AMD Patent Details Innovative Stacked Accelerator That Could Empower Next-Gen RDNA GPUs

AMD goes all-in with chip stacking applied sciences, just like the stacked 3D V-cache that’s headed to its refreshed Zen 3 processors for a claimed 15 % efficiency enhance, and in addition its third Gen EPYC processors with an excellent greater efficiency uplift. However what about on the GPU facet? A patent utility might shed some mild on AMD’s future plans.
AMD Accelerated Processing Die Image
Supply: AMD through USPTO
In a patent filed a yr in the past that was printed only a few days in the past, AMD outlines a further die stacked on high of the GPU, separate from the present chiplet construction. AMD calls this additional die an accelerated processing gadget (APD) and pitches it as a approach to improve machine studying capabilities. It might include reminiscence and a number of ML accelerators.

By stacking the extra die, there may be an inherent efficiency benefit that comes by means of high-speed interconnects to switch information and instructions. AMD notes the reminiscence portion may very well be configurable as both or each a cache and as straight accessible reminiscence. It additionally says it might include matrix multiplication arithmetic logic models (ALUs) configured to carry out matrix multiplication operations to spice up ML chores.

AMD Accelerate Processing Device Patent Image
Supply: AMD through USPTO
You may see the overall make-up within the patent sketch above. It is essential to notice {that a} patent utility shouldn’t be a foregone conclusion of upcoming know-how in precise merchandise. And if if this does come to fruition, it is not clear if it could be sure for high-performance computing (HPC) like its dual-die Intuition MI200 or future gaming merchandise, or each. However we’re actually intrigued on the prospect.
There’s already been chatter for AMD bringing stacked 3D V-cache to its gaming GPUs, and including one other stacked die to the combo is actually throughout the realm of chance. An APD die might serve a few of the identical functions as NVIDIA’s Tensor cores.
It is likely to be too late to count on this to be a part of RDNA 3, however maybe RDNA 4 is honest sport. That is after all assuming AMD turns its patent into an precise product. Within the meantime, you possibly can try AMD’s full patent for extra particulars.

Be the first to comment

Leave a Reply

Your email address will not be published.