AMD CEO hints on Instinct MI400
AMD has recently unveiled the Instinct MI300 series, a combination of CDNA3 and Zen4 architectures. Among the series, the MI300X stands out as a GPU-only accelerator boasting an 192GB of HBM3 memory. This particular model is specifically designed to cater to large language models (LLMs) that demand substantial memory capacity and bandwidth.
One of the key advantages of the MI300 series is its versatility, made possible by utilizing die stacking technology. The MI300X shares the chiplet architecture with the MI300A, a data center APU that integrates Zen 4 CPU cores and CDNA3 GPU cores within a single package. The MI300A is geared towards HPC and AI workloads, benefiting from a unified memory architecture and featuring 128GB of HBM3 memory.
AMD and its partners have already hinted at the development of the successor to the MI300 series, referred to as the MI400, which will be a high-end data-center accelerator series possibly based on CDNA4 architecture. Though the specific SKUs are yet to be announced, AMD CEO Dr. Lisa Su has confirmed the company’s plans for the upcoming MI400 series during the recent Q2 earnings call, without providing further details.
When you look across those workloads and the investments that we’re making, not just today, but going forward with our next generation MI400 series and so on and so forth, we definitely believe that we have a very competitive and capable hardware roadmap. I think the discussion about AMD, frankly, has always been about the software roadmap, and we do see a bit of a change here on the software side.
— AMD CEO Dr Lisa Su
The MI400 will be the fifth generation of Instinct accelerators and will continue the departure from the Radeon branding, which became even more apparent with the MI300 and the integration of the Zen CPU architecture into the series. It will follow the path set by the Arcturus based MI200 series (already deployed and commercially available) and the upcoming MI300 (codenamed Aldebaran).
Rumors surrounding the MI400 suggest that it might incorporate a new XSwitch interconnect technology, which could potentially position AMD in a more competitive position against NVIDIA NVLink based products.
| AMD Instinct Accelerators Specifications | |||||||
|---|---|---|---|---|---|---|---|
| VideoCardz | Year | Node | SKU | GPU CU Cores | CPU Cores | Memory (HBM) | TDP |
| 2024+ | CDNA4/Zen5 (?) | TBC | TBC | TBC | TBC | TBC | |
| 2023 | 5/6nm CDNA3/Zen4 | MI300A | TBC | ||||
| MI300X | – | ||||||
| MI300C | – | TBC | |||||
| MI300P | – | TBC | |||||
| 2022 | 6nm CDNA2 | MI250X | – | ||||
| MI250 | – | ||||||
| MI210 | – | ||||||
| 2020 | 7nm CDNA1 | MI100 | – | ||||
| 2018 | 7nm VEGA20 | MI60 | – | ||||
Source: SeekingAlpha, SemiAnalysis
Many thanks to Element115 for the tip!
