Multi-Chip Module or MCM design architecture could make its way to consumer graphics cards if a new patent filed by AMD is to be believed.
Multi-Chip Module or MCM design architecture could make its way to consumer graphics cards if a new patent filed by AMD is to be believed. The patent document reveals how AMD plans to build a GPU chiplet graphics card, and the process resembles MCM based CPU designs. With NVIDIA already invested heavily in MCM-based graphics cards, AMD was a little late, but not far behind.
The newly filed patent by AMD attempts to address the technological limitations or restrictions that prevented the company from embracing MCM design architecture for GPUs sooner. The company explains that it is finally ready with a High Bandwidth Passive Crosslink to solve the latency, bandwidth, and overall communication issues between multiple GPU chiplets on an MCM GPU board.
Monolithic or Singular Graphics Chip Designs To Be Eclipsed By MCM GPU Chiplets?
High latency between chiplets, programming models, and difficulty in implementing parallelism were the core reasons AMD couldn’t move ahead with MCM GPU Chiplet architecture, claims the newly filed patent by the company. To address multiple issues, AMD is planning to use an on-package interconnect it calls the High Bandwidth Passive Crosslink.
High Bandwidth Passive Crosslink would enable each GPU chiplet to communicate with the CPU directly as well as other chiplets. Each GPU would also feature its own cache. Needless to add, this design implies each GPU chiplet would appear as an independent GPU. Hence, an operating system could fully address each GPU in the MCM Architecture.
The change to the MCM GPU Chiplet design could happen after RDNA 3. This is because NVIDIA is already deep into MCM GPU with its Hopper Architecture. Moreover, Intel has been suggesting that it has succeeded with MCM design methodology. The company even offered a brief demonstration.
AMD Has Built Multi-Chip Module Based Products With ZEN 3 Architecture:
AMD’s ZEN-based processors have been great in the HEDT space. Its latest ZEN 3 Ryzen Threadripper CPUs have 32 Cores and 64 Threads. It was quite difficult for consumers to imagine a 6 Core 12 Thread CPU a few years back, but AMD has successfully delivered powerful multi-core processors. In fact, even the power of server-grade CPUs has trickled down to consumers.
Modern-day manufacturing of silicon wafers is undoubtedly tricky. However, the company has successfully evolved to a 7nm Fabrication Process. Meanwhile, Intel is still holding on to the archaic 14nm Production Process. Intel keeps coming up with branding such as SuperFin, but the technology hasn’t yet advanced considerably.
An MCM design approach instantly boosts yields too. A single, monolithic die has a rather poor yield. However, breaking the same die into multiple smaller chips instantly boosts the overall yield of the die. Thereafter arranging these GPUs chiplets in an array or configuration according to the needed specifications is obviously the way forward.
Given the obvious benefits and the economics involved, it is no wonder that every CPU and GPU manufacturer is interested in MCM chiplet design architecture. While NVIDIA and Intel have made a lot of progress, AMD is now setting its affairs in order.
Originally published at Appuals