Unveiling the AI-focused AMD Instinct MI300X GPU

The rapid acceleration of enterprise artificial intelligence usage has brought a dramatic demand surge for advanced chips, such as graphics processing units. In answering that demand, Advanced Micro Devices Inc.’s GPU product family, AMD Instinct, is designed to cater to the high demands of AI training.

“At AMD, we developed a GPU family called Instinct,” said Marty Poniatowski (pictured), presales director of East and Central Canada at AMD. “The latest version is the MI300X, and it has 192 [gigabytes] of memory, which means you can run against it very large Llama 2 models. It’s made for really high-end training, and it’s a great alternative because no customer wants to buy just from one manufacturer.”

Poniatowski spoke with theCUBE Research’s John Furrier and Dave Vellante at HPE Discover, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed the AMD Instinct GPUs and the company’s comprehensive approach spanning high-performance, adaptable software and innovative processors. (* Disclosure below.)

AMD Instinct MI300X: Meeting the high-performance GPU demand

The MI300X’s design and memory specifications have equipped it to handle intricate AI workloads. Integrations are already underway in Microsoft Azure, which has optimized its environments to leverage MI300X, allowing customers to experiment and deploy AI models efficiently, according to Poniatowski.

“Microsoft Azure has an MI300X environment that’s optimized with Hugging Face,” he said. “A lot of my customers are trying this out. They like to get them on-premises, but they can try them out right in Azure.”

The departure from solution lock-in is particularly heightened given the prevalence of multicloud and hybrid infrastructures. With open-source frameworks such as PyTorch and JAX, which integrate seamlessly with AMD’s ROCm software, users can operate across various GPUs without being tied to a single vendor, Poniatowski explained.

“Now you have multiple GPUs that you can run, and you have the software stack that’ll support any GPU,” he said. “So, now my customers feel protected. They’re not locked in from a hardware standpoint or from a software standpoint.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE Research’s coverage of HPE Discover: 

(* Disclosure: TheCUBE is a paid media partner for HPE Discover. Neither Hewlett Packard Enterprise Co. and Intel Corp., the primary sponsors of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE