AMD’s new Instinct GPUs might just blow Nvidia out of the water

AMD has announced the launch of its new Instinct MI325x accelerators as competition heats up with Nvidia in the AI space.

Unveiled at the chip maker’s Advancing AI conference in San Francisco, AMD said the new range is designed specifically to drive performance and efficiency in contending with “demanding AI tasks”.

The new Instinct range is aimed at supercharging foundation model training, fine-tuning, and inference, the company said.

The announcement from AMD marks the latest in a slew of progressive upgrades to its Instinct graphics processing unit (GPU) series in recent years and forms a key focus in its ongoing roadmap to drive AI infrastructure advances for hyperscale customers.

In a press briefing ahead of the Advancing AI event today, AMD executives waxed lyrical about the latest iteration, emphasizing performance capabilities compared to competitors such as Intel and Nvidia.

But it’s the latter of these that enterprises will be keen to see the comparable benchmarks for.

AMD said the MI325X accelerators provide “industry-leading” memory capacity and bandwidth, boasting 256GB of HBM3E supporting 6TB/s. All told, this equates to 1.8x more capacity and 1.3x more bandwidth than Nvidia’s H200 series.

In terms of LLM inference capabilities, the company told assembled press the MI325X can provide up to 1.3x the inference performance on the Mistral 7B model, for example, along with 1.2x inference performance on the Llama 3.1 70B range.

AMD revealed the MI325X accelerators are currently on track for production shipments in Q4 2024, and are expected to have widespread availability from a host of platform providers by Q1 2025.

These include Dell Technologies, HPE, Lenovo, Supermicro, and more.

AMD’s roadmap is accelerating

Looking ahead, AMD execs gave attendees a glimpse at what’s in store with regard to future offerings in 2025 and beyond. The chip maker previews the next-generation AMD Instinct MI350 series accelerators at the annual conference, highlighting its rapid-fire efforts in the next few months.

Based on the AMD CDNA 4 architecture, the MI350 series accelerators are designed to deliver a whopping 35x improvement in inference performance compared to CDNA 3-based accelerators.

This will also include increased memory capacity, the company confirmed, reaching 288GB of HBM3E memory per accelerator.

Development of the MI350 series is on track, AMD said, and will be available during the second half of 2025.