A chip stuffed with 146 billion transistors, it also claims to reduce the training time of large models such as ChatGPT and DALL‧E from months to weeks, saving millions of dollars in electricity bills. Right at CES 2023, Su Ma brought AMD’s “biggest chip to date” to the crowd.
This chip, codenamed Instinct MI300, is also AMD’s first data center / HPC-class APU.
parameter performance
What is the concept of 146 billion transistors?
Previously, the Intel GPU Ponte Vecchio server had 100 billion + transistors, while the new nuclear bomb Nvidia H100 integrates 80 billion transistors.
However, anyone familiar with AMD knows that an APU is just a package of CPU and GPU. From this point of view, it seems reasonable to much more than the number of transistors of competitors.
Specifically, this chip with 146 billion transistors uses a small chip (Chiplet) design. Based on 3D stacking technology, 9 5nm computer chips (CPU + GPU) are stacked on top of 4 small 6nm process chips.
According to AMD’s public information, the CPU uses the Zen 4 architecture, including 24 cores. The GPU uses AMD’s CDNA 3 architecture. Since the Zen 4 architecture is typically an 8-core design, it is widely speculated that 3 of the 9 computing chips are CPUs and 6 are GPUs. In addition, MI300 has 8 HBM3 VRAM which amounts to 128GB.
In terms of performance, AMD did not publish much information, only in comparison to the Instinct MI250X.
Su Ma said that compared to the MI250X, the AI performance per watt of the MI300 has increased 5 times, and the overall AI training performance has increased 8 times.
According to Tom’s Hardwre, AMD also revealed that MI300 can shorten the training time of large models such as ChatGPT and DALL‧E from months to weeks.
The Instinct MI300 is expected to be introduced in the second half of 2023. At that time, the chip will also be used on two new exaFLOP supercomputers.
One more thing
Also at CES this year, AMD used Apple’s M-series chips directly as a benchmark to launch “the world’s fastest ultra-thin processor” – the Ryzen 7040 series.
Specific models include: Ryzen 5 7640HS, Ryzen 7 7840HS, Ryzen 9 7940HS.
For the top-end R9 7940HS, Su Ma praised:
R9 7940HS is 34% faster than Apple M1 Pro in terms of multi-threaded performance; 20% faster than Apple M2 in processing AI tasks.
A more intuitive experience data is also given: ultra-thin notebook computers equipped with Ryzen 7040 series chips can play 30+ hours of videos continuously.
The first batch of notebook computers equipped with Ryzen 7040 processors will ship in March this year.
source:
