AMD’s new chiplet isn’t just a 24-core complex, it’s much more. We’re talking about the Instinct MI300, here are all the details!
We mentioned it in previous articles, and it was the protagonist of the CES 2023it’s the new one AMD Instinct MI300, and complex chiplet in which inside there is a very elaborate architecture that will open new horizons on the Server side and beyond.
It is based on architecture Zen 4 and CDNA 3, “merged” together. A single chip which in turn on a single package delivers well 128 GB of memory type HBM3 and that we will be able to see in the second half of 2023. Below is a slide that summarizes the main characteristics.
AMD Instinct MI300: what the new 24-core Chiplet hides
The new processor will be destined for the world of data centers and supercomputers, as it is logical to imagine, but it comes from a combination of x86 cores, with GPU and memory on a single package. The new AMD Instinct MI300, as extensively during the presentation the CEO Lisa Su he explained, consists of a structure of nove chiplet a 5 nanometerimpilati, and made tramite stacking 3D above four chiplets, all made a 6 nanometers. In addition, we have side by side 128 GB of memory type HBM3, aspect that allows the chip to function without resorting to external DRAM memory. The new Instinct MI300 is based on 146 billion transistors and it’s AMD’s most complex product ever.
These were the words of Lisa Su, during the conference;
To put this into perspective, there has been a lot of talk over the holidays about ChatGPT and what can be done with these great language models. These models take months to train on thousands of GPUs that consume millions of dollars of electricity. MI300 can reduce training time for these models from months to weeks, with significantly lower energy costs. And most importantly, it can also support much, much larger models that can be used for even more advanced and powerful AI services in the future.
The beating heart of the structure of the new AMD Instinct MI300, composed of 24 x86 cores based on Zen 4 architecture with GPU based on CDNA 3 architecture, connected together (the structures) thanks to the help of the fourth generation of interconnection Infinity, at very high speed. The performance per watt should therefore be up to 5 times better, and the performance of AI calculation up to 8 times. Currently housed in the supercomputer Captainto offer beyond 2 exaflops of power, and will be installed at the Lawrence Livermore National Lab del DOE (US Department of Energy). If you want to keep learning about the latest news from the hardware world, keep following us. Greetings from TechGameWorld.com.
The article AMD Instinct MI300: what the new 24-core chiplet hides comes from TechGameWorld.com.