Tag
Blackwell Ultra
Blackwell Ultra is NVIDIA’s high-end inference platform built on the Blackwell architecture, centered on B300 and GB300 NVL72 systems with larger HBM3e capacity, higher bandwidth, and rack-scale scaling. It matters for LLM inference, KV cache sizing, cloud cost, and datacenter deployment choices.
2 articles
Industry News/Apr 3
NVIDIA B300 vs H200: Specs and DeepSeek Perf
B300 packs 288GB HBM3e and up to 8TB/s bandwidth. Here’s how it compares with H200 for DeepSeek inference and cloud costs.
Industry News/Apr 2
NVIDIA Sets New MLPerf Inference Records
Blackwell Ultra hit new MLPerf Inference v6.0 highs, with GB300 NVL72 gaining 2.7x on DeepSeek-R1 server tests and 1.5x on Llama 3.1 405B.