Chips

GB300 vs GB200 NVL72: Specs and Verdict

GB300 NVL72 vs GB200 NVL72 compared: 20 TB vs 13.4 TB HBM3E, 1.5x dense FP4, 2x attention, ConnectX-8 networking, and which rack to pick for your workload.

Two towering liquid-cooled AI server racks side by side, one glowing amber and one teal, with light streaming between them, illustrating NVIDIA GB300 NVL72 versus GB200 NVL72.
Illustration: Edgewisely

TL;DR

  • GB300 NVL72 (Blackwell Ultra) carries 20 TB of HBM3E (up to 288 GB per GPU) vs the GB200 NVL72's 13.4 TB (~192 GB per GPU) — about 1.5x the memory.
  • GB300 delivers 1.5x the dense FP4 compute (1,080 vs 720 PFLOPS) and 2x attention-layer performance per GPU.
  • GB300 doubles per-GPU networking to 800 Gb/s via the ConnectX-8 SuperNIC (14.4 TB/s per rack, 2x GB200); the rack is otherwise identical — 72 GPUs, 36 Grace CPUs, 130 TB/s NVLink, liquid-cooled.
  • Verdict: pick GB300 for large reasoning models and long-context inference; GB200 is the cheaper, more widely available workhorse for training and standard inference.

GB300 vs GB200: the short answer

The GB300 NVL72 is NVIDIA's Blackwell Ultra rack; the GB200 NVL72 is the standard Blackwell rack. Both pack 72 GPUs and 36 Grace CPUs into one liquid-cooled, NVLink-connected system. The GB300 adds roughly 50% more HBM3E memory (20 TB vs 13.4 TB), 1.5x the dense FP4 compute, 2x attention performance, and double the per-GPU network bandwidth. The rack itself is the same size and shape.

What is the difference between the GB300 and GB200 NVL72?

The difference is the GPU inside, not the rack around it. Both systems are NVLink-72 designs: 72 GPUs acting as one large accelerator, with 36 Grace CPUs (2,592 Arm Neoverse V2 cores) and 130 TB/s of fifth-gen NVLink bandwidth. Liquid cooling, the tray layout, and the 72:36 GPU-to-CPU ratio are identical. If you're weighing the individual building-block GPUs rather than the full rack, our GB200 vs B200 breakdown covers that layer.

What changes:

  • Memory. Per NVIDIA's GB300 NVL72 page, the Blackwell Ultra GPUs hold up to 288 GB of HBM3E each, for 20 TB per rack. The GB200 NVL72's Blackwell GPUs hold ~192 GB each, for 13.4 TB. That extra memory is the headline reason to care.
  • Compute. GB300's Tensor Cores deliver 1.5x more dense FP4 FLOPS than Blackwell and 2x attention-layer acceleration, per NVIDIA.
  • Networking. GB300 integrates the ConnectX-8 SuperNIC at 800 Gb/s per GPU over PCIe Gen6, which NVIDIA's Blackwell Ultra technical blog puts at 14.4 TB/s of rack networking — 2x the GB200 NVL72.

GB300 vs GB200 NVL72: the specs that matter

Spec (per rack) GB300 NVL72 (Blackwell Ultra) GB200 NVL72 (Blackwell)
GPUs 72 Blackwell Ultra 72 Blackwell
Grace CPUs 36 (2,592 Neoverse V2 cores) 36 (2,592 Neoverse V2 cores)
GPU memory (HBM3E) 20 TB (up to 288 GB/GPU) 13.4 TB (~192 GB/GPU)
GPU memory bandwidth Up to 576 TB/s 576 TB/s
FP4 Tensor Core (sparse / dense) 1,440 / 1,080 PFLOPS 1,440 / 720 PFLOPS
FP8/FP6 Tensor Core (sparse) 720 PFLOPS 720 PFLOPS
NVLink (5th-gen) 130 TB/s 130 TB/s
Network bandwidth (per rack) 14.4 TB/s (ConnectX-8, 800 Gb/s/GPU) 7.2 TB/s (Quantum-X800 / BlueField-3)
CPU memory 17 TB LPDDR5X @ 14 TB/s 17 TB LPDDR5X @ 14 TB/s
Cooling Liquid, rack-scale Liquid, rack-scale
Per-GPU TDP Higher (not officially disclosed) Not officially disclosed

Note the FP4 line: both racks hit the same 1,440 PFLOPS sparse peak, but GB300's dense FP4 is 1.5x higher (1,080 vs 720). Dense is the number that matters for most real inference and training, where sparsity doesn't apply.

Is the GB300 NVL72 faster than the GB200?

Yes, for the workloads it was built for — but not uniformly. On dense FP4 compute GB300 is 1.5x faster, and on attention layers (the bottleneck in long-context and reasoning models) NVIDIA claims 2x. FP8, FP16, and sparse FP4 peaks are the same on both racks. So GB300's advantage shows up most in large reasoning-model inference and long sequences, less so in workloads that don't lean on attention or memory capacity.

The bigger practical win is memory. The extra 6.6 TB of HBM3E per rack lets GB300 hold larger models and longer KV caches in memory, which is what raises tokens-per-second-per-user and overall throughput per rack for reasoning workloads. It is the same capacity-versus-compute split you see one tier down between the H200 and B200.

When is the GB300 NVL72 available?

Blackwell Ultra was unveiled in March 2025, and NVIDIA states the products were expected from partners in the second half of 2025. The GB300 NVL72 product page now lists it as available, ordered through NVIDIA sales and shipped via OEM and cloud partners. GB200 NVL72 shipped first and is the more broadly deployed of the two. NVIDIA does not publish list pricing for either rack; both go through OEM and cloud channels, so real cost depends on the partner and contract.

Is the GB300 worth it over the GB200?

It depends on whether your bottleneck is memory. GB300's case rests almost entirely on capacity and attention throughput. If you serve large reasoning models, long-context inference, or mixture-of-experts models that strain HBM, the 20 TB of memory and 2x attention performance translate directly into more users per rack and longer contexts. If your models fit comfortably in GB200's 13.4 TB and you're compute- rather than memory-bound, GB200 gives you most of the platform for less, with wider availability today.

Power is the trade-off. Blackwell Ultra draws more per GPU than Blackwell — NVIDIA does not publish the exact per-GPU TDP, but the generational step is upward, so factor denser power delivery and cooling into any 2026 build.

Who should pick which?

  • Training frontier models: either rack works; GB200 is the cost-efficient default. Choose GB300 if your training runs are memory-bound or lean heavily on attention. For the prior generation's trade-offs, see A100 vs H100.
  • Large-scale reasoning / long-context inference: GB300. The 288 GB-per-GPU memory and 2x attention are the entire reason it exists.
  • Standard inference at scale: GB200. Same sparse FP4 peak, lower cost, available now.
  • Buying cloud capacity: rent GB200 unless your SLA needs the longest contexts or largest models — then pay up for GB300 instances.
  • Planning a 2026 data-center build: spec for GB300 power and cooling headroom even if you deploy GB200 first; both share the NVLink-72 topology, and GB300 is where NVIDIA's roadmap is heading.

Frequently Asked Questions

What is the difference between GB300 and GB200?

The GB300 NVL72 uses Blackwell Ultra GPUs; the GB200 NVL72 uses standard Blackwell GPUs. Both racks hold 72 GPUs and 36 Grace CPUs. GB300 adds about 50% more HBM3E memory (20 TB vs 13.4 TB), 1.5x the dense FP4 compute, 2x attention performance, and 800 Gb/s per-GPU networking via ConnectX-8.

Is the GB300 NVL72 faster than the GB200 NVL72?

For memory-bound and attention-heavy work, yes. GB300 delivers 1.5x more dense FP4 compute (1,080 vs 720 PFLOPS) and 2x attention-layer performance per GPU. Sparse FP4, FP8, and FP16 peaks are identical on both racks, so compute-bound workloads that don't stress attention or memory see a smaller gain.

How much more memory does the GB300 NVL72 have?

About 1.5x. The GB300 NVL72 carries 20 TB of HBM3E (up to 288 GB per GPU) versus the GB200 NVL72's 13.4 TB (roughly 192 GB per GPU). That extra 6.6 TB per rack is what lets GB300 serve larger models and longer-context inference with bigger KV caches in memory.

When is the GB300 NVL72 available?

NVIDIA unveiled Blackwell Ultra in March 2025 and said products would be available from partners in the second half of 2025. NVIDIA's GB300 NVL72 page now lists it as available, ordered through sales and shipped via OEM and cloud partners. GB200 NVL72 shipped earlier and is more widely deployed.

Should you buy GB300 or GB200?

Buy GB300 if you run large reasoning models or long-context inference, where its 20 TB of memory and 2x attention performance pay off directly. Buy GB200 for standard training and inference that fits in 13.4 TB — it costs less, is more widely available, and matches GB300 on sparse FP4, FP8, and FP16 peak compute.

Editor's note — sources:

  • NVIDIA GB300 NVL72 product page (https://www.nvidia.com/en-us/data-center/gb300-nvl72/) — GB300 specs: 20 TB HBM3E, FP4 1,440 sparse / 1,080 dense PFLOPS, 130 TB/s NVLink, ConnectX-8 SuperNIC at 800 Gb/s per GPU, 72-GPU / 36-Grace configuration.
  • NVIDIA GB200 NVL72 product page (https://www.nvidia.com/en-us/data-center/gb200-nvl72/) — GB200 specs: 13.4 TB HBM3E, FP4 1,440 sparse / 720 dense PFLOPS, 130 TB/s NVLink, Quantum-X800 InfiniBand / Spectrum-X / BlueField-3.
  • NVIDIA Developer Blog, "NVIDIA Blackwell Ultra for the Era of AI Reasoning" (https://developer.nvidia.com/blog/nvidia-blackwell-ultra-for-the-era-of-ai-reasoning) — up to 288 GB HBM3E per GPU, 20 TB per rack (1.5x GB200), 14.4 TB/s rack networking (2x GB200), ConnectX-8 800G SuperNIC over PCIe Gen6, availability from partners in H2 2025.

Get Edgewisely in your inbox

Business stories that matter, free. Enter your email — no password, no account to set up.
jamie@example.com
Subscribe