NVIDIA GB300 NVL72 vs GB200 NVL72

Both are rack-scale NVL72 systems — 72 GPUs and 36 Grace CPUs wired into a single fifth-generation NVLink domain — but a generation apart. The GB200 NVL72 is built on the original Grace Blackwell (GB200); the GB300 NVL72 is the newer Grace Blackwell Ultra (GB300).

The defining difference is memory. Each GB300 GPU carries 288 GB of HBM3e versus 192 GB on the GB200 — so a full GB300 rack holds roughly 20.7 TB of HBM3e against about 13.8 TB on the GB200. NVLink, GPU count, and the ConnectX-8 scale-out plane are the same across both. Because the largest models are increasingly memory-bound, the market has rotated to the higher-memory GB300, and that is the rack-scale system we supply.

How the two rack-scale generations line up. The headline move is memory — 288 GB vs 192 GB per GPU (≈20.7 TB vs ≈13.8 TB per rack) — on top of an otherwise-shared NVL72 platform: 72 GPUs, one fifth-generation NVLink domain, and ConnectX-8 scale-out.

The GB300 is a memory refresh of the rack-scale platform rather than a new fabric. Grace Blackwell Ultra raises on-package HBM3e from 192 GB to 288 GB per GPU — a 50% step — so the pooled memory of a single NVL72 rack grows from roughly 13.8 TB to ≈20.7 TB of HBM3e.

Why it matters: a rack-scale NVL72 behaves like one very large accelerator, and its ceiling is set by how much model and long-context KV cache stay resident across the NVLink domain. More HBM3e per GPU lets a larger model shard across the same 72 GPUs with less spillover, which is exactly the constraint the frontier of AI serving keeps running into. The NVLink generation (fifth-gen, 1.8 TB/s per GPU) and the ConnectX-8 scale-out plane (800 Gb/s per GPU) are common to both — the GB300’s advantage is memory, not fabric.

On compute the two are nearly matched — both deliver ~5 PFLOPS of dense FP8 per GPU — but Blackwell Ultra lifts dense FP4 from ~10 to ~15 PFLOPS per GPU, so a full 72-GPU rack reaches roughly ~1.1 EFLOPS of dense FP4 against about ~0.72 EFLOPS on the GB200 — the throughput lever for large-scale inference. Figures are dense; sparsity doubles them.

For a new rack-scale buildout, the GB300 NVL72 is the current generation and the one to deploy — more HBM3e per GPU, more pooled memory per rack, and the platform NVIDIA and its integrators are shipping now. The GB200 NVL72 was the prior rack-scale flagship; the market has rotated to the GB300’s higher memory, and that is what we carry — factory-integrated from Supermicro, Lenovo, HPE, and Pegatron.

For the full rack composition — 18 compute nodes of 4 GPUs, 36 Grace CPUs, the in-rack CDU, and scale-out fabric — see the GB300 NVL72 reference architecture. Pricing across every configuration is quoted per configuration on request — tell us the cluster you are standing up and we will return pricing and availability. Browse the GPU catalog to compare integrations.