AMD Patents Chiplet Design To Build Colossal GPUs

AMD jumped over to using a chiplet-based CPU design with the introduction of Zen 2 in the Ryzen 3000 series CPUs, enabling the chipmaker to cram more cores into a single CPU. Now, a new patent appears to reveal that AMD wants to do the same thing with GPUs (via ComputerBase).

Breaking Through the Performance Barrier

On the face of it, this move certainly makes sense. As you increase die sizes, yields tend to drop drastically due to defects in the silicon, leading to costly losses that impact prices. Semiconductor manufacturing equipment also has a reticle size limit, which essentially creates a barrier beyond which it simply isn't possible to make a bigger GPU. Coupled with the state of the process node (TSMC is currently at 7nm), there is a performance ceiling that foundries simply cannot manufacture beyond. 

However, breaking the big silicon die into multiple little ones addresses those pain points. But things aren't quite this simple, and AMD's patent explains why this hasn't been done yet. 

The long and short of it is that parallelism in GPU compute workloads is difficult to span across multiple chiplets. CPU workloads are big and do not require a lot of communication between each other -- which is why it's comparatively easy to make a chiplet-based CPU. 

Make it Look Like One, Big, Monolithic GPU

But GPU workloads are tiny, which leads to lots of traffic overhead on the internal communication fabrics. To solve this, AMD's approach uses a high-bandwidth interconnect to facilitate communication between chiplets -- AMD calls this crosslink an HBX. At the physical level, this will look much like the interposer of Zen 3 CPUs, except that on an electrical level, it is connected with a focus on L3 cache synchronization to account for the parallel workload.

The design is proposed such that the CPU is connected to the first GPU chiplet, and one passive interconnect ties the L3 cache and other channels together between chiplets. This means that as far as the CPU is concerned, it communicates with one big GPU rather than a bunch of little GPUs. 

A single controller that communicates with the smaller GPUs isn't a viable solution. The parallel workloads would generate too much traffic, and active switching would quickly create a (latency) bottleneck -- or demand far too big of a controller. As such, AMD proposes that a chiplet-based GPU needs a simple electrical path to be tied together to look like one big GPU, rather than a bunch of little ones addressed through a controller.

The beauty of this GPU model is that it remains compatible with existing coding languages, and the GPU model doesn't change from the developers' point of view. A few things will need to be adjusted in the driver, naturally, to account for the new architecture, but existing software should run without any major changes.

The patent also goes on to explain that the layout of the chiplet-based GPU does not have to take form with four chiplets. Other configurations, die sizes and shapes may be welcome depending on the client's needs, with even pentagon-shaped dies as an option. Naturally, the design should be scalable, though symmetry between the chiplets would be a limitation. 

Sounds Great. When is it Coming?

As great as it all sounds, we don't want to get your hopes up for such a product to come out anytime soon. SLI and Crossfire died because getting multiple GPUs to work together across different cards is a pain, and even with AMD's proposed solution of bringing the GPU chiplets closer together with a high-bandwidth interconnect, there is still a lot of work to be done. 

Chances are that if this manifests into a real-world product, it will first happen on a research-level scale, being aimed at supercomputers or scientific-purpose GPUs for users with high GPU power needs in single workstations. Such a ludicrous amount of GPU horsepower will likely need to be coupled to HBM memory just to keep up, so it's likely that you can rule out consumer products for some time to come. 

All that being said, it's also very possible that this will never become a thing. Tech companies file a lot of patents, and most of them never end up being used.