Tonight, in addition to AMD’s highly anticipated client product announcements for desktop CPUs, mobile CPUs, and mobile GPUs, AMD CEO Dr. It wowed the audience. AMD’s next-generation data center APU currently under development. With silicon literally in hand, a quick teaser laid out the basic specs for the parts along with reiterating AMD’s intent to take leadership in the HPC market.
First unveiled by AMD at the 2022 Financial Analyst Day in June 2022, the MI300 is AMD’s first attempt to combine AMD’s CPU and GPU technologies to build a true data center/HPC class APU. . In its layout at the time, the MI300 was a fragmented design, using multiple chiplets built on TSMC’s 5nm process, using 3D die stacking to place them on the base die, and stacking them all together. Combine with on-package HBM. Maximize available memory bandwidth for AMD.
AMD is no stranger to combining the power of its CPUs and GPUs, but it’s enough to look at CPU/APU in laptops, but they’ve never done it on a large scale. AMD’s current best-in-class HPC hardware is a combination of the discrete AMD Instinct MI250X (a GPU-only offering) and AMD’s EPYC CPU, which is exactly what Frontier supercomputers and other his HPC That’s what the project did. MI300 then is the next step in the process, combining two processor types into one package and not only wiring them in an MCM fashion, but also doing a full chiplet route using TSV stacked dies. , allowing very high performance. Bandwidth connections between various parts.
A key part of tonight’s announcement was to show off the MI300 silicon. It has reached early production and is currently being launched in AMD’s labs. AMD has previously promised a 2023 launch for the MI300, and getting silicon back from the fab and assembling it is a strong sign that AMD is well on its way to that delivery date.
In addition to the chance to see the Titanic chip in person (or at least in a video stream), Dr. Su’s short teaser also provided some new and intriguing details about the hardware. With 146 billion transistors, his MI300 is the largest and most complex chip AMD has ever built. Only comparable to current chip designs, this is significantly more transistors than Intel’s 100B transistor Xeon Max GPU (Ponte Vecchio) or NVIDIA’s 80B transistor GH100 GPU. In fairness to both, AMD packs both his GPU and CPU into this piece.
On the CPU side of the MI300, we’ve been confirmed to use 24 of AMD’s Zen 4 CPU cores, which finally gave us a basic idea of what to expect in terms of CPU throughput. The GPU side, on the other hand, uses an (yet) undisclosed number of CDNA 3 architecture CUs. All of this is paired with 128 GB of HBM3 memory.
According to AMD, the MI300 will consist of nine 5nm chiplets, sitting on top of four 6nm chiplets. The 5nm chiplets are definitely compute logic chiplets (i.e. CPU and GPU chiplets), but we can sort out exactly what is not available. A reasonable guess at this point would be 3 CPU chiplets (each with 8 Zen 4 cores) combined with 6 GPU chiplets. However, there are some cache chiplets that have not yet been described. On the other hand, if we take AMD’s ‘on top of’ statement literally, the 6nm chiplet will be the base die on top of all of these. Based on AMD’s rendering, it looks like 8 HBM3 memory stacks are being used. This means that the memory bandwidth is around 5TB/s.
Regarding performance expectations, AMD isn’t saying anything new at this point. Previous claims were more than a 5x improvement in AI performance per watt compared to the MI250X and an overall 8x improvement in AI training performance compared to his, which AMD said at the time of his CES There are also things I claim.
The main advantage of AMD’s design is the operational simplicity of putting CPU and GPU cores in the same design, as well as allowing both processor types to share a fast, low-latency unified memory space. This makes it fast and easy to move data between CPU and GPU cores, allowing each to handle the aspects of computing it does best. Similarly, it greatly simplifies HPC programming at the socket level by allowing both processor types to directly access the same memory pool. This is not just a unified virtual memory space with copies to hide physical differences, but a truly shared, physically unified memory space. .
AMD’s MI300, which launches in the second half of 2023, is expected to take on some of the competition. It’s a super tip. NVIDIA hasn’t done exactly the same level of integration as AMD so perhaps he MI300 would be a more ambitious project but there are merits (capacity etc.) to his NVIDIA decision to keep the split memory pool It’s not that there isn’t one. On the other hand, AMD’s schedule would put it much earlier than its biggest rival, Intel’s Falcon Shores XPU, which isn’t due until 2024.[のスケジュールでは、2024年までに予定されていない最大のライバルであるIntelのFalconShoresXPUよりもはるかに早く登場することになります[sschedulewouldhavethemcominginwellaheadofarchrivalIntel’sFalconShoresXPUwhichisn’tdueuntil2024
The company is keen to show off its most ambitious processor to date, so expect to hear more from AMD about the Instinct MI300 in the coming months.