NVIDIA GB200 NVL72
Overview
The NVIDIA GB200 NVL72 is a rack-scale system: 72 Blackwell chips and 36 Grace CPUs joined by fifth-generation NVLink into a single 72-chip domain with 900 GB/s of unidirectional bandwidth per chip. There is no scale-out network inside the rack because the rack itself is the compute unit, with roughly 13.4 TB of pooled HBM3e.
That one-hop, all-to-all fabric is why GB200 NVL72 dominates wide-expert-parallel serving of large MoE models: expert layers can be spread across dozens of chips without crossing InfiniBand or Ethernet. Disaggregated prefill/decode stacks such as Dynamo and SGLang PD map naturally onto the domain.
Specifications
| Vendor | NVIDIA |
|---|---|
| Architecture | Blackwell |
| Memory (usable) | 186 GB HBM3e |
| Memory bandwidth | 8 TB/s |
| FP4 dense TFLOP/s | 10,000 |
| FP8 dense TFLOP/s | 5,000 |
| BF16 dense TFLOP/s | 2,500 |
| Scale-up interconnect | NVLink 5.0 |
| Scale-up bandwidth per chip | 900 GB/s |
| Scale-up world size | 72 |
| Scale-up topology | Switched 18-rail Optimized |
| Scale-out network | N/A (NVLink domain only) |
| NIC | N/A (NVLink domain only) |
| TDP per chip | 1,200 W |
| All-in power per chip | 1.87 kW |
| Hyperscaler $/chip/hr | $1.86 |
| Neocloud $/chip/hr | $2.26 |
| Retail $/chip/hr | $2.60 |
Source: $/chip/hr rate tiers from the SemiAnalysis AI Cloud TCO Model
How InferenceX benchmarks it
InferenceX benchmarks GB200 NVL72 with multi-node disaggregated configurations, including wide-EP vLLM, SGLang and Dynamo TRT-LLM runs on frontier MoE models, and normalizes results as throughput per chip so rack-scale numbers stay comparable with 8-chip HGX nodes.
Frequently asked questions
- How much does GB200 NVL72 cost per hour in the cloud?
- The SemiAnalysis AI Cloud TCO model rates GB200 NVL72 at about $1.86/hr at hyperscalers, $2.26/hr at neoclouds and $2.60/hr at the retail tier. InferenceX performance-per-dollar pages use these rates to turn measured throughput into $/M tokens.
- How much memory does GB200 NVL72 have?
- GB200 NVL72 has 186 GB of usable HBM3e per chip with 8 TB/s of memory bandwidth. A 72-chip NVLink 5.0 domain pools 13,392 GB.
- What is the power consumption of GB200 NVL72?
- GB200 NVL72 has a 1,200 W TDP per chip, and about 1.87 kW all-in per chip once the host CPU, NICs and cooling share are included. InferenceX uses the all-in figure for energy-per-token math.
- Does GB200 NVL72 support FP4?
- Yes. GB200 NVL72 reaches 10,000 dense FP4 TFLOP/s (5,000 at FP8), and InferenceX tracks FP4-versus-FP8 serving accuracy and throughput on its precision compare pages.
- How fast is GB200 NVL72 for LLM inference?
- It depends on the model, framework, precision and interactivity target, so InferenceX publishes continuously refreshed throughput-versus-interactivity Pareto frontiers for GB200 NVL72 instead of a single number. The live dashboard and compare pages show current results on every covered model.
See live benchmark results
Every number above is static hardware data. Delivered tokens per second, cost per million tokens and energy per token are measured continuously on the dashboard:
Go deeper with the SemiAnalysis models
InferenceX measures delivered inference performance. The SemiAnalysis institutional models cover the market behind these chips: who ships them, who buys them, and what they cost to own.
SemiAnalysis Accelerator & HBM Model
SKU-level AI accelerator shipments, pricing and specifications, from foundry wafer starts and HBM supply through customer-level installed base, quarterly with multi-year forecasts.
SemiAnalysis AI Cloud TCO Model
The source of the hourly rates on this page: all-in GPU cost of ownership built up from server capex, power, colocation and cost of capital, with rental price scenarios and a full cluster finance suite.