Perplexity 发布了一项研究，详细介绍了他们如何在 NVIDIA 的 GB200 NVL72 Blackwell 机架上部署大型语言模型，特别是 Qwen3 235B。研究结果表明，与之前的 NVIDIA 硬件相比，GB200 平台在大型模型推理方面提供了显著的改进，具有更低的延迟和更高的吞吐量。这项研究强调了 GB200 在训练和高吞吐量推理方面的能力，特别是对于专家混合（MoE）模型。 AI
影响 NVIDIA 的 GB200 Blackwell 平台在 LLM 推理速度和成本效益方面显示出显著的提升，可能加速大型模型的部署。
排序理由 该集群包含 Perplexity 发布的研究，内容涉及 LLM 推理硬件。
AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →
This NVIDIA remains the strongest platform for large-model inference at scale. Prefill/decode disaggregation, Blackwell-native quantization, custom kernels, and rack-scale NVLink turn GB200 into faster answers lower serving cost. Read the full paper here
The benchmarks show the gap. NVLS all-reduce latency drops from 586.1µs on H200 to 313.3µs on GB200. In MoE prefill at EP=4, combine falls from 730.1µs to 438.5µs. For decode, GB200 sustains much higher throughput at high token speeds.
Prefill and decode stress hardware differently. Prefill is compute-bound, so Blackwell Tensor Cores, memory bandwidth, NVLink, and SHARP reductions help. Decode is latency/memory-bound, where GB200’s rack-scale NVLink domain opens up parallelism Hopper could not.
We published new research on how we serve post-trained Qwen3 235B models on NVIDIA GB200 NVL72 Blackwell racks. GB200 is a major step up over Hopper for high-throughput inference on large MoE models, not just a training platform. https://t.co/yYZuPRXWzr