AMD announced that its MI325X GPU will launch later this 12 months and that it beats NVIDIA's H200 GPUs in several ways.
NVIDIA is currently the market leader in data center chip sales and is estimated to carry over 70% of the AI chip market share, but Intel and AMD are also making great strides with their very own offerings.
At Taiwan's Computex 2024, AMD CEO Lisa Su announced that the Instinct MI325X GPU will likely be released later this 12 months. AMD says its successor to the MI300 will offer more memory and faster data throughput.
AMD's data center chips are gaining traction, with firms like Microsoft integrating them into their Azure cloud services, while Microsoft, Meta and Oracle have all adopted the MI300 platform. The company is targeting $4 billion in revenue by 2024.
Su said the MI325X significantly outperforms NVIDIA's popular H200 GPU by way of memory capability, bandwidth and performance.
The MI325X features as much as 288GB of HBM3e memory and 6TBps of bandwidth. That's greater than twice the memory of NVIDIA's H200 and offers 30% faster data throughput.
It achieves a theoretical peak throughput of two.6 petaflops for 8-bit floating point (FP8) and 1.3 petaflops for 16-bit floating point (FP16), 30% greater than the H200.
A single server consisting of eight of those GPUs combined on the Instinct MI325X platform has enough memory to run advanced AI models with as much as 1 trillion parameters, twice the scale supported by an H200 server.
Today at Computex, we unveiled a multi-year, expanded roadmap for the AMD Instinct accelerator, delivering annual improvements in world-class AI performance and memory capability each generation. pic.twitter.com/PQKyVAnAVW
— AMD (@AMD) June 3, 2024
While the H200 is the flagship NVIDIA GPUs currently available, Su didn’t mention how the MI325X compares to the Blackwell GPUs that NVIDIA will ship later this 12 months.
Blackwell's top-end models feature as much as 192GB of HBM3e memory with 8TB/s bandwidth. That's significantly less memory, but more bandwidth than AMD's top offering.
AMD desires to push ahead with the event of latest GPUs and produce a brand new GPU family to market yearly. This is the rhythm that NVIDIA CEO Jensen Huang says he can be aiming for.
In 2050, AMD will release the MI350, which uses a 3nm process and uses AMD's CDNA 4 architecture.
Su says the CDNA 4 architecture will represent a generational leap in AI computing, delivering 35x the inference performance in comparison with the present CDNA 3 architecture.
Moving all this data between GPU clusters and servers requires high-speed networks. Su concluded her speech by saying, “The way forward for AI networks should be open.”
She announced that AMD joined a consortium of high-performance computing firms last week to develop a high-bandwidth, low-latency networking standard to attach lots of of AI accelerators.
NVIDIA was not invited to take part in the project and has its own proprietary standard. Su said the UALink standard is a “great alternative to proprietary options.”
The fronts in the event of AI GPUs have been very unspectacular and NVIDIA is seemingly slowly feeling the pressure from AMD and Intel.
