The time has finally come: AMD has delivered on its long-awaited promise and the Exascale-class APU, the Instinct MI300A, is finally going into series production. Series production will begin this quarter and the APU is expected to be available in 2024 as the world’s fastest HPC solution. It has been a long time coming and people have been eager to experience the power of this new technology.
The AMD Instinct MI300A APU combines different architectures and interconnect technologies such as Zen 4, CDNA 3 and the latest generation of the Infinity architecture. The MI300A APUs offer a number of highlights.
Up to 61 TFLOPS FP64 computation
Up to 122 TFLOPS FP32 calculation
Up to 128 GB HBM3 memory
Up to 5.3 TB/s storage bandwidth
146 billion transistors
The MI300A is very similar to the MI300X, but with the difference that it uses memory- and Zen-4-optimized cores. Now let’s look at the details of this exascale performance for next-generation HPC and AI data centers.
An active chip has removed two CDNA 3 GCDs and replaced them with three Zen 4 CCDs, each of which has its own cache pools and core IPs. There are now a total of 24 cores and 48 threads on the chip, divided into 8 cores and 16 threads per CCD. In addition, there is a separate cache pool per CCD with a size of 32 MB and an L2 cache of 24 MB (1 MB per core). It should be noted that the CDNA 3 GCDs also had a separate L2 cache.
AMD has activated a total of 228 compute units on the GPU side, which are based on the CDNA 3 architecture. This corresponds to 14,592 cores, which means that there are 38 compute units per GPU chiplet. Here are some of the outstanding features of the AMD Instinct MI300 Accelerators summarized:
First integrated CPU GPU package
Target exascale supercomputer market
AMD MI300A (integrated CPU GPU)
146 billion transistors
Up to 24 Zen 4-cores
CDNA 3 GPU architecture 228 compute units (14,592 cores)
Up to 128 GB HBM3 memory
Up to 8 chiplets 8 memory stacks (5nm 6nm process)
AMD has again compared the MI300A with the H100, but this time in HPC-specific workloads. In terms of performance figures, the Instinct MI300A APU in OpenFOAM was able to achieve up to a 4-fold increase in performance. This is mainly due to the unified memory layout, GPU performance and overall available memory capacity and bandwidth. Compared to NVIDIA’s Grace Hopper superchips, the system also offers up to 2x performance per watt.
It has been confirmed that the Instinct MI300A APUs are now shipping and will also be used to power the upcoming El Capitan supercomputer. This is expected to offer up to 2 exaflops of computing power. It is worth noting that AMD is the only company to break the 1 exaflop barrier so far with the Frontier supercomputer and also has the most efficient system in the world.
Source: AMD
15 Antworten
Kommentar
Lade neue Kommentare
Urgestein
1
Veteran
Urgestein
Urgestein
Veteran
Urgestein
Urgestein
Urgestein
Veteran
Urgestein
Urgestein
Urgestein
Veteran
Alle Kommentare lesen unter igor´sLAB Community →