AMD introduced the Instinct MI325X AI accelerator to compete with Nvidia Blackwell and talked about the even more powerful Instinct MI355X

AMD has officially unveiled its flagship computing accelerator Instinct MI325X, which will compete with Nvidia Blackwell and has already entered production. At the same time, the manufacturer revealed details about the next generation accelerator – Instinct MI355X based on CDNA4 architecture.

Image source: AMD

Regardless, CDNA4 is a new architecture that represents a significant upgrade from the previous CDNA3. AMD described it as “reimagined from the ground up,” although experts say that may be a bit of an exaggeration.

The MI355X accelerator will be manufactured on TSMC’s new 3nm N3 process, requiring major changes compared to the N5, but the core design elements may remain similar to CDNA3. The HBM3e memory capacity will reach 288 GB. The accelerator will be equipped with 10 computing elements per GPU, and performance will reach 2.3 petaflops of computing power for FP16 operations and 4.6 petaflops for FP8, which is a 77% increase compared to the previous generation accelerator.

One of the key innovations of the MI355X will be support for FP4 and FP6 floating point numbers, which will double the processing power of FP8, allowing 9.2 petaflops of performance in FP4. For comparison, the Nvidia Blackwell B200 offers up to 9 Pflops of FP4 performance, while the more powerful GB200 offers 10 Pflops. Thus, the AMD Instinct MI355X could become a serious competitor for future Nvidia products, including thanks to 288 GB of HBM3E memory – this is 50% more than Nvidia Blackwell. Moreover, both devices will have memory bandwidth of up to 8 TB/s on the GPU.

As experts note, computing power and memory are not the only key parameters for AI accelerators. An important factor is the scalability of systems when using a large number of GPUs. So far, AMD has not revealed details about possible changes to the interconnect system between GPUs, which could be an important aspect in comparison with Nvidia’s Blackwell.

Along with the announcement of the Instinct MI355X, AMD confirmed that the Instinct MI325X accelerator has officially entered production and will go on sale this quarter. The main difference between the MI325X and the previous MI300X model was the increase in memory capacity from 192 to 256 GB. Interestingly, it was originally planned to equip the accelerator with 288 GB of memory, but apparently AMD decided to limit the increase to 33% instead of 50%. The new HBM3E memory provides a throughput of more than 6 TB/s, which is 13% more than the 5.3 TB/s of the MI300X.


AMD conducted a comparative analysis of the performance of the Instinct MI325X and Nvidia H200. The AMD accelerator turned out to be 20-40% faster than its competitor in launching already trained large language models, and showed performance parity in training neural networks.

AMD hasn’t disclosed the cost of its AI accelerators, but company officials said one of the goals is to provide a total cost of ownership (TCO) advantage. This can be achieved either through better performance at the same price, or through a lower price at the same performance. As an AMD spokesperson noted, “We are business people and will make responsible pricing decisions.” The Instinct MI355X is scheduled to ship in the second half of 2025.

admin

Share
Published by
admin

Recent Posts

Intel Arc B580 “pixel feeder” surfaces on Amazon

The next generation Intel video card, the Arc B580 made by ASRock, was discovered in…

1 hour ago

Microsoft wants every person to have an AI assistant, and every business to have an AI agent

At Microsoft's annual Ignite conference, the company unveiled its vision for the future of PC…

1 hour ago

Elon Musk presented the DOGE concept: massive cuts in the US government apparatus and simplification of government regulation

Billionaires Elon Musk and Vivek Ramaswamy presented the concept of creating a Department of Government…

3 hours ago

Google Gemini will be able to perform tasks in applications without opening them

Google Gemini artificial intelligence will gain new capabilities thanks to the “app functions” API, which…

6 hours ago

Elon Musk got off with $2,923 for failing to appear to testify in Twitter acquisition case

Elon Musk's 2022 Twitter acquisition trial is ongoing, and the billionaire was scheduled to testify…

7 hours ago