AI chips compared: H100, B200, Rubin, MI355X, TPU and Trainium accelerator specs

This table lists AI accelerator specs for data-center chips from Nvidia, AMD, Google, AWS, Intel, Cerebras, Huawei, Microsoft and Meta: Hopper H100 and H200, Blackwell B200 and B300, GB200 and GB300 NVL72 racks, the Rubin GPU and Vera Rubin NVL72, Instinct MI300X through MI455X, TPU Ironwood and TPU 8, Trainium2 and Trainium3, Gaudi 3, Cerebras WSE-3T, Ascend, Maia 200 and MTIA.

Every number comes from the vendor’s own datasheet, product page, technical documentation or announcement, linked on each entry with the date it was checked. We copy figures as the vendor states them. Vendors quote FP8 and FP4 throughput either with sparsity or without it (dense); each entry’s precision note says which, because the two can differ by a factor of two. Where a vendor publishes no figure for a field, the field is left blank rather than estimated.

Some vendors publish specs per chip, others per board, rack or pod. The form factor column says which level a row describes, so rack-scale rows (NVL72, CS-4, Groq 3 LPX) are not directly comparable with single chips. Memory is in GB, bandwidth in TB/s, FP8 in TFLOPS and FP4 in PFLOPS.

The table is checked against vendor sources three times a week. For launch coverage see AI chip news, for cloud rental rates per GPU-hour see GPU prices, and for where these chips are being deployed see AI data centers. How we source data is described in our editorial standards.