Delivering Blackwell's 30X performance gain required integration and optimization across six or seven chips.

Judge as of 2025-09-26. The claim is that Blackwell's '30X' performance gain (Nvidia's claimed up-to-30x LLM inference improvement of GB200 NVL72 over H100) depends on co-designing and integrating six or seven distinct chip types (for example Blackwell GPU, Grace CPU, NVLink switch chip, ConnectX NIC, BlueField DPU, Spectrum/Quantum switch). True: Nvidia technical documentation or credible analysis shows that the 30X figure is measured on a full rack-scale system combining about six or seven distinct chip types, and that it depends on that system integration (such as NVLink scale-up and the full software stack) rather than on the GPU die alone. Mostly true: the gain is system-level and depends on several chips, but the count is somewhat different (for example four or five chip types), or part of the gain comes from software or numerical precision (such as FP4) rather than chip integration. Mostly false: the 30X figure comes mostly from a single factor (such as lower-precision FP4 or a favourable benchmark configuration), so multi-chip integration is a minor contributor. False: the 30X figure does not depend on multi-chip integration, or Nvidia never claimed such a gain.

Set before the research began.

1. E1 NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference | NVIDIA Technical Blog NVIDIA (developer blog) · 18 March 2024 Own record

   “It delivers a 30x speedup for resource-intensive applications like the 1.8T parameter GPT-MoE compared to the previous H100 generation. This advancement is made possible with a new generation of Tensor Cores, which introduce FP4 precision and the many advantages that come with fifth-generation NVLink”

2. E2 NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference | NVIDIA Technical Blog NVIDIA (developer blog) · 18 March 2024 Own record

   “Results based on token-to-token latency = 50 ms; real-time, first token latency = 5,000 ms; input sequence length = 32,768; output sequence length = 1,024 output, 9x eight-way HGX H100 air-cooled: 400 GB IB Network vs 18 GB200 Superchip liquid-cooled: NVL72, per GPU performance comparison.”

3. E3 NVIDIA NVLink and NVIDIA NVSwitch Supercharge Large Language Model Inference | NVIDIA Technical Blog NVIDIA (developer blog) · 12 August 2024 Own record

   “The NVIDIA GB200 NVL72 system connects 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell GPUs in a rack-scale design, and with the fifth-generation NVLink, enables all 72 GPUs to act as a single GPU, enabling 30x faster real-time trillion-parameter inference compared to the prior generation.”

4. E4 Demystifying AI Inference Deployments for Trillion Parameter Large Language Models | NVIDIA Technical Blog NVIDIA (developer blog) · 12 June 2024 Own record

   “More specifically, NVIDIA Blackwell can deliver 30x more throughput at reading speeds of 20 tokens per user per second (5–6 words per second) using TP2EP16PP2 and a chunk size of 896 tokens.”

NVIDIA's technical blogs describe the 30x gain over H100 as a result for the whole rack-scale system. It is measured on a GB200 NVL72 rack of 36 Grace CPUs and 72 Blackwell GPUs linked by fifth-generation NVLink, compared with nine HGX H100 systems connected by InfiniBand (E2, E3). The same sources also credit new Tensor Cores with FP4 precision for the gain (E1), and the figure was measured under a specific parallelism and chunk-size configuration (E4), so part of the gain comes from numerical format and may also depend on software configuration. The evidence names about three chip types (Grace CPU, Blackwell GPU, NVLink) and does not confirm a count of six or seven. Under the criteria this makes the claim mostly true rather than true.