Nvidia ships Blackwell Ultra to cloud partners, AMD answers
Nvidia claims 2.4 times the inference throughput of the prior generation on large mixture-of-experts workloads. AMD says MI400 begins sampling next quarter.
Nvidia said Blackwell Ultra is now shipping to cloud partners, and put the gain at 2.4 times the inference throughput of the previous generation on large mixture-of-experts workloads. The figure is the company's own, measured on workloads it selected, and has not been independently verified.
AMD responded the same day, saying MI400 begins sampling next quarter with what it called competitive HBM bandwidth. It offered no throughput comparison of its own. Bloomberg's Emily Chang reported both claims together, treating the AMD statement as a direct answer to the Blackwell Ultra timing.
Sampling and shipping are different stages. Nvidia is describing hardware reaching partners now. AMD is describing parts going out for evaluation in a later quarter.