Alongside AMD’s widely expected client product announcements this evening for desktop CPUs, mobile CPUs, and mobile GPUs, AMD’s CEO Dr. Lisa Su also had a surprise up her sleeve for the large crowd gathered for her prime CES keynote: a sneak peak at MI300, AMD’s next-generation data center APU that is currently under development. With silicon literally in hand, the quick teaser laid out the basic specifications of the part, along with reiterating AMD’s intentions of taking leadership in the HPC market.
First unveiled by AMD during their 2022 Financial Analyst Day back in June of 2022, MI300 is AMD’s first shot at building a true data center/HPC-class APU, combining the best of AMD’s CPU and GPU technologies. As was laid out at the time, MI300 would be a disaggregated design, using multiple chiplets built on TSMC’s 5nm process, and using 3D die stacking to place them over a base die, all of which in turn will be paired with on-package HBM memory to maximize AMD’s available memory bandwidth.
AMD for its part is no stranger to combining the abilities of its CPUs and GPUs – one only needs to look at their laptop CPUs/APUs – but to date they’ve never done so on a large scale. AMD’s current best-in-class HPC hardware is to combine the discrete AMD Instinct MI250X (a GPU-only product) with AMD’s EPYC CPUs, which is exactly what’s been done for the Frontier supercomputer and other HPC projects. MI300, in turn, is the next step in the process, bringing the two processor types together on to a single package, and not just wiring them up in an MCM fashion, but going the full chiplet route with TSV stacked dies to enable extremely high bandwidth connections between the various parts.
Previous Instinct products, like the MI200, were GPU-based resources only, rather than the new hybrid design shown here. AMD says that using the hybrid CPU-and-GPU chiplet approach not only saves energy for a given compute load, but also makes programming the chip easier than before. AMD says this is the world’s first datacenter processor integrating CPU and GPU resources with all of this memory bandwidth on hand into a single package.
When it launches in the later half of 2023, AMD’s MI300 is expected to be going up against a few competing products. The most notable of which is likely NVIDIA’s Grace Hopper superchip, which combines an NVIDIA Armv9 Grace CPU with a Hopper GPU. NVIDIA has not gone for quite the same level of integration as AMD is, which arguably makes MI300 a more ambitious project, though NVIDIA’s decision to maintain a split memory pool is not without merit (e.g. capacity). Meanwhile, AMD’[s schedule would have them coming in well ahead of arch rival Intel’s Falcon Shores XPU, which isn’t due until 2024.
Expect to hear a great deal more from AMD about Instinct MI300 in the coming months, as the company will be eager to show off their most ambitious processor to date.
The key advantage of AMD’s design, besides the operational simplicity of putting CPU cores and GPU cores on the same design, is that it will allow both processor types to share a high-speed, low-latency unified memory space. This would make it fast and easy to pass data between the CPU and GPU cores, letting each handle the aspects of computing that they do best. As well, it would significantly simplify HPC programming at a socket level by giving both processor types direct access to the same memory pool – not just a unified virtual memory space with copies to hide the physical differences, but a truly shared and physically unified memory space.
The 3D design allows for incredible data throughput between the CPU, GPU and memory dies while also allowing the CPU and GPU to work on the same data in memory simultaneously (zero-copy), which saves power, boosts performance, and simplifies programming. It will be interesting to see if this device can be used without standard DRAM, as we see with Intel’s Xeon Max CPUs that also employ on-package HBM.
AMD’s representatives were coy with details, so it isn’t clear if AMD uses a standard TSV approach to fuse the upper and lower dies together, or if it uses a more advanced hybrid bonding approach. We’re told AMD will share more details about the packaging soon.