Unveiling the Future: AMD's Next-Gen Accelerators Set to Rival NVIDIA

AMD’s latest Instinct accelerator, the MI325X, has recently been outed amid an aggressive roadmap that details future versions of its CDNA architecture, especially as revealed by the company’s engineers at Computex this year. The timing of those revelations – against the backdrop of their biggest rival, NVIDIA, announcing its own wares at the same time – is particularly poignant. Here is DIVE’s own version of the teapot computation, detailing how AMD’s breakthroughs might allow it to scalably push computing performance and efficiency of its current flagship CPU (the EPYC ‘Genoa’) to a new record, competing against NVIDIA’s giant titan.

Enter the MI325X: AMD's Answer to NVIDIA's Dominance

The real star of AMD’s product announcement is the MI325X, equipped with HBM3E, that will be a mid-generation follow-on to the MI300, and take AMD accelerators even higher by comparison with NVIDIA accelerators. The MI325X will have up to 288GB of memory and as much as 6TB/second of local memory bandwidth. AMD is raising the bar for what accelerators can be.

Doubling Down on Memory Capacity

It’s not the incremental product update that many would have expected of the MI series, but a leap: the MI325X, with the jump from HBM3 memory to the new HBM3E accelerator memory, doubles the available memory space in the supercomputer accelerator. The new accelerator supports configurations with a whopping 2.3TB (terabytes) of RAM, enabling throughput easily enough to power through a trillion parameter large language model. Unlike the MP whole chip in the Cerebrus supercomputer, which is based on Intel’s process node, AMD clearly signalled beyond a simple perch-keeping gesture with the MI325X: AMD is keeping up with NVIDIA, but then also passing it in the areas that NVIDIA needs it most.

AMD's CDNA Roadmap: A Vision for the Future

However, beyond the MI325X itself, my greater takeaway from AMD’s Next Horizon event was their thorough reveal of CDNA architecture and Instinct product roadmap (updates run as far out as 2026). There is nothing so breathless as breathless openness, and AMD is opening up about a roadmap they sincerely hope to stick to, as they continue to innovate and expand efficiency. CDNA 4, the successor to CDNA 3, with which the MI350 series is due, will support an even lower level of precision for its data elements, as well as a move to the 3nm process and radically improved performance-per-Watt.

Targeting AI Inference and Efficiency

This CDNA 4 pivot also very methodically acknowledges the needs of computing in the future, especially in AI. With a lower precision data format in FP4, it’s a path of compute density and memory-constrained throughput, storage and memory pressure. All of these elements heavily influence AI inference workloads which AMD is targeting to improve an extraordinary 35x (yes, 35 times) when comparing its current MI300X models versus its incoming MI350. Again, an impressive claim, especially since NVIDIA owned the lion’s share of this space for a long time now.

The Race Heats Up: AMD vs. NVIDIA

In the world of computing, AMD and NVIDIA rivalry is nothing new. That their announcements both fell within a day of each other further illustrates the ferociousness of the competition. Already, AMD states it is instituting a CDNA roadmap cadence in 2024 to bring out a new generation of architectures more efficiently and more often. It would have big implications if the accelerator game truly changes, especially for applications that demand increasing amounts of computational power and efficiency.

Looking Beyond: The CDNA "Next" and MI400 Series

However, looking further into AMD’s roadmap, the expected CDNA ‘Next’ architecture and the MI400 series accelerator are expected to undergo significant architecture upgrades by 2026. This is AMD’s next-next-generation, and this kind of planned architecture upgrade would signal their commitment to sustainable growth and provide evidence of their long-term plans. As technology evolves, these future platforms will shape the future of computing, leaving NVIDIA no choice but to respond if they want to maintain their market share.

Towards a New Era of Accelerators

So if AMD really can bring everything together as suggested at Computex, we’re in for a fascinating ride for early 2019 and onwards. Pitted against NVIDIA, this makes AMD a formidable competitor in the accelerator space and fighting fit. As to what comes next, one thing seems sure: there’s plenty more fighting to go.

About NVIDIA

It is also important to keep AMD’s rivalry with NVIDIA in mind as we explore its roadmap and innovations. NVIDIA is one of the largest and most competitive computing and graphics companies in the world. Over the past two decades, it has led the accelerator revolution with its GPUs. Today, the company’s foundational role in the growth of AI, gaming, 3D graphics, self-driving cars, robotics and more means that AMD has a lot to live up to. As it and NVIDIA continue to push the envelope of what’s possible, the stage is set for technological change that promises to bring exciting new developments for the computing industry and consumers alike.