geordiemp
Member
The AMD patent is about shared L1 cache which gives an average of 22 % more performance for shader workload, and for some of the work loads almost 50 % (BVH calcs in ray tracing).
Note that the technique has a shared and private mode for the L1 cache, as not every instruction benefits from sharing data between caches.
What is not clear is the mesh sharing of data, as this also brings L2 into consideration and how big the cache is on die overall.
The current rumours is that this will be a main of AMD reveal of RDNA2, as 22 % performance increase by optimising an L1 cache is a big deal, what we dont know is how big L2 cache will be to fgurther increase the IPC, as it is also suggested L2 is on the mesh connections between caches.
I doubt the recent rumours of 128 MB of L2 cache, that would be infinity in die terms but would take up so much die it seems overkill, and I would think as cache is meshed to talk to each other 128 MB is a combination of caches, but not long now I guess until the mystery clears.
22 % IPC would make a 23 to 24 TF card punch close to a 30 TF card, its not all about TF. Certainly interesting and cant wait. This generation will be about speed and caches from AMD at least.
Has Ps5 got infinity cache ?
Who knows, to add infinity cache mesh connection only increases the die by 0.09 mm2 per cu, so on die shots we would never know as 4 mm2 is pretty hard to detect. Its likely the mesh interconnects for all the cache would increase wafer cost and more layers rather than die size.
So, Ps5 MIGHT have L1 infinity cache, only if the technology was available in time to include it in the design.
We know Ps5 wont have much L2, as the die size says we will be lucky to get 8 MB or 16 MB being overtly optimistic.
Cerny certainly would of gone for it as the cache scrubbers, coherency and Sony patent on data handling of pixel vertices to shaders tells us it was a strong Sony focus point.
Also note the 50 % BVH improvement, I guess we have seen some ps5 BVH ray tracing already, but its not a given.
Will XSX have infinity cache
Just as much chance as Ps5 to be fair. Hotchips did not show it, but MS could not talk about L1 at hotchips. The block diagram showed no common L1 but again it could of been altered to adhere to the NDA.
Playing devils advocate, why do a hotchips and not show the major IPC improvement ? I guess all will come clear soon.
MS will almost certainly been offered the technology like Sony, and again its a timescale thing, did console makers have time to include it ?
The other confusing aspect of MS is rhe XSX chip is dual purpose as a server, and running 4 instances it may be arranged to have L1 private as no point in sharing L1 cache among 4 different games.
But L1 ininity cache would add about 56 x 0.09 mm2 to the die, so we would not know unless MS tell us. We know the L2 cache is 5 MB, so just scaled by PHY.
We have not seen much ray tracing BVH running in game on XSX, except minecraft but not sure if that was playable and how it ran.
Lets hope both consoles have infinity cache, at least L1, and BVH higher capacity meaning we will see some ray tracing in next gen on consoles.
Last edited:
