On the ISSCC 2023 earlier this week, AMD mentioned the way forward for computing over the subsequent decade. CEO Dr. Lisa Su was the lead presenter and confirmed that AMD has carried out admirably in supercomputer, server, and GPU efficiency traits over the previous few a long time. Nevertheless, in all probability extra fascinating are the well-crafted plans that present how AMD goals to maintain the pedal to the metallic and use superior applied sciences to counteract the truly fizzling out of semiconductor course of shrink advantages.
Picture 1 of 3
Within the above efficiency slides, AMD claims it has efficiently doubled mainstream server efficiency each 2.4 years since 2009. Nevertheless, it doesn’t share any projections with respect to this market. AMD is assured to see additional into the longer term with its GPU efficiency traits (slide 2 within the gallery above). Right here you’ll be able to see that it claims to have doubled GPU efficiency each 2.2 years since 2006. The chart exhibits this development is locked in till no less than 2025.
AMD’s supercomputer efficiency is essentially the most profitable when it comes to advances over time, with the final chart above exhibiting that, because the late 90s, AMD processors have been instrumental in doubling supercomputer efficiency each 1.2 years. Furthermore, AMD predicts we are going to attain Zetascale supercomputer efficiency in roughly a decade. AMD additionally took time to spotlight effectivity beneficial properties, and the extraordinary battle to maintain Moore’s Legislation alive as logic density tapers off.
Generously, AMD laid out a few of its key plans which can assist it drive ahead with effectivity and efficiency beneficial properties over the approaching decade. Superior packaging goes to be a powerful driver of each efficiency and effectivity, in keeping with AMD. We have now already seen a number of the outcomes of AMD touring down this path with using chiplets and 3D V-Cache, and it’ll proceed.
Picture 1 of 3
Some superior packaging avenues which can be explored embody the mixing of 3D CPU & GPU silicon “for next-level effectivity.” Moreover, AMD reckons that “even tighter integration of compute and reminiscence” will lead to increased bandwidth with decrease energy. AMD can even goal processing in reminiscence. A slide shared at ISSCC confirmed a processor with an HBM module stacked on high. AMD says that if key algorithmic kernels could be executed in reminiscence, that takes vital burden and latency out of the system.
One other large goal for effectivity financial savings, and thus potential efficiency boosts, are chip I/O and communications. Particularly, utilizing optical communications tightly built-in with the compute die is anticipated to supply a worthwhile effectivity increase.
AMD additionally took a while to boast in regards to the AI efficiency beneficial properties which have been delivered by its processor portfolio during the last decade. The presentation mentioned some use instances for AI computing and highlighted the potential efficiency beneficial properties AI can ship for simulations.
After all, AMD isn’t alone in the advantages of superior packaging, chiplets, die stacking, in-memory computing, optical computing, and AI acceleration. It’s good, although, to see that it has stable plans for fierce competitors with rivals, and to provide ever sooner and extra environment friendly chips for PC fanatics.