AMD announced that its AI accelerator, the Instinct MI500, expected in 2027, will offer 1000 times the performance of the MI300X. This new line of accelerators will rely on CDNA 6 architecture and be manufactured by a 2 nm process. Unfortunately, AMD did not specify how the new solution can offer a 1000-fold performance increase over the MI300X. It’s worth mentioning that this isn’t the most modern accelerator, yet it was launched only at the end of 2023.
The MI300X is based on the CDNA 3 architecture, has 19,456 stream processors, 1,216 Matrix Core engines, and FP8 performance of either 2.61 or 5.22 PFLOPS (in normal and Structured Sparsity modes). FP4 performance isn’t indicated because it’s unsupported, and AMD might compare it based on these metrics.
The demand for computational power is growing faster than ever. Meeting this demand requires continuous advances in performance, significantly surpassing the current level. The MI400 marked a turning point in providing leadership learning in all fields, including inference and scientific computing. We keep moving forward. The development of our next-generation MI500 series is in full swing. With the MI500, we take another significant step in performance enhancement. It’s built on our next-generation CDNA 6 architecture, manufactured by a 2-nm process, and uses high-speed HBM4E memory.
How much memory the accelerator will have, AMD has not yet stated, but even the MI450 will have 432 GB of it. Incidentally, this accelerator will show a performance of 40 PFLOPS in FP4 mode.
