NVIDIA Blackwell InferenceMAX Benchmarks: Performance & Efficiency

Here’s a summary of the MAX v1 benchmarks and performance​ improvements ⁤highlighted⁣ in the provided text:

* Throughput: Blackwell delivers over 10,000 TPS (tokens per second) per GPU at 50 ⁢TPS per user interactivity. This is 4x higher per-GPU ⁣throughput compared to​ the NVIDIA H200 GPU.
* Power Efficiency: Blackwell provides 10x throughput per megawatt compared to the previous generation.
*​ Cost efficiency: ​ The cost per million tokens is 15x lower with the Blackwell architecture compared to the previous generation.
* Performance Balance: Blackwell balances cost, energy efficiency, throughput, and responsiveness, offering the highest ROI across real-world ‍workloads. It utilizes the Pareto ​frontier to map performance across these dimensions.
* InferenceMAX: Uses the Pareto frontier to​ map performance, reflecting ​Blackwell’s ⁣ability to balance ‍production priorities.

The ‌text emphasizes that Blackwell isn’t just ‍about peak performance in ‌one area,but about delivering efficient and cost-effective performance across a range of production needs.