Semiconductor engineer Tom Wassick has found what he believes to be 3D V-Cache performance on one in every of AMD’s finest GPUs, the RX 7900 XT. The engineer took a peak contained in the 7900 XT’s die with infrared imaging and located the identical sort of 3D V-Cache connection factors used on AMD’s Zen 3 and Zen 4 structure. Wassick noticed the connections on the MCD die.
Wassick can’t say if these TSV connection factors might be used for caching functions particularly, however AMD has made no identified plans to develop its 3D packaging capabilities past vertically stacked cache at this level. That makes it look like these connection factors can be used with some form of 3D cache in thoughts to extend gaming efficiency and/or compute efficiency.
The invention comes after quite a few unconfirmed rumors that AMD would add 3D V-Cache tech to its GPUs.
3D V-Cache has been used to nice success on AMD’s Ryzen and EPYC CPUs up to now. The know-how depends on a hybrid bonding approach that fuses a further 64MB slab of cache on high of a Ryzen or EPYC compute die to extend L3 cache capability. At the moment, this 3D stacking approach has allowed AMD to double the quantity of L3 cache obtainable to its desktop Ryzen 9 7900X3D and 7950X3D elements whereas tripling it on its Ryzen 7 5800X3D, 7800X3D shopper chips and EPYC Milan-X server processors.
The efficiency advantages from this know-how have been spectacular, with 3D-V-Cache chips gaining a full generational improve in efficiency in functions that profit closely from giant chunks of cache. A very good instance of that is with the Ryzen 7 5800X3D the place we noticed a 28% uplift in gaming efficiency towards the Ryzen 9 5900X, and seven% quicker efficiency than the Core i9-12900KS.
AMD’s server counterparts are much more spectacular, with Milan-X benchmarks from AMD and Microsoft displaying efficiency enhancements of properly over 50% towards customary Milan elements. Nevertheless, this know-how cannot magically improve efficiency at will. Solely cache-sensitive workloads will see such a conduct.
What else are you able to see? A linear array of “spots” that look remarkably just like the maintain out zones on X3D, and which might be on the identical 17-18 um pitch. May they be contemplating stacked MCD performance (or perhaps they’re one thing else)?January 27, 2023
We don’t know how 3D V-Cache would function in a GPU utility. However in principle, the primary ideas of 3D V-Cache ought to nonetheless apply. Having extra cache capability would allow quicker processing of cache-sensitive workloads for the reason that GPU has to make fewer journeys to its slower GDDR6 reminiscence.
We have already seen a very good instance of this with AMD’s Infinity Cache within the RX 6000 collection the place AMD was ready to make use of slower GDDR6 reminiscence and retain the identical efficiency as Nvidia’s RTX 30 collection GPUs, that includes power-hungry GDDR6X reminiscence, due to that infinity cache conserving the GPU fed with knowledge.
Nevertheless, we do not know if the identical behaviors will apply with 3D V-Cache. This can all rely upon how delicate AMD’s GPU architectures are to further cache capability, and what number of functions will profit.
One other downside AMD should cope with is thermals. We have seen this situation extensively on AMD’s Ryzen X3D processors, the place the extra slab of cache hinders thermal dissipation, leading to decrease CPU frequencies and better temperatures on the similar time (compared to a non-X3D half). There is a excessive probability AMD would cope with the identical points on 3D V-Cache GPUs, and be pressured to scale back clock speeds to maintain temperatures in verify.
Nonetheless, its cool to see AMD probably trying into the concept of including 3D-Vache to its GPUs. We may very well be taking a look at AMD’s subsequent silver bullet, to “magically” improve gaming efficiency.