Intel patent hints at future disaggregated, chiplet-based GPU designs

Editor’s take: Intel’s patent submitting supplies a glimmer of hope that we could someday see disaggregated GPUs grow to be a actuality, nevertheless it’s nonetheless early days. The corporate’s upcoming Arc Battlemage structure, slated for early 2025, will possible characteristic a monolithic design as soon as once more. Nonetheless, the exploration of chiplet approaches may lay the groundwork for future GPU generations that make the most of this cutting-edge tech.

Intel, in response to a not too long ago found patent submitting, is exploring using disaggregated architectures, which contain dividing GPU designs into smaller, specialised “chiplets.”

For these not accustomed to the time period, disaggregated, or chiplet-based architectures, signify a departure from conventional monolithic GPU designs. As an alternative of housing every little thing on a single, large chip, the assorted elements are break up throughout a number of smaller chiplets, every tailor-made for particular duties like computing, graphics rendering, or AI workloads. These chiplets are then interconnected utilizing cutting-edge packaging applied sciences.

Earlier this month, Intel was lastly granted a patent for its disaggregated GPU structure, which can possible be the primary industrial GPU structure with logic chiplets, additionally permitting for the power-gate of chiplets not used to course of workloads. pic.twitter.com/XsNjjdVIOu

– Underfox (@Underfox3) October 26, 2024

There are a number of potential benefits to this modular method. For starters, it may result in considerably improved energy effectivity, which is a giant win in an period the place vitality prices and environmental influence are main considerations. Since particular person chiplets will be powered down when not in use, because of power-gating capabilities, total vitality consumption will be diminished.

However that is not all – disaggregated GPU designs additionally supply elevated flexibility, customization, and modularity. Producers may combine and match completely different chiplets to create specialised GPU configurations optimized for numerous workloads or market segments.

In fact, Intel is not the one one eyeing this modern GPU design technique. Major rival AMD has additionally filed patents associated to chiplet-based GPU architectures up to now. Final 12 months, even Nvidia was rumored to make use of a multi-chiplet design for its Blackwell GPUs, although newer experiences debunked that.

Both approach, it is getting more and more clear that each one three firms are eyeing such structure. AMD’s newest Intuition MI300 collection AI GPUs already supply a twin chiplet-based design, as do a number of of Nvidia’s server choices. Nonetheless, we’ve but to see a real chiplet-based GPU hit the buyer market.

Making the leap from monolithic to disaggregated GPU designs isn’t any small feat, although. There are vital manufacturing and engineering challenges to beat, to not point out the necessity for superior interconnect applied sciences to seamlessly hyperlink all these tiny chiplets.