AMD Pushes the Boundaries of Multi-Chip Module GPUs With New Patent

AMD takes the GPU battle up a notch with a new patent on chiplet GPU design that could potentially drive down production costs and give significant performance gains. The multi-chip module (MCM) based GPU patent could allow AMD to steal a march over rival chipmakers NVIDIA and Intel. 

The semiconductor industry has entered a new era of chiplet revolution. Chipmaker Advanced Micro Devices (AMD) is betting big on a modular chipset architecture based on multi-chip module (MCM) design for its Radeon GPUs. If successful, the move will mark a shift away from expensive monolithic design and power significant computing growth. 

MCM or hybrid integrated circuit (IC) is an electronic package where multiple ICs are integrated onto a unifying substrate. They offer packaging efficiency of more than 30%, improve performance, and reduce costs. The individual ICs that form powerful MCM are known as chiplets and have gained traction due to faster time-to-market, reduced manufacturing costs, and improved yields. 

Though Nvidia Opens a new window has reportedly been experimenting with MCM GPUs, they’re hard to implement due to programming models, the high latency between chiplets, and the difficulty in implementing parallelism. AMD made the first move by submitting Opens a new window a new patent on chiplet GPU designs to the U.S. Patent and Trademark Office (USPTO) on December 31, 2020. 

According to AMD’s patent, the chipmaker plans to build a GPU chiplet graphics card using high bandwidth passive crosslinks (HBPC). With HBPC, the first GPU chipset will be directly communicably coupled to the CPU, while each of the GPU chips in the array would be coupled to the first GPU via a passive crosslink. The GPU chiplets will be placed on a single interposer, and the passive crosslink will work as communication wires between chiplets. 

Also Read: Apple Could Beat Intel Chips With Its New Mac Processors

The patent explainedOpens a new window , “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.”

Also, conventional GPUs have shared last-level cache (LLC), but AMD’s patent reveals that each GPU chiplet should have its own LLC to avoid memory synchronization issues.

AMD has already seen success in multi-chip module (MCM) design with its Ryzen processors that use compute chiplets built on TSMC’s high-performance 7nm manufacturing node. With the success of compute-based chiplet, experts believe that AMD could quickly develop cheaper graphics-based or AI-powered chiplet for different workloads and gain an edge over NVIDIA and Intel. 

Following AMD’s footsteps in the chiplet technology, Intel introduced its first chiplet solution, Foveros, in 2019. It is an advanced 3D face-to-face integration technology with which the company developed its new Lakefield mobile processorOpens a new window in June 2020. NVIDIA is too expected to make waves in the MCM design with its Hopper architecture.

Since MCM for GPU designs claims to improve processor performance, reduce manufacturing costs and complexities, manufacturers might depart from expensive monolithic designs. This move can also be viewed as an attempt to grab a sizable GPU market share from entrenched hardware rival Nvidia.  

What are your thoughts on AMD’s new patent? Comment below or let us know on LinkedInOpens a new window , TwitterOpens a new window , or FacebookOpens a new window . We’d love to hear from you!