This is all for generative AI
AMD introduced the Instinct MI300 line of accelerators. The model of the same name was first introduced to us back in January, but now AMD has expanded the line and revealed new details.
AMD itself calls Instinct MI300 an accelerator, although in general it can also be called a hybrid processor. True, with certain reservations, since these monsters have little in common with consumer APUs. Actually, the company also uses the term APU.
So now only AMD can.
It is worth noting that there are two models in the line, and only one of them is essentially a hybrid processor, since the second one is devoid of a CPU, that is, it is a GPU-based accelerator.
So, the accelerator, which in January was simply called MI300, is now called Instinct MI300A. It is he who is of the greatest interest from the point of view of technical implementation, since there is nothing similar on the market.
This is a gigantic APU that includes 24 Zen 4 cores, a CDNA 3 GPU, and 128GB of HBM3 memory. AMD does not disclose the GPU configuration, but WWCF Tech claims that we are talking about 228 CU units. If their configuration has not changed relative to MI200, then we are talking about 14,592 stream processors.
AMD uses 5nm and 6nm process technology to manufacture various chiplets. Such an APU contains 146 billion transistors. This is more than the Apple M2 Ultra.
The Instinct MI300X has an even larger transistor budget – 153 billion transistors. Apparently, this is the largest and most complex GPU in the world. There’s also 192 GB (HBM3) with a throughput of 5.2 TB/s! The bandwidth of the Infinity Fabric bus reaches 896 GB / s.
For the Instinct MI300X, it also doesn’t reveal the number of CUs, but the same source says 304, which is 19,546 stream processors.
The AMD Instinct MI300 line of APUs will be compatible with a brand new socket known as SH5, which will allow up to four of these chips to be stacked together using Infinity Fabric.
It is on the basis of Instinct MI300 that the El Capitan supercomputer with a performance of 2 exaflops will be built. AMD also notes that the Instinct MI300X is the world’s most advanced generative AI accelerator. In particular, thanks to the huge amount of memory, AMD customers can install large language models, such as the Falcon-40, with 40 billion parameters on a single accelerator.
The Instinct MI300A APUs will be available immediately after the presentation, while the Instinct MI300X will be available next quarter.
As a reminder, today AMD also introduced 128-core Epyc Bergamo and Epyc Genoa-X processors with more than 1 GB of cache .
