Blackwell, Decoded
The Systems · Module 6

GB300 NVL72 Rack

The flagship reasoning system9 min readModule 6 of 8

Everything so far — the dual-die GPU, the engines, NVLink — comes together in NVIDIA’s flagship reasoning machine. The GB300 NVL72 is a single liquid-cooled rack that links 36 Grace CPUs and 72 Blackwell Ultra GPUs into one NVLink domain that behaves as one colossal GPU.

72
Blackwell Ultra GPUs
36
Grace CPUs
>1
exaFLOP of FP4 compute
37 TB
fast memory per rack

#What’s in the rack

The rack holds 18 compute trays and 9 NVLink Switch trays. Each compute node packs 4 Blackwell GPUs and 2 Grace CPUs. Every GPU carries 279 GB of HBM3e and delivers 15 petaFLOPS of dense FP4 — and the whole rack is liquid-cooled to pack that density in.

Table 2 — GB300 NVL72 vs GB200 NVL72 (per rack)
SpecGB300 NVL72 (Ultra)GB200 NVL72
GPUs | CPUs72 | 3672 | 36
Compute / NVLink-switch trays18 / 918 / 9
FP4 Tensor Core (dense / sparse)1,080 / 1,440 PF720 / 1,440 PF
FP8 / FP6 (dense / sparse)360 / 720 PF360 / 720 PF
INT8 (dense / sparse)12 / 24 POPS360 / 720 POPS
FP16 / BF16 (dense / sparse)180 / 360 PF180 / 360 PF
FP64 Tensor Core100 TF2,880 TF
HBM3e memory20 TB13.5 TB
HBM bandwidth576 TB/sup to 576 TB/s
NVLink bandwidth130 TB/s130 TB/s
CPU cores2,592 Arm V22,592 Arm V2
Read the table like an engineer

Notice Blackwell Ultra trades away INT8 and FP64 throughput (100 TF FP64 vs 2,880) to pour its silicon into FP4 — 1,080 vs 720 dense petaFLOPS. It’s a chip deliberately specialized for low-precision reasoning inference, not double-precision science.

#The headline gains vs Hopper

Lower cost / token25×Energy efficiency30×Reasoning inference35×AI factory output50×
Each bar is a different metric, all measured against Hopper-generation systems. (NVIDIA projections, subject to change.)

#The AI factory, and how Dynamo earns 50×

NVIDIA frames an AI factory as a trade-off between two things: total throughput per megawatt (how much the factory produces) and per-user speed (how fast each person’s answer comes). You can’t max both — the best operating points form a Pareto frontier curve, and factory output is roughly the area under it.

NVIDIA Dynamo is a real-time orchestrator that dynamically repartitions GPU, memory and NVLink resources, letting the system glide along that curve instead of being stuck at one fixed point. By tuning parallelism strategies across Blackwell Ultra GPUs, Dynamo reaches 50× the AI-factory output of Hopper while keeping latency low.

output ≈ area Throughput / MW Per-user speed → Hopper frontier Blackwell Ultra + Dynamo optimal operating point
Dynamo keeps the factory at the corner of the curve, maximizing output without sacrificing responsiveness.

#Networking: wiring racks into a data center

One GB300 NVL72 is already a single huge computer; AI factories chain many together. Each GPU gets 800 Gb/s of throughput via a ConnectX-8 SuperNIC, over either Quantum-X800 InfiniBand or Spectrum-X Ethernet, with BlueField-3 DPUs. GB300 is the first to use PCIe Gen6 between GPU and SuperNIC — eliminating a separate PCIe switch — and the SuperNIC encrypts GPU-to-GPU traffic at line rate.

#Bonus: real-time video generation

Video is staggeringly token-heavy: a single five-second clip is about 4 million tokens (versus a 128,000-token LLM context window) and takes ~90 seconds on Hopper. Blackwell Ultra delivers a 30× speedup, enabling real-time generation from NVIDIA Cosmos world foundation models for Physical AI.

One more comparison

With 65× more AI FLOPS than an HGX H100 system, the GB300 NVL72 enables dramatically more inference per rack — the reason NVIDIA pitches it as a strategic necessity, not just an upgrade.

Continue →