NVIDIA GB300 Superchip

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

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

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

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

OEM & Variants

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

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

FAQ

Related Entries

References

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.

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.