AMD Rents Its Own Chips From Google to Design New Chips

Modern chip design processes require a lot of compute performance. In a press release from Google Cloud via PR Newswire, we learn that in a bid to extend the  capabilities of its own datacenters, AMD will use Google Cloud to run some of its electronic design automation (EDA) workloads. Among instances that AMD plans to use will be the compute-optimized C2D VM instance that is powered by AMD's own EPYC CPUs. Meanwhile, other instances that AMD will use don't seem to be AMD-powered.

AMD's engineers design dozens of chips in parallel, which requires formidable compute capabilities but it does not require this power all of the time. Adding more machines to on-premises datacenters is an option, but if they are not needed then they are inefficient from a power and financial points of view. By lending CPU horsepower from Google Cloud, AMD gets the right amount of compute performance for its short-term projects, without impacting server allocation for long-term projects. AMD also keeps its investments in check.

There are other advantages that AMD gets by working with Google Cloud. In addition to the compute-optimize C2D VM instance, AMD plans to use Google Cloud's networking, storage, artificial intelligence (AI), and machine learning (ML) capabilities. Perhaps, AI and ML are as important parts of the announcements as the C2D VM.

Today it takes years to develop a new CPU or GPU architecture. Physical implementation of the chip (chip floorplan, layout, placement, routing, etc.) may take an additional 24 months or more to design. Every aspect of the chip's implementation has to be evaluated and simulated. Evaluating hundreds of design options and exploring feasibility of all potential combinations cannot be done by human engineers in a reasonable amount of time and this is where AI and ML come into play.

Synopsys last year introduced its DSO.ai platform that uses AI to figure out the best possible place and route configuration in terms of performance, power, and area (PPA) without performing lengthy simulations. Samsung was the first company to adopt the Synopsys DSO.ai platform for chip design, Google, and Nvidia are also using it. While we cannot say whether AMD uses Synopsys's Design Space Optimization AI, Cadence's Cerebrus ML platform, or has its own way of leveraging AI and ML for chip designs, we can certainly say that AI and ML are becoming important factors in chip design flows these days.

"In today's semiconductor environment, the speed, scale, and security of the cloud unlock much needed flexibility," said Sachin Gupta, GM and VP, Infrastructure, at Google Cloud. "We are pleased to provide the infrastructure required to meet AMD's compute performance needs and equip the company with our AI solutions to continue designing innovative chips."