full version lives on Best GPU for LLM — VRAM calculator, GPU comparison table, and live Amazon pricing.

What GPU Do You Need for SDXL in 2026? (5 Picks)

Thurmon Demich — Wed, 27 May 2026 01:14:21 +0000

Cross-posted from Best GPU for AI — visit the original for our VRAM calculator, GPU comparison table, and current Amazon pricing.

You open ComfyUI, load an SDXL checkpoint, add a ControlNet, and hit generate. Ten seconds later your GPU runs out of memory. Sound familiar? SDXL is significantly more demanding than SD 1.5, and choosing the wrong card means constant OOM errors or painfully slow generation. Here is what you actually need.

Who this is for

This guide covers GPU selection specifically for SDXL workflows — base generation, inpainting, ControlNet, upscaling, and LoRA training. If you are running SD 1.5 or Flux, see our dedicated guides for Stable Diffusion and Flux.

SDXL VRAM requirements

Workflow	Minimum VRAM	Recommended VRAM
SDXL base (1024x1024)	8GB	12GB
SDXL + ControlNet	10GB	16GB
SDXL + ControlNet + upscaler	12GB	16GB
SDXL + multiple ControlNets	14GB	16-24GB
SDXL LoRA training	10GB	16GB
SDXL Dreambooth	12GB	24GB

The baseline SDXL model uses about 6.5GB of VRAM. Each ControlNet adds 1.5-2.5GB. Upscalers add another 1-2GB. The overhead stacks up fast.

VRAM chart available at the original article

Best GPUs for SDXL ranked

GPU	VRAM	SDXL (1024px)	SDXL + ControlNet	Price
RTX 5090	32GB	~3.5 s/img	~4.5 s/img	~$2,000+
RTX 4090	24GB	~5.5 s/img	~6.5 s/img	~$1,600
RTX 5070 Ti	16GB	~7.0 s/img	~8.5 s/img	~$750
RTX 4070 Ti Super	16GB	~8.5 s/img	~10 s/img	~$700
RTX 4060 Ti 16GB	16GB	~12 s/img	~14 s/img	~$400
RTX 3060 12GB	12GB	~16 s/img	~19 s/img	~$250 used

The 16GB sweet spot

For SDXL, 16GB VRAM is the practical sweet spot. It handles:

Base SDXL generation at 1024x1024 with headroom
One ControlNet layer without memory pressure
Upscaling with Tile ControlNet or ESRGAN
LoRA training at reasonable batch sizes

Three cards hit this mark at different price points:

RTX 4070 Ti Super (~$700) — Best value for SDXL. Fast generation, 16GB VRAM, and strong compute. The card most SDXL users should buy.

RTX 5070 Ti (~$750) — Slightly faster with GDDR7 bandwidth. Worth the small premium if you generate images frequently.

RTX 4060 Ti 16GB (~$400) — Budget option that still has 16GB. Slower generation but handles the same workflows.

When you need 24GB or more

If you stack multiple ControlNets, run SDXL at 2K+ resolution, or train Dreambooth models regularly, 16GB gets tight. The RTX 4090 at 24GB eliminates memory concerns entirely and generates images nearly twice as fast as 16GB cards.

Which GPU should you buy?

Basic SDXL generation with occasional ControlNet: The RTX 4060 Ti 16GB at $400 handles this without issues. Generation is slower but you will not hit OOM.

Regular SDXL work with ControlNet workflows: The RTX 4070 Ti Super at $700 is the sweet spot. Fast enough for iterative creative work, 16GB covers complex pipelines.

Professional SDXL production or Dreambooth training: The RTX 4090 at $1,600 gives you 24GB and top-tier speed. Worth it if image generation is your daily workflow. For trainer-specific picks, our best GPU for Dreambooth guide covers VRAM and batch-size trade-offs in more detail.

Maximum headroom and fastest output: The RTX 5090 at $2,000+ guarantees you never think about VRAM again.

Common mistakes to avoid

Buying an 8GB card for SDXL. It technically works for base generation, but any ControlNet or upscaler pushes you over the edge. 12GB is the real minimum, 16GB is recommended.
Running SDXL at FP32 precision. Always use FP16 or BF16. FP32 doubles VRAM usage for zero visible quality improvement in generated images.
Ignoring VAE tiling for high-res work. If you upscale or generate above 1024px, enable VAE tiling to avoid OOM during the decode step. For dedicated upscaler workloads beyond SDXL's hires-fix, see our best GPU for AI upscaling guide.
Choosing raw compute over VRAM. A faster card with 8GB is worse for SDXL than a slower card with 16GB. VRAM determines what you can run; speed determines how fast.

Final verdict

Budget	GPU	Why
$250	RTX 3060 12GB (used)	Minimum viable SDXL card
$400	RTX 4060 Ti 16GB	Budget 16GB for SDXL
$700	RTX 4070 Ti Super	Best value for SDXL
$1,600	RTX 4090	Power user + training

For most SDXL users, a 16GB card is the right call. The RTX 4070 Ti Super delivers the best balance of speed, VRAM, and price. Check our full guides on Stable Diffusion GPUs and Flux GPUs for broader comparisons.

SDXL eats VRAM for breakfast. Buy 16GB minimum, and you will never fight OOM errors again.

Related guides on Best GPU for AI

The full version lives on Best GPU for AI — VRAM calculator, GPU comparison table, and live Amazon pricing.