AMD unveils Instinct MI300 apu with 153 billion transistors for AI language models
AMD announces several new processors and GPUs for AI data centers and servers, including the 153 billion transistor AMD Instinct MI300X accelerator. The brand also announces EPYC Bergamo CPUs based on the Zen 4c architecture.
AMD MI300
The AMD Instinct MI300 series consists of two HPC accelerators. The MI300 APU that AMD previously announced is now called MI300A. This chip has nine CDNA 3 GPU chiplets and also contains 24 Zen 4 CPU cores. The accelerator therefore has both GPU and CPU chiplets, for a total of thirteen chiplets and 146 billion transistors in one package. The MI300A has up to 128GB of shared HBM3 memory and, according to AMD, is made for AI workloads.
The MI300X is a completely new model in the MI300 series and consists only of GPU cores. AMD thus replaces the three CPU chiplets of the MI300A APU below with two additional CDNA 3 chiplets; The APU consists of twelve chiplets in total. According to the brand, the accelerator is intended for running generative AI language models. The accelerator contains up to 192GB of HBM3 memory with a bandwidth of 5.2TB/s.
The brand offers a ready-made composition of MI300 chips in the form of the AMD Instinct Platform for data centers. This system contains eight MI300 chips and a total of up to 1.5TB of HBM3 memory. AMD has started shipping MI300A samples to select customers. From the third quarter of this year, this will also happen with MI300X chips.
EPYC Bergamo with Zen 4c cores
Furthermore, at the same event, AMD will announce new EPYC server processors based on Zen 4c cores, where the ‘c’ stands for ‘cloud’. These cores are more compact than regular Zen 4 cores, which are already in the current EPYC Genoa processors. Instead of a maximum of twelve CCDs with eight cores each, AMD now uses eight CCDs with sixteen cores. This should make the new processors more energy efficient.
This concerns three variants as part of the Bergamo generation of CPUs. The top model is the EPYC 9754 with 128 cores and 256 threads. The processor has a TDP of 360W. AMD promises base and boost clock frequencies of 2.25 and 3.10GHz respectively. The 9754S has largely the same specifications, but does not support simultaneous multithreading and therefore has a maximum of 128 threads. The chips are said to be up to 2.7 times as efficient as the previous generation of EPYC processors, AMD claims during its presentation.
AMD’s entry-level Bergamo variant, the EPYC 9734, also has 112 cores and 224 threads. The base and boost clock frequencies are almost the same, at 2.20 and 3.00GHz respectively. The TDP of that server processor is 320W.