AMD announced that it is working on the AMD Instinct MI430X graphics processing unit for AI supercomputing.
At the AMD AI event at the Moscone Center in San Francisco, AMD revealed the GPU as its most advanced high-performance computing (HPC) AI accelerator.
AMD said its HPC FP64 compute is 288 teraflops, or 8.7 times the rival Nvidia Vera Rubin, with HBM4 capacity of 432 gigabytes, or 50% better than Vera Rubin. The HBM4 bandwidth is 23.3 terabytes per second, or 6% better than Vera Rubin.
On the low end, AMD also announced the MI350P, an air-colled GPU in a power envelope of 450 to 600 watts TBP. It can address up to 260 billion model parameters at FP4, said AMD’s senior vice president Vamsi Boppana, during the AMD AI keynote event.
The Anthropic deal is specifically around the MI430X project, Su said.