Perplexity details running Qwen3 235B on NVIDIA Blackwell
By Steven Van ·
The research covers disaggregated prefill/decode and tensor parallelism, and finds Blackwell a major gain over Hopper for large MoE inference.
Perplexity has published research on running Qwen3 235B, a large mixture-of-experts model, on NVIDIA's GB200 NVL72 racks. The writeup covers disaggregating prefill and decode, tensor parallelism, and NVLink throughput, and finds Blackwell a major step up over Hopper for high-throughput inference on large MoE models.
Perplexity
AI-powered answer engine — ask anything and get cited, real-time answers from across the web.
View Perplexity →