Notebookcheck
, , , , , ,
search relation.
, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
 

Patent shows future AMD RDNA-based GPUs could leverage multi-chip modules a la Intel Xe HP Arctic Sound and NVIDIA Hopper, but RDNA 3 will likely be monolithic

Future AMD GPUs could sport MCM designs. (Image Source: AMD)
Future AMD GPUs could sport MCM designs. (Image Source: AMD)
AMD could be offering multi-chip module (MCM) designs in GPUs in the near future. A patent filed in 2019 details how an MCM design can help lower costs and improve yields. Since most GPU-bound programs work best with monolithic cores, AMD's patent describes how the MCM design can be made oblivious to the CPU thereby alleviating the need to rewrite code to take advantage of this design.

Working For Notebookcheck

Are you a techie who knows how to write? Then join our Team! English native speakers welcome!

News Writer - Details here

After having successfully demonstrated the efficacy of a multi-chip module design (MCM) on the CPU side, AMD may very well look forward to repeat the success on the GPU side of the equation as well. A patent titled "GPU Chiplets using High Bandwidth Crosslinks" filed in 2019 and now available on the Free Patents Online (FPO) website indicates that AMD does have some tricks up its sleeve to effectively transition from the conventional monolithic design.

The patent's abstract reads,

A chiplet system includes a central processing unit (CPU) communicably coupled to a first GPU chiplet of a GPU chiplet array. The GPU chiplet array includes the first GPU chiplet communicably coupled to the CPU via a bus and a second GPU chiplet communicably coupled to the first GPU chiplet via a passive crosslink. The passive crosslink is a passive interposer die dedicated for inter-chiplet communications and partitions systems-on-a-chip (SoC) functionality into smaller functional chiplet groupings."

In the patent application, AMD notes that monolithic dies are becoming expensive to manufacture. It also notes that parallel workloads in GPUs are difficult to properly distribute across multiple chiplets and keep memory across them in sync. Adding to this onerous task is that most applications are written keeping a single GPU in mind. Therefore, the challenge is to preserve the current programming model as much as possible while still offering a chiplet design to reduce manufacturing costs and improve yields.

To overcome these concerns, AMD's patent envisages using the first GPU in the series of chiplets to directly couple to the CPU. The other GPU chiplets in the array would then be coupled to the primary GPU via a high-bandwidth passive crosslink (HBPC). The HBPC is essentially a passive interposer die dedicated for inter-chip communication across the SoC.

With regards to memory sync, the patent notes that though each GPU chiplet would feature its own last-level cache (LLC), all such LLCs are coupled in a way that enables coherency across all GPU chiplets. Since only the primary GPU chiplet receives requests from the CPU, the entire complex still appears monolithic to the CPU and consequently to the OS.

At the moment, AMD already offers multi-GPU designs on a single board. Previously, we have seen the AMD Radeon R9 295X2 and the Radeon Pro Duo. The newer Radeon Pro Vega II Duo is exclusive to the Apple Mac Pro and features two Radeon VIIs with 64 compute units (CUs) each connected by an Infinity Fabric link. 

Back in 2018, AMD's SVP of Radeon Technologies Group (RTG) David Wang opined that unless a multi-GPU package is made invisible, ISVs may show some reluctance. However, he also expressed optimism that "anything's possible".

Now, it is not clear when we would see such a design in flesh. Official CDNA and RDNA roadmaps only point to an "Advanced Node" for 2022, so it is likely that we may not see CDNA 2 or RDNA 3 sporting multi-chip module (MCM) designs.

That being said, the competition is already showing similar trends with both Intel and NVIDIA looking to offer MCM parts with Xe HP Arctic Sound and Hopper, respectively.

, , , , , ,
search relation.
, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
 
AMD CDNA roadmap. (Source: AMD)
AMD CDNA roadmap. (Source: AMD)
AMD RDNA roadmap. (Source: AMD)
AMD RDNA roadmap. (Source: AMD)
CPU, RAM, and GPU MCM communication. (Source: FPO)
CPU, RAM, and GPU MCM communication. (Source: FPO)
Multi-chip GPU design. (Source: FPO)
Multi-chip GPU design. (Source: FPO)
GPU chiplet floor plan. (Source: FPO)
GPU chiplet floor plan. (Source: FPO)
A four-chiplet configuration with passive crosslinks. (Source: FPO)
A four-chiplet configuration with passive crosslinks. (Source: FPO)
Inter-chiplet communication flow diagram. (Source: FPO)
Inter-chiplet communication flow diagram. (Source: FPO)
 
static version load dynamic
Loading Comments
Comment on this article
Vaidyanathan Subramaniam
Vaidyanathan Subramaniam - Managing Editor - 1395 articles published on Notebookcheck since 2012
Though a cell and molecular biologist by training, I have been drawn towards computers from a very young age ever since I got my first PC in 1998. My passion for technology grew quite exponentially with the times, and it has been an incredible experience from being a much solicited source for tech advice and troubleshooting among family and friends to joining Notebookcheck in 2017 as a professional tech journalist. Now, I am a Lead Editor at Notebookcheck covering news and reviews encompassing a wide gamut of the technology landscape for Indian and global audiences. When I am not hunting for the next big story or taking complex measurements for reviews, you can find me unwinding to a nice read, listening to some soulful music, or trying out a new game.
contact me via: @Geeky_Vaidy
Please share our article, every link counts!
> Notebook / Laptop Reviews and News > News > News Archive > Newsarchive 2021 01 > Patent shows future AMD RDNA-based GPUs could leverage multi-chip modules a la Intel Xe HP Arctic Sound and NVIDIA Hopper, but RDNA 3 will likely be monolithic
Vaidyanathan Subramaniam, 2021-01- 3 (Update: 2021-01- 3)