DEV Community: aixx

NVIDIA GB300 Superchip

aixx — Sun, 01 Nov 2026 09:41:54 +0000

Overview

NVIDIA GB300 Superchip (Grace Blackwell Ultra Superchip) is NVIDIA's Blackwell Ultra generation superchip combination: 1× Grace CPU (72-core Arm Neoverse V2) and 2× Blackwell Ultra (B300) GPUs co-packaged with NVLink coherent interconnect — each GPU comes with 288GB HBM3E , i.e., 576GB GPU memory per Superchip . Delivered as a GB300 NVL72 The building-block unit of AI factories delivered as a full rack (72 GPUs + 36 Grace), launched at GTC on March 18, 2025 alongside Blackwell Ultra[1][2].

NVIDIA's generational claim is straightforward: versus the previous Blackwell generation (GB200 era), 1.5× dense FP4 performance 、 2× attention performance ; The official page now shows dual figures for NVL72 full-rack FP4 Tensor Core — "1,440 | 10,802 PFLOPS" (verified 2026-10[1]; both figures per the official page footnote); vs the Hopper generation, NVIDIA officially states 5× throughput per megawatt 、 30× real-time video generation[1]。

The channel reality must be addressed first : the GB300 generation not received export licenses for China — unavailable through legitimate channels in mainland China; the open market centers on the H20 and compliant derivative form factors[4]. Meanwhile, the GB300 no single-card or single-Superchip retail form and ships only as full NVL72 racks through OEMs/integrators — it is AI-factory infrastructure, not a channel commodity.

Use Cases

AIXX Index (0–100): estimated by this Pedia via weighted public specifications — memory 40% + interconnect 30% + ecosystem 30%, with weights adjusted by use case. 90+ flagship tier · 75+ Excellent · 60+ usable · below 60 not recommended. The methodology is transparent and verifiable, and does not constitute a performance guarantee.

History

March 18, 2025 , announced at GTC Blackwell Ultra : The GB300 Superchip and GB300 NVL72 full rack debuted, with official delivery by major OEMs and cloud providers from mid-year (per public reports[2]).

First half of 2025 , Supermicro, Dell, Lenovo and other OEMs plus rack integrators such as Wistron successively disclosed GB300 delivery schedules (per public reports[3]).

Second half of 2025 , GB300 NVL72 instances/clusters went live at major overseas cloud providers, and Blackwell Ultra entered its ramp-up phase (per public reports[3]).

2026 , the official page marked "Available Now," with Wistron's new Texas plant folded into Grace Blackwell Ultra capacity (official page, verified 2026-10[1]); the China export license never came — the compliant mainland market is out of reach[4].

Technical Specifications

Item	Specifications (GB300 Superchip / NVL72 · per NVIDIA official figures, verified 2026-10[1])
Combination	1× Grace CPU (72-core Arm Neoverse V2) + 2× Blackwell Ultra GPUs (NVLink coherent interconnect)
Memory per GPU	288GB HBM3E (Blackwell Ultra generation; previous-gen B200 was 192GB)
Memory per Superchip	2× 288GB = 576GB
CPU Memory	LPDDR5X (17TB total CPU-side across NVL72 rack · 14TB/s, official figures[1])
FP4 Tensor Core	Official page lists NVL72 as "1,440
Generational performance	Dense FP4 1.5× previous generation · attention 2× (GTC launch figures[2])
NVL72 rack	72 GPUs + 36 Grace (2,592 cores) · 20TB total GPU memory · 576TB/s aggregate bandwidth (official figures[1])
NVLink	Fifth generation · 130TB/s across the NVL72 rack (official figures[1])
Networking	ConnectX-8 SuperNIC · 800Gb/s per GPU (official figures[1]); Quantum-X800 InfiniBand / Spectrum-X Ethernet
Cooling	Fully Liquid-cooled rack architecture (official figures[1])
Status	Available Now (official page, verified 2026-10[1])

Specifications per the official NVIDIA GB300 NVL72 page (verified 2026-10[1]); single-Superchip-level parameters are derived from the publicly disclosed rack composition (36× Grace / 72× GPUs), as NVIDIA publishes no standalone Superchip retail spec sheet; this wiki does not convert the dual FP4 figures itself — defer to the official page footnote.

Generational comparison · GB200 vs GB300 Superchip

Item	GB200 Superchip	GB300 Superchip
GPU configuration	2× B200（Blackwell）	2× B300 Ultra（Blackwell Ultra）
Memory per GPU	192GB HBM3e	288GB HBM3E（+50%）
dense FP4	Generational baseline	1.5× (official figures[2])
attention	Generational baseline	2× (official figures[2])
Networking per GPU	ConnectX-7 · 400Gb-class	ConnectX-8 · 800Gb/s (official[1])
System form	NVL72 (homogeneous rack)	NVL72 (fully Liquid-cooled, official[1])
China channel	No license	No license[4]

Comparison based on public figures; GB200-generation data can be found in our GB200 Superchip / GB200 NVL72 entries. The memory generational gap (192→288GB) on MoE inference capacity can be computed directly from parameter counts.

Family Members

Grace Blackwell / Blackwell Ultra family tree:

Cooling & Deployment

The GB300 NVL72 uses a fully liquid-cooled rack architecture (official figures[1]): one rack of 72 GPUs + 36 Grace, delivered as a complete set with coolant distribution units (CDUs) plus primary/secondary loops — liquid-cooling retrofit of the machine room, power capacity (hundreds of kW per rack, public reporting), and floor loading are the three deployment prerequisites . This isn't "buying a machine" but "building a workshop": delivery lead times, piping work, and commissioning typically run in months (integrator figures).

A single Superchip has no standalone deployment form — there is no way to "install GB300 in an ordinary server"; for 8-GPU board-level needs, look to HGX B300 (air-cooled availability is defined by OEM server vendors).

Model Compatibility

Scenario	Fit for NVL72's 20TB aggregated memory (estimated: weights × precision + KV/framework overhead)
Trillion-parameter training	20TB GPU memory + 130TB/s NVLink across the rack[1]: the current mainstay foundation for frontier LLM training
FP4 inference	Dense FP4 1.5× previous generation[2]: a generational throughput gap for high-concurrency generative inference
Long context	Attention 2× (official figures[2]): a clear beneficiary for long-context/multimodal inference
Video Generation	Officially 30× real-time video generation vs Hopper[1]: a deployment base for video-generation models
Not applicable	Compliant procurement in mainland China (no license[4]); small-scale deployments (full-rack threshold doesn't fit)

Methodology: memory usage ≈ parameter count × bytes per parameter, plus KV and framework overhead; same source as the selector's "memory math" and reproducible.

Gallery

GB300's minimum delivery unit is the full rack, and all official public materials are rack/system-level — NVIDIA has not published a single-Superchip render, so this entry does not substitute other images.

Video Overview

Video walkthrough of this entry is in the production queue (see the video slot schedule for Chinese/multilingual workflow) — this section updates automatically once it goes live.

Software & Ecosystem

Continuity of the full CUDA stack: the GB300 generation runs the NVIDIA AI Enterprise suite and the NIM microservices ecosystem, with TensorRT-LLM specifically optimized for Blackwell Ultra's FP4/attention features (official figures[1][2]). The Grace CPU side is Arm-based (kin to the DGX Spark's GB10), with CUDA's unified memory model covering the Grace-GPU coherent interconnect. On networking, ConnectX-8 + Quantum-X800/Spectrum-X form the "AI factory" networking standard (official figures[1]).

Ecosystem position: GB300 is the flagship vehicle of the "AI factory" narrative — buying it bundles NVIDIA's full reference architecture (networking/software/operations). At selection time, that's part of the premium and part of the lock-in.

Purchasing & Inspection

Common risks

For rack-level procurement, engage third-party inspection and full-rack integration testing, with acceptance reports matched one-to-one to serial numbers/rack IDs.

How to Stress Test

Rack-Level Validation Tools: DCGM (rack-wide GPU health/power/temperature monitoring), nccl-tests (measured NVLink/IB Interconnect Bandwidth), gpu-burn (full-rack Burn-in Test), bandwidthTest (HBM bandwidth). On the model side, use the MFU Metric for acceptance: model compute utilization of target workloads on the full rack against vendor commitments—this is the only watermark-free acceptance metric for rack-level procurement. A dedicated validation segment for this card is in production.

Shipping & Delivery

Stage	Key Points
Form Factor	Rack-level delivery (rack-grade weight and volume)—survey hoisting/forklift/aisle width against data center engineering in advance
Liquid-cooled Pipelines	Quick-connector transport protection and cleanliness management; prioritize pipeline flushing and water quality testing upon arrival
Cross-border Compliance	Flagship AI infrastructure is export-control-sensitive — cross-border transfers require full compliance documentation throughout (no license for China[4])
Insurance	Insure at full rack value (public reports put GB200-generation racks at roughly the US$3 million level[3]; for GB300, go by the quote); consider adding engineering insurance

OEM & Variants

Variants / OEM	Form Factor & Positioning
GB300 NVL72	72 GPU + 36 Grace full rack · flagship form factor (official figures[1])
Supermicro/Dell/Lenovo and others	OEM rack solutions (per public reports[3])
Wistron	Rack integrator · new Texas plant added to capacity (official page[1])
HGX B300	8-Socket Board Form · Integrated by OEM Server Vendors (Air-cooled Availability Defined by OEM)
DGX SuperPOD	GB300-Class Cluster Scaling · Data Center Flagship System

Limitations & Controversies

No License for China : The GB300 generation has no export license for China[4] — unavailable in the compliant mainland market; this is the premise and boundary of all procurement discussion in this entry.
Rack Form Factor Barrier : No single-GPU/single-Superchip retail—irrelevant for small-scale users and technical validation teams; HGX B300/B300 NVL16 are the accessible forms.
FP4 Metric Dual Listings : The official page's 1,440 | 10,802 PFLOPS dual figures (verified 2026-10[1]) coexist with the GTC keynote's "1.1 exaflops dense FP4" claim (public reports[2]) — this wiki does not convert between them; contract terms should follow official confirmation documents.
Liquid Cooling and Power Barriers : Fully Liquid-cooled rack + hundred-kW-class power supply (public reporting)—traditional Air-cooled data centers cannot support it, and retrofit costs often exceed rack-price expectations.
ecosystem lock-in : NVLink/CX-8/software suite full-stack NVIDIA—the efficiency dividend of the "AI factory" and vendor lock-in are two sides of the same coin.

Market Notes

Date	Type	Market	Price Band	Basis
2026-10	Rack pricing	Overseas OEMs/Solution Providers	No standing official price · quote-based (the previous-gen GB200 NVL72 was publicly reported at roughly the US$3 million level[3], a useful reference anchor)	Full-rack delivery
2026-10	Cloud Rental	Overseas Clouds	Blackwell Ultra instances subject to cloud providers' pricing pages (billed per GPU-hour)	Third-Party Metrics
2026-10	Marketplace Status	Mainland China	Not applicable (no license[4]) — the marketplace centers on B300 NVL16 and similar form factors	Marketplace Search › (/hall.html)

Basis: this wiki has no transaction data for this model; rack-level pricing is subject to formal OEM/solution-provider quotes, with publicly reported figures serving only as order-of-magnitude anchors. Market intel and listings: AIXX Exchange (/hall.html) (Zero-commission Matching; not a party to transactions).

Timeline

Time	Event
2025-03-18	GTC Unveils Blackwell Ultra (GB300 Superchip / NVL72)[2]
First Half of 2025	OEMs/rack integrators have successively disclosed delivery schedules (per public reports[3])
Second half of 2025	GB300 instances went live on overseas clouds, entering the ramp-up phase (per public reports[3])
2026	Official page shows "Available Now"; Wistron's Texas capacity folded in (verified 2026-10[1])
Ongoing	No China License—Mainland Compliant Market Misses the Blackwell Ultra Generation[4]

FAQ

Related Entries

References

[1]	NVIDIA Official · GB300 NVL72 Product Page (72 GPU+36 Grace/288GB/20TB/130TB/s/CX-8/Fully Liquid-cooled/FP4 Dual Metrics/Available Now) · https://www.nvidia.com/en-us/data-center/gb300-nvl72/ (https://www.nvidia.com/en-us/data-center/gb300-nvl72/) (verified 2026-10)
[2]	NVIDIA GTC 2025 Keynote · Blackwell Ultra Launch (dense FP4 1.5×/attention 2× Metrics) · https://www.nvidia.com/en-us/events/gtc/ (https://www.nvidia.com/en-us/events/gtc/) (per public reporting)
[3]	Public Reporting · GB200 NVL72 Rack at ~$3 Million Level and GB300 Delivery Updates · https://www.reuters.com/technology/ (https://www.reuters.com/technology/) (per public reporting)
[4]	Public Reporting · China Export Controls (Blackwell Ultra Generation Not Licensed) · https://www.reuters.com/technology/ (https://www.reuters.com/technology/) (per public reporting)
[5]	Selector tool · VRAM budgeting methodology (parameters × precision + KV/framework overhead) · /wiki/xuanxing.html (/wiki/xuanxing.html) (this site's methodology)
[6]	This Site · GB200 NVL72 / GB200 Superchip / B300 Entries (Generational Comparison Data Sources) · /wiki/gb200-nvl72.html (/wiki/gb200-nvl72.html) (This Site's Entry)

Specifications are anchored to NVIDIA's official page; "public reporting" items are media paraphrases and this database does not vouch for their accuracy; the database does not convert between dual FP4 performance metrics—procurement should rely on official confirmation documents and measured results.

Vendor	NVIDIA
Architecture	Grace + 2× Blackwell Ultra
CPU	Grace · 72-Core Neoverse V2
GPU Memory	2× 288GB HBM3E = 576GB
Generational performance	dense FP4 1.5× · attention 2×
Rack Form Factor	NVL72：72 GPU + 36 Grace · 20TB
NVLink	Fifth Generation · 130TB/s Rack-Wide
Networking	ConnectX-8 · 800Gb/s per GPU
Cooling	fully liquid-cooled rack architecture
Status	Available Now (Verified 2026-10)
China channel	Not Licensed (Not Applicable)
Market	No Official Price · Marketplace Search › (/hall.html) (rack via OEM/solution-provider quotes)

Originally published on AIXX Pedia, the AI compute hardware encyclopedia maintained by the AIXX exchange team. Specs, pricing context and verified-inventory insights are updated there first.

HGX B300 NVL16

aixx — Sun, 01 Nov 2026 09:40:11 +0000

Overview

HGX B300 NVL16 of the Blackwell Ultra generation, in 16-GPU NVLink fully connected form : Sixteen B300 (Blackwell Ultra) GPUs form an NVLink domain within the baseboard/compute tray assembly—each card carries 288GB HBM3E, FP4 throughput is substantially higher than the previous generation, and aggregate NVLink 5 switching bandwidth sits around 14.4TB/s (16-GPU domain basis, NVIDIA HGX platform figures[1]). It occupies the middle tier of the form-factor spectrum: below the 8-GPU HGX baseboard system, above the GB300 NVL72 rack.

To be precise about positioning: NVIDIA's officially promoted HGX B300 baseboard is the 8-GPU configuration, while NVL16 refers to the 16-GPU NVLink fully connected configuration —The compute half of the GB300 NVL72 rack (2×8-GPU compute trays + NVLink switching) is exactly this form factor, and system vendors also offer project-based delivery of a 16-GPU NVLink domain (public sources[2]). For buyers it's a compromise of "rack-scale compute, baseboard-level granularity": one step up from an 8-GPU system in NVLink domain size, one step down from a full rack in datacenter retrofit requirements.

In the 2026 spot market, the B300 NVL16 form factor entered circulation in step with GB300 NVL72 volume: project leftovers, OEM solutions, and tray-level trading coexist. This entry carries no market band — actual price subject to the seller's quote.

Use Cases

AIXX Index (0–100): estimated by weighted public specifications — compute 40% + interconnect 30% + deployment friendliness 30%. 90+ flagship · 75+ Excellent · 60+ usable · below 60 not applicable.

History

2025 , as the Blackwell Ultra generation rolls out: B300 (Blackwell Ultra) ships with GB300 NVL72 racks and the HGX B300 platform—each B300 carries 288GB HBM3E with NVFP4 throughput improved over B200, and the HGX platform positioning defines a 16-GPU NVL form factor (NVIDIA HGX platform page[1]).

2025-2026 , the form-factor split took shape: 8-GPU HGX B300 baseboards serve small-to-mid deployments, GB300 NVL72 serves hyperscale clusters, and the 16-GPU NVLink domain form factor (NVL16) lands as the middle tier via both system vendor project builds and the rack's compute half (public sources[2]).

2026 , spot circulation: the B300 NVL16 form factor entered spot-market view alongside GB300-generation volume — OEM project leftovers and tray-level trading in parallel. Cross-border trade policies for high-end models continue to evolve; at procurement, defer to regulatory policy and professional advice (per public reporting, not legal advice).

Technical Specifications

Item	Specifications (HGX B300 NVL16 · official basis)
GPU	16× NVIDIA B300（Blackwell Ultra，SXM）
Board Memory	288GB HBM3E
In-domain memory	16×288GB = 4.6TB HBM3E (estimated: per-GPU ×16)
Board bandwidth	8TB/s HBM3E
Per-GPU compute	Around 15 PFLOPS FP4 dense (NVIDIA HGX platform figures[1])
NVLink	Fifth-generation NVLink · aggregate NVLink Switch bandwidth around 24TB/s (platform figures[1])
Form Factor	Baseboard/compute tray group (16-GPU NVLink domain), predominantly liquid-cooled
Generation Positioning	Same-origin form as the compute half of the GB300 NVL72 rack; 8-GPU HGX B300 is the smaller tier

Specifications per NVIDIA HGX platform figures (as of October 2026)[1]; system power draw, power delivery, and tray physical specs follow the system vendor's configuration. Not listed here = unverified—no guessing.

B300 Three-Tier Form-Factor Comparison

Item	8-GPU HGX Baseboard	NVL16 (this entry)	GB300 NVL72
NVLink Domain	8 GPUs	16 GPUs	72 GPUs
In-domain memory	2.3TB (estimated)	4.6TB (estimated)	~20TB (estimated)
Deployment Barrier	Air-cooling viable	Mostly liquid-cooled	Full-rack liquid-cooling retrofit
One-line buying advice	Small-cluster starting point	Hundred-GPU building block	Very large clusters, full rack

In-domain memory = per-GPU memory × GPU count (estimated); see respective entries for B300/GB300 NVL72.

Family Members

The B300 form-factor spectrum splits into three tiers by NVLink domain size; NVL16 is the middle tier. Family members:

Cooling & Deployment

Cooling for the NVL16 form factor centers on Mostly liquid-cooled : 16-GPU NVL16 systems come in both air-cooled and liquid-cooled versions (8U rack form, per industry sources); confirm the OEM solution's cooling type before purchase. The liquid-cooled option also involves a CDU loop (smaller scale than rack-level — see rackmount CDU), with facility-side water conditions and quick-disconnect workmanship to be checked item by item (see the Liquid-cooled CDU entry).

Compared with a full NVL72 rack, the NVL16 is one tier friendlier to deploy: no rack-level power retrofit or lift planning needed — it lands at the rack level. But the NVLink switching (16-GPU in-domain full interconnect) remains a baseboard-level component — NVLink domain health is the first inspection item at acceptance (per this library's compilation).

Use-Case Fit

Workload	B300 NVL16 Performance Basis	AIXX Index
Hundred-billion-parameter training/fine-tuning	16 GPUs in one domain, no cross-node communication within the NVLink domain	88
Extra-large model single-domain inference	4.6TB in-domain memory (estimated) fits extra-large models + long contexts	89
Hundred-GPU cluster building block	6×NVL16 + fabric makes a hundred GPUs, in-domain efficiency on par with rack solutions	90
Traditional FP64 HPC	Not Blackwell's home turf — weigh double-precision workloads separately	70

Workload tiers are Pedia editorial notes based on public specs (estimates; not a performance commitment).

Gallery

Image is the official HGX platform image; field photos of NVL16 systems will be added per inspection batch.

Software & Ecosystem

Item	Requirements / Status
Base Software	NVIDIA AI Enterprise family (shared with the HGX platform)
Collective Communication	NCCL auto-detects NVLink domain topology
Cross-domain networking	ConnectX-8/7 + IB (Quantum family) or Spectrum-X Ethernet
Storage	GPUDirect Storage pass-through (see the NVMe All-Flash Storage entry)
System Vendor	Dell/Supermicro/Lenovo/Inspur and other HGX B300 system solutions (per public product lines)

The HGX platform software ecosystem is common to the 8-GPU form; versions evolve over time — defer to official NVIDIA releases.

Purchasing & Inspection

Three inspection points when buying a B300 NVL16:1NVLink domain acceptance as a wholeThe 16-GPU domain is where the value lies: nvidia-smi nvlink all ports Active, NVLink switch firmware status, sampled in-domain bandwidth tests — a single degraded point hurts the whole domain; verify the entire domain before signing off.2Cooling solution confirmationFirst ask whether the system is a liquid-cooled or air-cooled variant, and whether CDU accessories ship with the liquid-cooled solution — "liquid cooling included by default" is a common misunderstanding; spell it out in the contract.3Tray/baseboard asset verificationMatch serial numbers against accompanying documents item by item; for NVL16 units pulled from a rack, verify tray condition against the original rack configuration.For the full six-test verdict framework see this entry "How to Stress-Test" (#benchtest) section.

How to Stress Test

Six-test burn-in framework, with all benchmarks anchored to the B300 NVL16: unboxing identification → NVLink domain → cooling loop → compute baseline → storage passthrough → extended Burn-in Test.

Shipping & Delivery

OEM & Variants

Deployment paths for the NVL16 form factor, with same-generation comparisons:

Path	Form Factor	Differentiators	Status as of 2026
Rack compute half	16-GPU NVLink domain inside GB300 NVL72 (2 compute trays + switching)	Ships with the rack, not sold separately	Rack basis
OEM 16-GPU solution	Project systems built on the HGX B300 platform	Configuration and cooling vary by OEM	Project-based circulation
8-GPU HGX baseboard	Standard HGX B300	One tier smaller domain, air-cooling viable	Spot-market mainstream

Form-factor information compiled from NVIDIA HGX platform figures[1] and public sources[2]; specifics are subject to the system vendor's official documentation.

Glossary Cards

In NVIDIA product names, NVL + a number = the GPU count in a fully connected NVLink domain: NVL16 means 16 GPUs in one NVLink domain. The bigger the number, the bigger the domain, and collective communication within the domain bypasses the network stack — the essential difference between NVL16 and NVL72 is simply domain size.

The enhanced Blackwell generation: 288GB HBM3E per GPU (up from B200), NVFP4 throughput a further step up — the core GPU of GB300 NVL72 and the HGX B300 platform. SXM package, baseboard-level integration.

The NVLink domain's switching component: baseboard/tray-level switch chips weave 16 GPUs into full interconnect — equal bandwidth between any two GPUs in the domain. NVL16's aggregate switching bandwidth is at the 24TB/s class (platform basis), the physical foundation of the "giant GPU" experience.

The compute unit inside racks/systems: the GB300 NVL72 packs 18 compute trays per rack (4 GPUs + 2 CPUs per tray) — the NVL16 form corresponds to 2 compute trays + NVLink switching, and tray-level trading is the main granularity in the secondary market.

Limitations & Controversies

Form-factor positioning, stated clearly first : NVIDIA officially leads with the 8-GPU HGX baseboard and the GB300 NVL72 rack—"NVL16" is largely an industry term for the 16-GPU NVLink domain form factor (the rack's compute half / system vendor project builds). When procuring, verify the specific vendor's configuration and tray composition; don't treat it as a "standard SKU" (public sources[2]).
Two Cooling Versions Confirmed : NVL16 systems come in both air-cooled and liquid-cooled versions (8U rack form factor)—the liquid-cooled variant offers higher performance ceilings but requires CDU support, while the air-cooled variant is friendlier to existing data centers. Spell out the cooling attribute in the contract.
The Domain-Size Dilemma : The selling point is interconnect efficiency one tier above an 8-GPU domain, but scaling a 16-GPU domain to hundreds of GPUs still requires fabric—budget for both layers of cost: 'in-domain efficiency + out-of-domain networking' (compiled by this Pedia).
Second-hand Market Circulation is still in its early stages : The Blackwell Ultra generation only rolled out in 2025, so residual-value benchmarks for used NVL16 systems aren't mature yet—verify tray provenance, liquid-cooling component lifespan, and NVLink component condition item by item (this Pedia's inspection experience).
Export-control-sensitive model : The Blackwell Ultra generation falls under close export-control scrutiny (per public reporting)—consult professional counsel on cross-border transfer compliance first; this Pedia does not constitute legal advice.

Market Notes

B300 NVL16 systems trade at two levels—OEM project designs and rack teardown trays—and wide configuration and cooling variations create large price dispersion—no verifiable unified market snapshot yet, AIXX Pedia honesty principle: no market band; prices are subject to the seller's quote 。

Spot search: AIXX Exchange Hall › (/hall.html?model=B300) (Zero-commission Matching; no involvement in transactions; no knowledge of or speculation on transaction prices — hence data neutrality). This section will switch to a standard market band once enough snapshots accumulate.

Price milestone timeline

FAQ

Not a single standard product in the strict sense: NVIDIA officially centers on the 8-GPU HGX baseboard and GB300 NVL72 rack; NVL16 is a general term for the '16-GPU fully NVLink-interconnected configuration'—implemented as a rack compute half or OEM project design. When purchasing, verify the specific design and tray contents.

It comes down to the NVLink domain: NVL16 is a single 16-GPU domain (fully interconnected within, no cross-machine communication), while two 8-GPU systems are two independent domains connected via the network. Collective communication efficiency differs by a tier in both training and inference—that's the source of the NVL form factor's premium.

NVL16 systems come in air-cooled and liquid-cooled versions (per industry sources): the liquid-cooled version offers a higher thermal ceiling but needs CDU loop support; the air-cooled version (8U rack) is friendlier to existing facilities. Decide based on your data center and OEM design.

It's a line of scale and barriers: NVL16 is a building block toward hundred-GPU clusters, deployable at rack level; NVL72 is a full-rack solution requiring data center liquid-cooling retrofits up front. Both architecture generation (Blackwell Ultra vs first-gen Blackwell) and domain size (16 vs 72) differ—decide by cluster scale and facility conditions.

Three things: full NVLink domain acceptance testing (all ports Active + in-domain bandwidth), cooling solution attributes and loop condition, and tray/baseboard asset verification. The Blackwell Ultra generation is still early in circulation—tray provenance and mixed-configuration history should be written into the contract.

Related Entries

Model comparison and selection

References

NVIDIA official site — HGX platform page (B300 platform compute and NVLink bandwidth figures, Oct 2026)　https://www.nvidia.com/en-us/data-center/hgx/ (https://www.nvidia.com/en-us/data-center/hgx/)
ESaiTech — HGX B300 NVL16 Architecture Guide (public industry material, form-factor positioning reference)　https://esaitech.com/blogs/insights/hgx-b300-nvl16-architect… (https://esaitech.com/blogs/insights/hgx-b300-nvl16-architecture-guide-for-enterprise-ai)
NVIDIA official site — GB300 NVL72 product page (rack compute half figures)　https://www.nvidia.com/en-us/data-center/gb300-nvl72/ (https://www.nvidia.com/en-us/data-center/gb300-nvl72/)
Marketplace Search — AIXX Exchange (Zero-commission Matching; no market snapshot for B300 NVL16 yet, actual price subject to the seller's quote)　https://www.aixx.vip/hall.html?model=B300 (/hall.html?model=B300)

Video Overview

Community real hardware · Source: YouTube

YouTubeThe Largest Server I've Seen · Supermicro HGX B300 Hands-OnSource: Craft Computing*YouTube*NVIDIA HGX B300: Tested On-Site by CoreWeave EngineersSource: CoreWeave (javascript:void(0))Playback may not be available in some regions

Vendor	NVIDIA (delivered by HGX platform-authorized OEMs)
Category	16-GPU NVLink fully connected form
GPU	16× B300（Blackwell Ultra SXM）
Board Memory	288GB HBM3E
In-domain memory	16×288GB=4.6TB (estimated)
Board bandwidth	8TB/s
NVLink	Fifth generation · aggregate switching bandwidth at the 14.4TB/s class
Cooling	Air-cooled / Liquid-cooled dual versions
Positioning	Same source as GB300 NVL72 compute half / building block toward hundred-GPU clusters
Market	No snapshot · Search spot listings in the hall › (/hall.html?model=B300) (prices subject to the seller's quote)

Originally published on AIXX Pedia, the AI compute hardware encyclopedia maintained by the AIXX exchange team. Specs, pricing context and verified-inventory insights are updated there first.

H100 NVL

aixx — Sun, 01 Nov 2026 09:31:49 +0000

Overview

H100 NVL It is the China-market variant of the NVIDIA H100 PCIe dual-slot Air-cooled form factor : 94GB HBM3 memory per card, 3,341 TFLOPS FP8, 60 TFLOPS FP64 Tensor, configurable 350-400W TDP, NVLink bridging at 600GB/s (2/4-way) — no HGX baseboard system required and plugs directly into a standard x86 server (official spec snapshot basis [1]).

How it relates to the HGX H100 baseboard version (SXM): same die, different form factor — the NVL version is a PCIe dual-slot Air-cooled card, while the HGX baseboard version is an SXM Liquid-cooled/Air-cooled system (8 GPUs fully interconnected via NVLink). For buyers: PCIe deployment has a lower barrier — no baseboard system needed; plugs straight into standard x86 servers. The 94GB memory is a 17.5% increase over the H100 PCIe's 80GB.

Comparison with the H200 NVL: memory 94GB vs 141GB (-33%), bandwidth 3.9TB/s vs 4.8TB/s, FP8 identical at 3,341 TFLOPS — the H200 NVL is the direct memory upgrade of the H100 NVL (official comparison basis [1]). This entry carries no market pricing — actual price subject to the seller's quote.

Use Cases

AIXX Index (0–100): Estimated in this Pedia via weighted public specs — memory 40% + interconnect 30% + ecosystem 30%.

History

In 2023 , H100 NVL launch: the Hopper-generation PCIe inference flagship — 94GB HBM3 memory, a 17.5% uplift over the H100 PCIe's 80GB, with FP8 Tensor Core at 3,341 TFLOPS in the same class (official spec snapshot basis [1]).

2024–2025 , inference at scale: explosive demand for online LLM inference — plug-and-play PCIe plus 94GB of memory made the H100 NVL one of the volume mainstays on the inference side (per this database's compilation).

2025-2026 , Blackwell takes over: RTX PRO 6000 Blackwell and the B200/B300 series launch — the H100 NVL moves into stock and parallel circulation (per this database's compilation). Cross-border trade policy is still evolving; this database reflects public reporting and does not constitute legal advice.

Technical Specifications

Item	Specs (H100 NVL · official figures)
Architecture	NVIDIA Hopper
Board Memory	94GB HBM3
Memory Bandwidth	3.9TB/s
FP8 / INT8	3,341 TFLOPS / 3,341 TOPS
FP16	1,671 TFLOPS
FP64 / FP64 Tensor	30 / 60 TFLOPS
FP32	60 TFLOPS
Interconnect	NVLink 600GB/s
Codecs	7 NVDEC
Power	350-400W (configurable)
Form Factor	PCIe dual-slot Air-cooled
MIG	Up to 7 instances @ 12GB

Specifications per NVIDIA official snapshot (October 2026 basis [1], verified line-by-line against the Spec Card).

Three form factors of the H100 family

Item	H100 NVL（PCIe）	HGX H100 8-GPU (SXM)	H200 NVL（PCIe）
Memory	94GB HBM3	80GB HBM3 × 8	141GB HBM3e
Interconnect	NVLink bridging at 600GB/s	NVLink 900GB/s all-to-all	NVLink bridging at 900GB/s
FP8	3,341 TFLOPS	3,358 TFLOPS	3,341 TFLOPS
Power	350-400W	700W (configurable)	600W (configurable)
One-line buying advice	Inference / single-server direct install	Large-scale training clusters	High-memory inference upgrade

Family Members

The H100 family is tiered by form factor; the NVL version is the PCIe inference flagship:

Cooling & Deployment

The H100 NVL is a PCIe dual-slot Air-cooled card (350-400W configurable) — plugs straight into standard x86 servers, no HGX baseboard system required (official spec snapshot basis [1]). Deployment quick-reference: Power (the 400W tier needs an adequate PSU and 16-pin CEM5 power cables), PCIe Gen5 direct attach (x16 direct to CPU, avoid switches that eat bandwidth), airflow (a 400W Full Load dual-slot card puts more airflow pressure than the 300W class — watch intake temperatures).

Multi-GPU deployment: 2-GPU NVLink bridging delivers 600GB/s of interconnect — inference throughput scaling doesn't require an HGX baseboard. 4-GPU bridging scales further but needs a dedicated bridge kit. Density: 2-4 GPUs per 2U server is common (airflow pressure from 400W×4=1.6kW needs careful planning).

Model Compatibility

Model / Scenario	Memory requirements (FP16 estimates)	H100 NVL deployment scenarios	AIXX Index
70B-class inference (INT8)	~70GB	94GB per card with ample headroom + MIG partitioning for multi-tenant	88
13B-class inference (FP16)	~26GB	Single card easily + MIG partitioning for multiple instances	90
Long-context inference	KV cache scales linearly with length	94GB offers more KV cache room than the 80GB bridge	84
Large model training	—	PCIe interconnect limitation — for large-scale training, look at the HGX SXM series	62

Gallery

Software & Ecosystem

Item	Requirements / Status
CUDA	Hopper generation (Compute Capability 9.0) — supported since CUDA 11.8
Frameworks	Full support across PyTorch / TensorFlow / vLLM / TensorRT
MIG	Up to 7 instances @ 12GB
Driver	NVIDIA data center driver line (shared with the HGX H100)

Purchasing & Inspection

H100 NVL procurement inspection points (three items):1Memory capacity verification94GB HBM3 — verify with nvidia-smi; the price gap between 94GB and 80GB is large, and mislabeling is a common risk.2MIG partitioning testTest the 7-instance partitioning — the core acceptance item for multi-tenant inference.3Codec testTest concurrent 7×NVDEC transcoding — the core acceptance metric for video workloads.Acceptance Quick Reference QUICK CHECKH100 NVL four-step quick check:1Identitynvidia-smi: H100 NVL model, 94GB memory, ECC status — match each item against the contract.2Full-load Burn-in Test400W tier Full Load: stable temperatures, zero ECC errors (per the case-law framework).3FP8 computeFP8 benchmark test (≥80% counts as Pass, estimated figures) — the real throughput anchor for inference production lines.4MIG partitioningTest the 7-instance partitioning — the core acceptance item for multi-tenant scenarios.Shipping & Delivery LOGISTICS1Cross-border complianceThe H100 NVL's export-control status toward China mirrors the H100 — consult professional counsel for cross-border procurement compliance; this database does not constitute legal advice.Glossary Cards GLOSSARYNVL namingIn NVIDIA's product lineup, the NVL suffix = a PCIe card with NVLink bridging — the H100 NVL achieves 600GB/s GPU-to-GPU Interconnect via 2/4-way NVLink bridging (not the SXM baseboard all-to-all).

H100 NVL memory specs: 94GB HBM3 (up 17.5% from the H100 PCIe's 80GB HBM2e), 3.9TB/s bandwidth — a significant lift of the memory ceiling for LLM inference in a PCIe form factor.

NVLink Bridge: physical bridging hardware between 2 or 4 GPUs — bypassing PCIe Switches, connecting card-to-card directly over NVLink. The H100 NVL's bridge bandwidth is 600GB/s (versus 900GB/s all-to-all on the HGX baseboard).

Limitations & Controversies

PCIe interconnect ceiling : NVLink bridging is capped at 2/4 GPUs (600GB/s); interconnect density for large-scale training is far below the HGX SXM baseboard (per this database's compilation).
FP64 is not the focus : 30 TFLOPS FP64 — higher than the A100 (19.5 TFLOPS FP64) for double-precision HPC, but not Hopper's headline (per public reporting).
Training vs inference positioning : PCIe form factor + bridged interconnect — for large-scale training look to the HGX SXM series; the H100 NVL's sweet spot is inference and compliance.
Planning for configurable power : 350-400W configurable — the power tier places high demands on facility power and cooling flexibility; spell out the tier in the purchase contract.
Cross-border trade policy variables : The H100 NVL's export-control status toward China mirrors the H100 — consult professional counsel for cross-border procurement compliance; this database does not constitute legal advice.

Market Notes

H100 NVL spot supply runs in parallel: brand-new OEM units and pulls from decommissioned systems — AIXX Pedia honesty principle: no market band; prices are subject to the seller's quote 。

Spot search: AIXX Exchange Hall › (/hall.html?model=H100%20NVL) (Zero-commission Matching, hence neutral data).

Price milestone timeline

FAQ

The dividing line is deployment form factor: PCIe direct install (standard x86 tower/servers) → NVL; HGX baseboard systems (8-GPU NVLink all-to-all) suit large-scale training clusters. 94GB vs 80GB is a memory upgrade.

A direct memory upgrade: 141GB vs 94GB (+50%), 4.8TB/s vs 3.9TB/s bandwidth, FP8 in the same tier — the H200 NVL offers better cost per GB of memory for inference.

PCIe interconnect limitation (600GB/s bridging vs HGX 900GB/s all-to-all) — for large-scale training look at the HGX SXM series; the H100 NVL's sweet spot is inference and compliance.

Four steps: identity verification (nvidia-smi 94GB/ECC) → 400W-tier Full Load Burn-in Test (temperature/zero ECC errors) → FP8 compute benchmark (≥80% Pass, estimated figures) → MIG 7-instance partitioning test.

The H100 NVL's export-control status toward China mirrors the H100 — under licensing control since 2022-10. No new-card sales through legitimate mainland channels; circulation falls under stock and grey-market. Assess compliance risk up front; this database does not constitute legal advice.

Related Entries

Model comparison and selection

References

NVIDIA Official — H100 NVL Spec Snapshot (verified line-by-line against our Spec Card, as of October 2026)　https://www.nvidia.com/en-us/data-center/h100/ (https://www.nvidia.com/en-us/data-center/h100/)
NVIDIA Official Website — H100 Product Line (family comparison)　https://www.nvidia.com/en-us/data-center/h100/ (https://www.nvidia.com/en-us/data-center/h100/)
Listing Search — AIXX Exchange (Zero-commission Matching; no market snapshot available for the H100 NVL yet, actual price subject to the seller's quote)　https://www.aixx.vip/hall.html?model=H100%20NVL (/hall.html?model=H100%20NVL)

Video Overview

Vendor	NVIDIA
Architecture	Hopper
Memory	94GB HBM3
Bandwidth	3.9TB/s
FP8 / INT8	3,341 TFLOPS / TOPS
FP16	1,671 TFLOPS
FP64 / FP64 Tensor	30 / 60 TFLOPS
Interconnect	NVLink 600GB/s
MIG	Up to 7 instances @ 12GB
Encode/decode	7 NVDEC
Power	350-400W (configurable)
Form Factor	PCIe dual-slot Air-cooled
Market	No snapshot · Search spot listings in the hall › (/hall.html?model=H100%20NVL) (prices subject to the seller's quote)

Originally published on AIXX Pedia, the AI compute hardware encyclopedia maintained by the AIXX exchange team. Specs, pricing context and verified-inventory insights are updated there first.

GB200 NVL72

aixx — Sun, 01 Nov 2026 09:29:51 +0000

Overview

GB200 NVL72 is NVIDIA's Rack-scale Superchip System : a single rack packs 72 Blackwell GPUs + 36 Grace CPUs with full liquid cooling; the 72 GPUs form a single fully-connected NVLink domain via fifth-generation NVLink (1.8TB/s per GPU, 30.4TB/s aggregate), delivering up to 1.44 EF FP4 of inference/training compute (per the official product page and Supermicro system documentation[1][2]).

It pushed the definition of "a server" to rack scale: compute trays (4×B200 + 2×Grace per tray) and NVLink switch trays (9 per rack) are fully connected over copper interconnects within the rack with no blocking, so training and inference run as if on one giant GPU (as announced at GTC, March 2024[3]). It was the workhorse rack of global AI infrastructure buildout in 2025 — leading the NVL72 family in deployment volume, discussion, and expected secondhand value.

On the 2026 spot market, GB200 NVL72 circulates in two forms: full racks and tray-level disassembly . Full-rack project delivery dominates, though compute trays and switch trays also trade at tray level. Procurement is heavy engineering work: liquid-cooling primary/secondary loops, power, rigging, and the software stack (NVLink domain health) must each be verified. This entry carries no market band — rack prices are actual price subject to the seller's quote.

Use Cases

AIXX Index (0–100): weighted from public specs — compute 40% + interconnect 30% + deployment barrier 30%. 90+ flagship tier · 75+ Excellent · 60+ usable · below 60 not applicable.

History

March 2024 , NVIDIA unveiled the Blackwell platform and the GB200 NVL72 : Built on the Grace Blackwell superchip (2×B200 + 1×Grace), with rack-scale NVLink full interconnect — the "72 GPUs as one GPU" narrative rewrote AI infrastructure from that point on (per NVIDIA's official page and GTC announcements[1][3]).

2024–2025 , the at-scale deployment phase: first racks went live at major North American cloud providers, with global shipments ramping through 2025 — liquid-cooling adoption climbed rapidly alongside NVL72 family volume (industry view: AI data center liquid-cooling penetration estimated above 30% in 2025, per TrendForce public reports[4]). Across the supply chain, cold plates, CDUs, manifolds, and quick connectors all scaled in step (see the AIXX Pedia entry on Liquid-cooled CDUs for details).

2025 , generational handoff: the GB300 NVL72 (Blackwell Ultra, 72×B300) launched and began shipping, moving GB200 NVL72 from new product to a phase where installed-base deployments and secondhand expectations run in parallel — both rack generations circulate in the 2026 market (see the GB300 entry).

2026 , spot positioning: GB200 NVL72 enters the circulation market as full racks or tray-level units. Cross-border policy for rack-scale equipment is still evolving; verify with regulators and professional counsel at procurement time (public reporting; not legal advice).

Technical Specifications

Item	Specifications (GB200 NVL72 · official figures)
GPU	72× NVIDIA Blackwell (B200 superchip dies)
CPU	36× NVIDIA Grace (72-core Armv9)
GPU Memory	Up to 13.4TB HBM3E (per rack total)
CPU Memory	Up to 17TB LPDDR5X (ECC)
Compute	Up to 1.44 EF FP4 (per rack, official figures)
Interconnect	NVLink 5.0 · 1.8TB/s per GPU · 72-GPU single domain · 130TB/s aggregate (official page)
Compute Trays	4×B200 + 2×Grace per tray, 18 trays per rack (per Supermicro system documentation[2])
Switch Trays	NVLink switch trays × 9 per rack (copper, fully connected, non-blocking)
Cooling	Rack-level liquid cooling (cold plates + CDU loop, see the Liquid-cooled CDU entry)
Launch	March 2024 (GTC, Blackwell platform)

Specifications per NVIDIA's official product page (as of October 2026)[1]; total system power and power-delivery details follow the system vendor's (ODM/OEM) configuration. Anything not listed here = not verified — no guessing.

NVL72 family, two generations side by side

Item	GB200 NVL72	GB300 NVL72
GPU	72× B200（Blackwell）	72× B300（Blackwell Ultra）
Board Memory	192GB HBM3E class	288GB HBM3E
Networking	ConnectX-7 generation	ConnectX-8 generation (800G)
Positioning	Installed-base workhorse	Current new-rack mainstay
One-line buying advice	Hunting value in the installed base	Flagship for new purchases

See the GB300 NVL72 entry for the GB300 line[1]; the two rack generations share the same liquid-cooling architecture lineage (for CDU return-loop details, see the Liquid-cooled CDU entry).

Family Members

GB-series racks are tiered by generation and form factor; GB200 NVL72 is the Blackwell-generation rack. Family members:

Cooling & Deployment

GB200 NVL72 is a fully liquid-cooled rack : cold plates cover every GPU/CPU/NVLink switch die, and the in-rack secondary loop exchanges heat with the facility's primary cooling water via a CDU (Coolant Distribution Unit) — the liquid-cooling loop is the first deployment barrier: primary-side supply temperature/flow/water quality, secondary-side quick disconnects and manifolds, and CDU N+1 redundancy must be verified item by item against the design (see the Liquid-cooled CDU entry). Retrofit cost and facility conditions are the real barriers to NVL72 adoption — air-cooled server rooms cannot take these racks directly.

Power and rigging are the other two barriers: rack power density far exceeds traditional air-cooled racks (have the OEM/ODM configuration re-check power and busbar capacity), and rack weight and transport rigging require advance planning for floor loading and routing. On the software side, NVLink domain health (NVLink switch trays + copper cable backplane) is the key acceptance item — a single bad copper cable can degrade the 72-GPU interconnect domain (Pedia editorial notes).

Use-Case Fit

Workload	GB200 NVL72 performance profile	AIXX Index
Trillion-Parameter MoE Training	72-GPU single domain + 13.4TB HBM3E per rack, expert parallelism without cross-rack hops	95
Large-model online inference	FP4 1.44 EF + in-domain memory sharing over NVLink, flagship batch concurrency throughput	93
Video/multimodal generation	FP4/FP8 inference pipelines saturate the NVLink domain bandwidth	88
Traditional FP64 HPC	Not Blackwell's home turf — double-precision-centric workloads should look at the Hopper generation or CPU clusters	70

Workload tiers are Pedia editorial notes based on public specs (estimates; not a performance commitment).

Gallery

Actual photos are added continuously as inspection batches are stocked; AI-generated images are never passed off as real photos.

Software & Ecosystem

Item	Requirements / Status
Base Software	NVIDIA AI Enterprise / Base Command family (cluster-level scheduling and monitoring)
Collective Communication	NCCL auto-detects the NVLink domain; full in-rack interconnect works with zero configuration
Cross-Rack Networking	Quantum-2/X800 (InfiniBand) or Spectrum-X (Ethernet) per fabric route, with ConnectX-7/8 NICs
Storage	GPUDirect Storage pass-through (see the NVMe All-Flash Storage entry)
Operations	Liquid-cooling loop inspections + NVLink domain health checks are the two daily routines
System Vendor	Full-rack NVL72 solutions from Supermicro, Dell, Lenovo, Inspur, and others (per Supermicro documentation[2])

Rack-scale systems upgrade "server ops" to "data-hall-level ops" — liquid cooling and the NVLink domain are two new specializations; software evolves over time, defer to NVIDIA's official releases.

Purchasing & Inspection

GB200 NVL72 procurement inspection checklist (three items):1End-to-end liquid-cooling loop acceptanceItemize and pressure-test CDUs, manifolds, quick disconnects, and cold plates — a leak anywhere is an incident under full load; align primary-side water quality and temperature with the facility first (see the Liquid-cooled CDU entry).2Full NVLink domain health checkThe 72-GPU single domain is the lifeline: nvidia-smi nvlink on all ports, switch tray firmware, and copper cable backplane checked one by one — one bad cable hurts the whole rack; the entire domain must pass before signing.3Tray-level asset verificationCount serials tray by tray across 18 compute trays, 9 switch trays, and power modules; for tray-level traded units, match tray condition against original rack configuration — mixed-tray assemblies must be stated in the contract.For the full six-test verdict framework see this entry "How to Stress-Test" (#benchtest) section.

How to Stress Test

Six-test methodology, all numeric verdicts anchored to GB200 NVL72: unboxing & identification → liquid-cooling loop → NVLink domain → compute baseline → storage pass-through → extended burn-in.

Shipping & Delivery

OEM & Variants

GB200 NVL72 is delivered primarily as full racks by NVIDIA-authorized OEMs:

System Vendor	Reference Solutions	Differentiators	Status as of 2026
Supermicro	SuperCluster GB200 NVL72 (per official documentation[2])	Tray configuration and delivery cadence	One of the mainstream full-rack delivery options
Dell	PowerEdge rack-level solution	Tier-one global service network	Full-rack delivery
Lenovo/Inspur and others	NVL72 full-rack solution	Regional delivery and service coverage	Project-based circulation
Rack/tray-split circulation	Compute tray / switch tray level	Mixed-tray assembly risk; inspect each tray individually	Spot secondary market

System configurations follow each vendor's official documentation[2].

Glossary Cards

Rack-scale systems named NVL (NVLink): compress the traditional cluster of multiple servers plus switches into a single rack, with 72 GPUs forming a single NVLink domain — from a programming perspective, it is one giant GPU card, and cross-GPU communication no longer goes through the network stack.

The compute unit of the GB200: 2 B200 GPUs and 1 Grace CPU (72-core Armv9) packaged via NVLink-C2C interconnect — a unified CPU-GPU memory coherence view, and the building block of the compute tray.

9 NVLink switch trays inside the rack: weave 72 GPUs into a single fully interconnected domain — copper backplane interconnect (no optical modules in-rack), the biggest architectural difference between NVL72 and traditional 8-GPU servers.

Two-stage Liquid-cooled architecture: the secondary loop (in-rack coolant, running through cold plates) exchanges heat with the primary loop (facility cooling/chilled water) via the CDU — the CDU is the interface device between the two loops and core supporting infrastructure for NVL72 deployment.

The secondary market form of rack-scale equipment: full racks are broken down and traded at tray level. For mixed-tray assembly (trays from different sources combined into one rack), focus on verifying NVLink domain integrity and tray condition consistency — spelling out configuration provenance in the contract is the baseline for self-protection.

Limitations & Controversies

Liquid cooling is a hard requirement Air-cooled facilities cannot host it directly — primary-side water conditions, CDU redundancy, and quick-disconnect workmanship; any failing link means a Full Load incident (per this Pedia's compilation; see the Liquid-cooled CDU entry for details).
Capital and facility barriers of full-rack-only sales The procurement granularity of NVL72 is the rack — budget, floor loading, power upgrades, and crane installation all come first. Smaller-scale needs should look at HGX baseboard systems (see the B300 NVL16 entry).
FP64 is not its home turf The Blackwell architecture focuses on FP4/FP8/BF16; traditional double-precision HPC users should weigh FP64 figures carefully when selecting (public architecture figures).
Second-hand Market Circulation is still in its early stages There is no mature yardstick for residual value assessment of rack-scale equipment — liquid-cooling component lifespan, NVLink copper cable condition, and tray mixing history are all variables. Inspection costs for a used NVL72 should be factored into the total price (per this Pedia's inspection experience).
Export-control-sensitive model The GB200 NVL72 is a key focus of export controls (per public reporting) — for cross-border circulation, always verify compliance status against professional advice first. This Pedia only notes policy developments and does not constitute legal advice.

Market Notes

The GB200 NVL72 circulates at two levels — full racks and split trays — and tray composition, liquid-cooling support, and facility conditions make single-rack prices highly dispersed. There is no verifiable unified market snapshot yet. AIXX Pedia honesty principle: no market band; prices are subject to the seller's quote 。

Spot search: AIXX Exchange Hall › (/hall.html?model=GB200) (Zero-commission Matching; no involvement in transactions; no knowledge of or speculation on transaction prices — hence data neutrality). This section will switch to a standard market band once enough snapshots accumulate.

Price milestone timeline

FAQ

For new purchases, consider budget and lead time: the GB300 (B300 288GB/card + CX-8 800G) is the flagship of new racks; the GB200 is the previous-generation rack with the largest installed base, with second-hand value depending on tray condition and liquid-cooling support. Both generations share the same liquid-cooling architecture, offering high facility condition reuse.

Three things: liquid cooling (primary-side water conditions + CDU deployment; air-cooled facilities cannot host it directly), power (full-rack power density far exceeds traditional racks; review busbars and transformers against the complete system configuration), and floor loading plus crane installation paths. On the software side, accept the NVLink domain and direct-attached storage per this entry's benchmarking criteria.

Tray-level circulation exists in the secondary market (18 compute + 9 switch trays can be split), but NVLink domain integrity depends on the rack-level backplane and switch trays — a single tray outside a rack has no NVLink domain value. Mixed-tray assemblies must pass full-domain stress testing before acceptance.

The dividing line is scale and budget: dozens of GPUs, air-cooled/hybrid facilities — HGX baseboard systems are more flexible; hundreds of GPUs, prioritizing single-domain interconnect efficiency, with liquid-cooling-ready facilities — NVL72. See each entry for a detailed comparison.

Two special inspections + one asset: full liquid cooling loop chain (CDU/manifold/quick connectors/cold plate pressure testing), full NVLink domain inspection (all ports Active + cross-tray bandwidth), tray-level asset verification (serial numbers and mixed-tray history). Complete the six stress tests in this entry before signing off.

Related Entries

Model comparison and selection

References

NVIDIA official site — GB200 NVL72 product page (primary spec reference, as of 2026-10)　https://www.nvidia.com/en-us/data-center/gb200-nvl72/ (https://www.nvidia.com/en-us/data-center/gb200-nvl72/)
Supermicro — SuperCluster GB200 NVL72 Datasheet (OEM configuration document: tray composition and interconnect figures)　https://www.supermicro.com/datasheet/datasheet_SuperCluster_… (https://www.supermicro.com/datasheet/datasheet_SuperCluster_GB200_NVL72.pdf)
HPE — GB200 NVL72 QuickSpecs (second OEM source)　https://www.hpe.com/us/en/collaterals/collateral.a50009224en… (https://www.hpe.com/us/en/collaterals/collateral.a50009224enw.html)
TrendForce — public report on AI data center liquid-cooling penetration (2025-08, industry figures)　https://www.trendforce.com/ (https://www.trendforce.com/)
Listings search — AIXX Exchange (Zero-commission Matching; no market snapshot for GB200 NVL72 yet, actual price subject to the seller's quote)　https://www.aixx.vip/hall.html?model=GB200 (/hall.html?model=GB200)

Video Overview

Vendor	NVIDIA (delivered by authorized OEMs)
Category	Rack-scale Superchip System
GPU	72× Blackwell B200
CPU	36× Grace (72-core Armv9)
GPU Memory	Up to 13.4TB HBM3E per rack
Compute	Up to 1.44 EF FP4 (full rack)
Interconnect	NVLink 5 · 72-GPU single domain · 130TB/s aggregate
Trays	18 compute + 9 NVLink switch
Cooling	Rack-scale Liquid-cooled (CDU loop)
Launch	March 2024 (GTC)
Market	No snapshot · Search spot listings in the hall › (/hall.html?model=GB200) (prices subject to the seller's quote)

Originally published on AIXX Pedia, the AI compute hardware encyclopedia maintained by the AIXX exchange team. Specs, pricing context and verified-inventory insights are updated there first.