"AMD's largest chip in history" explosion site CES: 146 billion transistors, which can greatly reduce ChatGPT training time

Yuyang Alex Posted from Aofei Temple
Qubit | Public account QbitAI

A chip contains 146 billion transistors.

It also claims to be able to shorten the training time of large models such as ChatGPT and DALL·E from months to weeks , saving millions of dollars in electricity bills.

Just at CES 2023, the technology Spring Festival Gala, Su Ma brought AMD’s “largest chip to date” to blow up the scene.

This chip, codenamed Instinct MI300 , is also AMD’s first data center/HPC-level APU.

Parameter performance

What is the concept of 146 billion transistors?

Previously, Intel's server GPU Ponte Vecchio had more than 100 billion transistors, while Nvidia's new nuclear bomb H100 integrated 80 billion transistors.

However, those who are familiar with AMD know that APU is simply a CPU and GPU packaged together. From this perspective, it seems reasonable that the number of transistors far exceeds that of competitors (manual dog head).

Specifically, this chip with 146 billion transistors adopts a chiplet design .

Based on 3D stacking technology, nine 5nm computing chips (CPU+GPU) are stacked on four 6nm process chiplets.

According to information disclosed by AMD, the CPU uses the Zen 4 architecture, including 24 cores. The GPU uses AMD's CDNA 3 architecture.

Since the Zen 4 architecture is usually an 8-core design, it is generally speculated that 3 of the 9 computing chips are CPUs and 6 are GPUs.

In addition, MI300 has 8 HBM3 video memories with a total of 128GB.

In terms of performance, AMD did not release much information and only compared it with the Instinct MI250X.

Su Ma said that compared with MI250X, MI300’s AI performance per watt has been improved by 5 times, and the overall AI training performance has been improved by 8 times.

According to Tom's Hardware, AMD also revealed that MI300 can shorten the training time of large models such as ChatGPT and DALL·E from months to weeks.

Instinct MI300 is expected to be delivered in the second half of 2023. By then, the chip will also be deployed in two new ExaFLOP supercomputers .

One More Thing

Also at this year's CES, AMD also directly benchmarked Apple's M-series chips and launched the "world's fastest ultra-thin processor" - Ryzen 7040 series.

Specific models include: Ryzen 5 7640HS, Ryzen 7 7840HS, Ryzen 9 7940HS.

As for the top-end R9 7940HS, Su praised it highly:

In terms of multi-threaded performance, the R9 7940HS is 34% faster than the Apple M1 Pro ; in AI task processing, it is 20% faster than the Apple M2 .

It also provides more intuitive experience data: an ultra-thin notebook equipped with Ryzen 7040 series chips can play videos continuously for more than 30 hours.

The first batch of notebook computers equipped with Ryzen 7040 processor will be shipped in March this year.

In any case, this news made the melon-eaters extremely happy:

Apple's decision to develop its own PC chips has intensified competition among chip manufacturers. This situation should have happened 10 years ago.

参考链接：
[1]https://www.youtube.com/watch?v=OMxU4BDIm4M
[2]https://www.tomshardware.com/news/amd-instinct-mi300-data-center-apu-pictured-up-close-15-chiplets-146-billion-transistors
[3]https://www.anandtech.com/show/18721/ces-2023-amd-instinct-mi300-data-center-apu-silicon-in-hand-146b-transistors-shipping-h223
[4] https://www.macrumors.com/2023/01/05/amd-new-chips-against-m1-pro/

click here