News Categories

NVIDIA could circumvent Moore’s Law with a new multi-module GPU design

By Koh Wanzi - on 7 Jul 2017, 10:38am

NVIDIA could circumvent Moore’s Law with a new multi-module GPU design

NVIDIA Pascal GPU. (Image Source: NVIDIA)

NVIDIA researchers have published a paper outlining the potential benefits of a new GPU design that could enable it to sidestep the limitations of Moore’s Law.

The paper also includes authors from Arizona State University, the University of Texas, and the Barcelona Supercomputing Center, and details a multi-module GPU design that could help NVIDIA continue to cram more transistors onto a single GPU package, eclipsing what traditional single-chip designs are capable of.

If NVIDIA were to continue with its current approach, the company says that GPU performance improvements would eventually plateau. Existing monolithic GPU designs are not sustainable in the long term, and stuffing more transistors into smaller spaces will result in additional challenges such as diminishing die yields and growing R&D costs.

NVIDIA’s proposed solution is a multi-chip module (MCM) that connects multiple smaller GPU modules (GPMs) on the same package, relying on high-speed I/O technology to allow them to communicate with each other.

NVIDIA multi-chip module

In the company’s own words, this means “partitioning GPUs into easily manufacturable basic GPMs, and integrating them on package using high bandwidth and power efficient signaling technologies”.

The key takeaway is that these GPMs would be simpler and easier to make, thus allowing for less complex designs compared to a monolithic approach.

To evaluate the potential performance, researchers simulated two virtual GPUs, each with 256 streaming multiprocessors (SMs). One used the current monolithic design, while the other followed the MCM design. This was purely hypothetical, as a 256 SM GPU is simply "unbuildable" at the moment, according to NVIDIA.

Still, the MCM design reportedly performed within 10 per cent of the monolithic GPU, and was also almost 27 per cent quicker than a similarly configured SLI setup.

Performance comparison of simulated MCM vs monolithic GPU. (Image Source: NVIDIA)

NVIDIA isn’t the only company looking into a multi-chip approach. In March, Intel floated the idea of a heterogeneous chip design that would mix and match chips from different generations.

Similarly, AMD has Infinity Fabric, an interconnect technology that joins together multiple quad-core Zen modules. More interestingly, Raja Koduri, SVP and Chief Architect at the Radeon Technologies Group, has specifically said that multi-GPU designs are possible with Infinity Fabric.

But if MCM GPUs are going to reach consumers, a lot of work will need to be done to ensure effective scaling.

We probably won’t see this technology any time soon, but all the signs suggest that it’s just a matter of time.

Source: NVIDIA

Join HWZ's Telegram channel here and catch all the latest tech news!
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.