Thurmon Demich

Posted on Jul 23 • Originally published at bestgpuforai.com

RTX 5090 vs 4090 for Flux.2 in 2026 (32GB FP8 Compared)

#rtx5090 #rtx4090 #flux2 #comparison

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

Quick answer: For Flux.2 workloads specifically, the RTX 5090 is roughly 30-40% faster than the RTX 4090 and the gap is bigger than it is on Flux.1 or SDXL. The reason is narrow but real — Blackwell has native FP8 tensor cores that Flux.2 was explicitly tuned for, while the 4090 falls back to a software FP8 path on Ada. If Flux.2 is your daily driver, the 5090 is the right buy. If you're a hybrid Flux.1 + occasional Flux.2 user, the 4090 still earns its keep.

Who this guide is for

This is specifically for people deciding between the RTX 5090 and RTX 4090 with Flux.2 as the primary workload. If you want the broader head-to-head across all AI workloads (LLMs, SDXL, fine-tuning), my RTX 4090 vs RTX 5090 for AI comparison covers that — this article narrows the lens to Flux.2 only, because the per-workload gap is meaningfully different from the general-AI picture.

I'm also assuming you've already seen the best GPU for Flux.2 buyer guide and shortlisted these two flagships. If you're still on Flux.1 and wondering whether the upgrade is worth the GPU jump, hold that question — I'll get to it in the contrarian section.

Specs side-by-side

Spec	RTX 4090	RTX 5090
VRAM	24GB GDDR6X	32GB GDDR7
Memory bandwidth	1,008 GB/s	1,792 GB/s
TGP	450W	575W
Architecture	Ada Lovelace	Blackwell
Compute capability	8.9	10.0
FP8 tensor cores	Software path (emulated via FP16/FP32)	Native hardware FP8
Street price	~$1,600	~$2,000

Two rows in that table do the heavy lifting for Flux.2 specifically: the bandwidth jump from 1,008 to 1,792 GB/s, and the FP8 row. Flux.2 at 32B parameters is bandwidth-bound during the diffusion steps, so the 78% bandwidth uplift translates into a near-linear speedup on the parts of the pipeline that matter most.

The FP8 difference is the one most people misread. Ada (compute cap 8.9) does technically support FP8, but through software emulation that schedules through the FP16/FP32 tensor units. Blackwell (compute cap 10.0) exposes FP8 directly as a first-class tensor op. On models that were trained or quantized with FP8 in mind — which is exactly what NVIDIA shipped for Flux.2 in May 2026 — Blackwell pulls ahead of Ada by more than the raw bandwidth gap alone would predict.

VRAM chart available at the original article

Real Flux.2 gen times

These are approximate ComfyUI numbers using the official NVIDIA Flux.2 FP8 checkpoint, Euler sampler, no ControlNet unless noted. My own measurements at 1024×1024 / 30 steps:

Flux.2 workload	RTX 4090	RTX 5090	5090 advantage
Flux.2 Dev FP8, 1024×1024, 30 steps	~14-16 s	~9-11 s	+35-40%
Flux.2 Schnell FP8, 1024×1024, 4 steps	~2.8 s	~1.9 s	+35%
Flux.2 + 1 ControlNet (Depth)	~18 s	~12 s	+33%
Flux.2 + 2 ControlNets (Depth + Canny)	tight on 24GB, ~24 s with swapping	~15 s, comfortable	+37% (plus no swap risk)
Flux.2 LoRA training, rank 32, batch 1	~22GB used, ~3.1 it/s	~24GB used, ~4.3 it/s	+38%
Flux.2 LoRA training, rank 64, batch 2	OOM at 24GB	~28GB used, ~3.6 it/s	5090 only

A few things worth flagging in those numbers:

The 35-40% gap is bigger than the 25-30% you'd see on Flux.1. That's the native-FP8 effect. On Flux.1, both cards run FP16 or community FP8 paths that don't hit Blackwell's tensor units as directly.
The ControlNet stacking row is where 24GB starts to bite. Two ControlNets plus the text encoder plus FP8 Flux.2 weights pushes the 4090 into VRAM swapping territory. The 5090's 32GB has another 8GB of headroom that keeps the whole pipeline resident.
Rank-64 LoRA training simply doesn't fit on a 4090 at batch size 2 with Flux.2. You can drop to rank 32 and batch 1, but if you're serious about LoRA training on Flux.2, the 5090 unlocks training configurations the 4090 just won't run.

Which should YOU buy?

You generate Flux.2 daily, stack ControlNets, train LoRAs: RTX 5090. The 32GB headroom plus native FP8 is exactly the workload these tradeoffs were designed for, and the per-image speedup pays back fast if you iterate a lot.
You're a Flux.2 hobbyist, single-ControlNet workflows, no training: RTX 4090 is honestly fine. 14-16s per image is iterable, and the $400 saving is real money for the rest of your build.
You do mostly Flux.1 and only occasional Flux.2: RTX 4090. Flux.1's FP16 weights fit comfortably in 24GB and the bandwidth gap matters less on the smaller 12B model. The 5090 premium isn't worth it for this profile — see my Flux.1 GPU guide for the broader Flux.1-first picture.
You're considering Flux.2 vs SD 3.5 and aren't sure yet: Flux.2 vs SD 3.5 hardware covers which model's hardware footprint fits which budget — relevant if you haven't fully committed to Flux.2 yet.
You bought a 5090 already and want sanity-check: my 6-month RTX 5090 retrospective covers what real users actually got from the upgrade across image-gen, LLMs, and training. Flux.2 is one of the workloads where the upgrade has aged the best.

The contrarian read: if you only run Flux.1, the 5090 isn't worth it

I want to be direct about this because it's the question I get most often. If your workflow is Flux.1 Dev (12B parameters, FP16-native), an RTX 4090 already runs it in ~6-8s per image with full 24GB headroom for ControlNets and IP-Adapter. The 5090 will cut that to ~4-5s, but you're spending $400 to save a few seconds per image on a workload that doesn't stress either VRAM or FP8. That's a bad upgrade.

The 5090's case rests almost entirely on workloads built for it — Flux.2 FP8, 30B+ LLMs, and rank-64+ LoRA training. If those aren't on your roadmap, save the $400.

Common mistakes when picking between these two

Buying the 5090 "to future-proof" for Flux.1 workflows. As above — if Flux.1 is your real workload, the 4090 already over-serves it. You're paying $400 for headroom you won't use until your model choice changes.
Underestimating the PSU upgrade. The 5090 wants 575W TGP and NVIDIA officially recommends a 1000W PSU. If you're on an 850W unit (typical 4090 build), add ~$150-200 for a proper PSU upgrade. That's real money on top of the GPU premium.
Treating "FP8 support" as a yes/no checkbox. Both GPUs technically run Flux.2's FP8 weights — but the 4090 emulates FP8 through FP16 tensor units while the 5090 runs it natively. That's why the 35-40% gap is bigger than raw bandwidth predicts. Don't compare these GPUs on Flux.2 the same way you'd compare them on FP16 Flux.1.
Ignoring the 32GB headroom on multi-ControlNet workflows. A 4090 can technically run Flux.2 + 2 ControlNets, but with the text encoder loaded it starts swapping to system RAM, and gen time balloons from 16s to ~24s. The 5090 keeps the whole pipeline resident in VRAM. If your workflow is "two ControlNets + IP-Adapter," that's the 5090's strongest case.

Final verdict

Workload pattern	Better buy	Why
Daily Flux.2 + ControlNet stacks + LoRA training	RTX 5090	Native FP8 + 32GB headroom + 1792 GB/s bandwidth
Single-ControlNet Flux.2, no training	RTX 4090	14-16s per image is fine, save $400
Mixed Flux.1 / occasional Flux.2	RTX 4090	Flux.1 doesn't stress either card; 5090 premium unjustified
Pure LLM use case (no image gen)	See 4090 vs 5090 general AI guide	Different bandwidth-vs-VRAM tradeoff

For Flux.2 specifically, the 5090 is one of the rare cases where a flagship-to-flagship upgrade actually earns its 25% price premium — but only if Flux.2 is your real workload, not just one of several models in your rotation.

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.

DEV Community