Digital Foundry Article Technical Discussion [2025]

1742548528771.png


The issue isn't one of less BW for GPU as that BW is consumed by CPU, but that the proportion of BW reduction is greater than the CPU uses. It's an issue of RAM topology and access, seemingly with a high overhead when the CPU accesses the RAM. It's unclear whether the type of access - larger, less frequent accesses or frequent, small accesses - impact GPU-available BW differently, but I expect so.

Thus the question is more if there have been improvements so the overhead of a shared RAM pool on PS5 and going forwards is less? We've only seen this for PS4 AFAIK. Does that mean it was a PS4 specific problem, or have other consoles and shared-RAM devices experienced the same? Ideally the total BW will remain static and it'll yust be divide across accessing processors.

An investigation into Ubisoft's The Crew in 2013 revealed there's a degree of allocation here; the graphics component has its own, faster 176GB/s memory bus labeled Garlic, while the CPU's Onion has a peak of 20GB/s via its caches. Even so, there's a tug-of-war at play here - a balancing act depending on the game.
Is it the Onion/Garlic structure imposing this limit, and would other vegetables provide a better solution?
 
5Pro Analysis from Ollie
I couldn’t disagree more with this. The only difference is RTGI, which while important, isn’t nearly enough to justify the praise. No PSSR, the resolution is just as low as on the PS5 performance mode, if not lower, and the graphics aren’t improved.

For a system with a faster GPU, 2 to 4 times faster rat tracing, and PSSR, you’d expect a lot more.

Really meh upgrade.
 
I couldn’t disagree more with this. The only difference is RTGI, which while important, isn’t nearly enough to justify the praise. No PSSR, the resolution is just as low as on the PS5 performance mode, if not lower, and the graphics aren’t improved.

For a system with a faster GPU, 2 to 4 times faster rat tracing, and PSSR, you’d expect a lot more.

Really meh upgrade.
Unfortunately, a faster GPU, 2-4 times faster RT, PSSR, can all be bound by available bandwidth.

I would say that It’s certainly says a lot about the cost of RTGI if you don’t do it the way that Metro does it. We have seen many games attempt RTGI on consoles and it’s just very heavy.
 
AC shadows is great 40 fps with RTGI on base PS5 and XSX with virtual geometry. This is great. This is above what I was thinking we will see with RTGI.

And 60 fps with RTGI and virtual geometry on PS5 Pro is good too.

Maybe we will see other game with virtual geometry and RTGI with better performance but for the moment this is not the case.

This is great to see this level of geometry, nearly no geometry popin with this level of Global illumination lightning.

This is not PS4/Xbox one level of geometry with RTGI. This is a real curent gen title.
 
Last edited:
AC shadows is great 40 fps with RTGI on base PS5 with virtual geometry. This is great. This is above what I was thinking we will see with RTGI.

And 60 fps with RTGI and virtual geometry on PS5 Pro is good too.

Maybe we will see other game with virtual geometry and RTGI with better performance but for the moment this is not the case.

This is great to see this level of geometry, nearly no geometry popin with this level of Global illumination lightning.

This is not PS4 level of geometry with RTGI. This is a real curent gen title.
And with destructible environments as well. This isn't just a good job on the GPU side, but on the CPU side too at 60 fps on Pro (would be 60 on base probably but the resolution would maybe get too low).
 
The issue isn't one of less BW for GPU as that BW is consumed by CPU, but that the proportion of BW reduction is greater than the CPU uses. It's an issue of RAM topology and access, seemingly with a high overhead when the CPU accesses the RAM. It's unclear whether the type of access - larger, less frequent accesses or frequent, small accesses - impact GPU-available BW differently, but I expect so.

Thus the question is more if there have been improvements so the overhead of a shared RAM pool on PS5 and going forwards is less? We've only seen this for PS4 AFAIK. Does that mean it was a PS4 specific problem, or have other consoles and shared-RAM devices experienced the same? Ideally the total BW will remain static and it'll yust be divide across accessing processors.

With 3D V-cache it's indeed possible to add some cache to the CPU later without hitting the die size limit, but right now 3D V-cache is probably more expensive than just making a larger die (depends on how large the cache is, of course).

So the question goes back to whether the shared memory architecture is sustainable. I think the jury is probably still out, but I do think it's probably more likely to go back to some sorts of traditional two memory pool architecture, as the memory requirement for CPU and GPU are basically the opposite (CPU likes low latency but GPU loves high bandwidth).

The benefit of a shared memory pool is that it's easier to access the memory from both CPU and the GPU, without something in the way. There's also less duplication of data. This is great when you need the CPU to do some process before sending the data to the GPU (e.g. decompress the image data), but now it seems that the GPU is going to be capable of doing many of these tasks. With a proper architecture design, it's not hard to stream data directly from the storage to the GPU memory (e.g. DirectStorage), so it's also less a bottleneck than before.
 
And with destructible environments as well. This isn't just a good job on the GPU side, but on the CPU side too at 60 fps on Pro (would be 60 on base probably but the resolution would maybe get too low).

Hair strand system too...

Tons of other improvement compared to other AC games but seen on other Open world like procedural time of day/weather, great cloud shystem, great particle system and so on.
 
The issue isn't one of less BW for GPU as that BW is consumed by CPU, but that the proportion of BW reduction is greater than the CPU uses.
Thanks Shifty that's exactly the image I was talking about and maybe I worded it wrong saying it doesn't scale linearly but that's exactly what I was trying to say. The amd powered handhelds could maybe help see if other devices suffer the problem, I'm not sure if the different ram (gddr vs lpddr) would have any effect.

As for the vegetables if we don't use onions there will be less tears, and maybe no garlic so no one gets garlic breath. I vote for carrots, the old wives tale says it helps eye sight so maybe it will help get the visuals sharper ;)
 
View attachment 13347


The issue isn't one of less BW for GPU as that BW is consumed by CPU, but that the proportion of BW reduction is greater than the CPU uses. It's an issue of RAM topology and access, seemingly with a high overhead when the CPU accesses the RAM. It's unclear whether the type of access - larger, less frequent accesses or frequent, small accesses - impact GPU-available BW differently, but I expect so.

Thus the question is more if there have been improvements so the overhead of a shared RAM pool on PS5 and going forwards is less? We've only seen this for PS4 AFAIK. Does that mean it was a PS4 specific problem, or have other consoles and shared-RAM devices experienced the same? Ideally the total BW will remain static and it'll yust be divide across accessing processors.


Is it the Onion/Garlic structure imposing this limit, and would other vegetables provide a better solution?
Because GDDR6 has 2 independent, 16bit buses per chip, as opposed to a single 32bit bus like GDDR5, I always wondered if that helped alleviate some of the memory contentions that PS4 had on the newer consoles. In theory, because the memory controller has twice as many buses to fulfill requests, smaller memory accesses from the CPU that underutilize the bus wouldn't effect total bandwidth as much.

It would interesting to test that Chinese console that was PS4Pro-esque with GDDR5 versus one of those rejected PS6 SoC PCs
 
Last edited:
Back
Top