Is PS4 hampered by its memory system?
- Thread starter oldschoolnerd
- Start date
- Status
I remember there has been a discussion here about DDR3 vs GDDR5 and the general consensus and data reveals that there is no significant difference of latencies. IIRC GDDR5 has relatively higher clock latency but GDDR5 also runs at a much higher clocks, in the 5.5~7Ghz range while current DDR3s run in the 1.6~2.3 Ghz range. If you factor in the higher clock latencies with the actual clocks themselves together they cancel each other out. If you measure the time itself I think they are generally the same.
Here's the info.
http://www.hynix.com/datasheet/pdf/dram/H5TQ1G4(8_6)3AFP(Rev0.1).pdf
http://www.hynix.com/datasheet/pdf/graphics/H5GQ1H24AFR(Rev1.0).pdf
Examples here are
GDDR5 timings as provided by Hynix datasheet:
CAS = 10.6ns
tRCD = 12ns
tRP = 12ns
tRAS = 28 ns
tRC = 40ns
DDR3 timings for Corsair 2133@11-11-11-28
CAS = 10.3ns
tRCD = 10.3ns
tRP = 10.3ns
tRAS = 26.2ns
tRC = 36.5ns
So latency wise you probably shouldn't see a disadvantage of GDDR5 as they're in the same ballpark.
Last edited by a moderator:
oldschoolnerd
Newcomer
I don't think the split achieves that, I think it stops the GPU monopolising all the bandwidth. I think there are two measures getting confused. One is the bandwidth, one is the latency of any given request. Each cycle the memory bus can only do one thing. Service one client. Switching between them has a latency cost. I'm not sure there is a limit on that cost (eg 150 vs 20). I could be wrong....
All of this applies equally to the contention between the 8 CPU cores. You have 9 entities trying to access the memory at the same time, but 8 of them have priority over the GPU. I think contention between cores should logically dwarf any impact from the GPU, and it still doesn't seem to be an issue. The memory arbitration is probably designed to deal with this.
oldschoolnerd
Newcomer
Ignore any raw latency difference then, it's the contention that induces additional latency.
pMax
Regular
I have said something different: I have said to group threads that access similar data in the same cluster, in order to maximize the potential good effects of shared L2.
The dev should, on the other hand, optimize the code on the opt stage to keep a coherent access pattern (easier cache reuse among threads in same cluster?).
It might actually reduce average accesses to RAM from CPUs, due to average better local data coherency.
oldschoolnerd
Newcomer
So are all the GPU cores making memory requests funnelled through one aggregator? Makes sense to do that. But it's going to be very greedy...and still there is the problem I suggested in the OP. When the memory controller is faced with arbitrating a conflict does it interrupt the in-flight GPU request, or does it make the CPU wait? Option one has to lead to increased contention, option 2 leads to increased latency for the cpu. ?
oldschoolnerd
Newcomer
I wasn't really saying differing latencies were a big factor, more of a contributing factor. To be honest I don't know what those KPIs mean. But I can see 10% ish in most of them, which in a stall situation could add 10% to the bottom line...which when you have logical dependencies in your threads would have more of an impact.
Having said that, if the latencies were identical my point still stands ...the more simultaneous requests made to the memory bus, the more contention. The PS4, without a dual pool configuration has loads of requests hitting their single pool.
Perhaps there are some oddities with gddr5 that were not expected by AMD or Sony. Why else is AMD backing off its gddr5 based Kaveri PC apu?
There is that too, but I'm just looking at skyrim with 4k texture mods which can run @ 30fps on a 7850 and I ask why do so many ps4 game have not impressive textures? We all know very very high textures res on PCs only has a moderate impact on gpus.
There is that too, but I'm just looking at skyrim with 4k texture mods which can run @ 30fps on a 7850 and I ask why do so many ps4 game have not impressive textures? We all know very very high textures res on PCs only has a moderate impact on gpus.
Last edited by a moderator:
oldschoolnerd
Newcomer
It doesn't. But the more simultaneous requests you make the more contention you have. If you'd found a way to get 3/4 of your requests going to a separate pool, you'd have less of a problem in the first place...Why does contention induce additional latency just for PS4?
oldschoolnerd
Newcomer
It might do, but it's complex to do hey? And if you were aiming to squeeze every last drop of performance, you'd be doing that anyway, whatever the platform.
Instead of looking at it on a bandwidth basis, look at it from a commandclock/cycles basis. How many memory cycles (ballpark range) would something that saturates a 7850 bandwidth (like Skyrim with 4k texture mods) take?
Are the remaining comandclock/ memory cycles sufficient to feed the latency sensitive cpu with a tiny cache? If not they'd have to compromise on feeding the gpu in a cycles basis.
Are the remaining comandclock/ memory cycles sufficient to feed the latency sensitive cpu with a tiny cache? If not they'd have to compromise on feeding the gpu in a cycles basis.
Last edited by a moderator:
It means you've given CPU a budget to work with. This then means you can think about priority, latency and what not. Every 'frame' (clock cycles equivalent) you can say: prioritise up to x cycles for CPU, then service GPU. It's probably a huge oversimplification, but it seems ... Unlikely to say the least that there wasn't a clear design decision that was made here.
And incidentally, I thought the days where a memory bus does 'only one thing' were long behind us, but I am certainly no low level expert.
There are dozens upon dozens of research papers on this subject that presents dozens upon dozens of solutions to tackle this issue. None I have read have stated this is a difficult problem to overcome just that giving simple preference to the cpu or gpu when accessing memory isn't a robust solution.
That being said, its going to be hard to determine the effect of memory contention on the PS4, if you don't look at the fact that PS4 isn't a discrete cpu/gpu set up but probably a HSA design which minimizes data copying between different pools of memory normally employed by a discrete system. Thereby reducing bandwidth pressure, memory contention as well as latency.
That being said, its going to be hard to determine the effect of memory contention on the PS4, if you don't look at the fact that PS4 isn't a discrete cpu/gpu set up but probably a HSA design which minimizes data copying between different pools of memory normally employed by a discrete system. Thereby reducing bandwidth pressure, memory contention as well as latency.
Last edited by a moderator:
There are dozens upon dozens of research papers on this subject that presents dozens upon dozens of solution to tackle this issue. None I have read have stated this is a difficult problem to overcome just that simply giving preference to the cpu or gpu when accessing memory isn't a robust solution.
That being said it going to hard to determine the effect of memory contention on the PS4, if you don't look at the fact that PS4 isn't a discrete cpu/gpu set up but probably a HSA design which minimizes data copying between system and video ram needed by discrete system. Thereby reducing bandwidth pressure as well as latency.
That being said it going to hard to determine the effect of memory contention on the PS4, if you don't look at the fact that PS4 isn't a discrete cpu/gpu set up but probably a HSA design which minimizes data copying between system and video ram needed by discrete system. Thereby reducing bandwidth pressure as well as latency.
What is this thread, internet warriors have discovered what game devs and Sony/AMD engineers have not? Maybe Sony found some holes they can access to reduce contention?
What is this thread, internet warriors have discovered what game devs and Sony/AMD engineers have not? Maybe Sony found some holes they can access to reduce contention?
Why are you saying this? I don't get it. We are the last to discover it on earth.
We are discovering it long after the Sony/AMD engineers have complete knowledge and understanding of it. Long after game studios have discovered it. We are the last on earth to discover it, and its just a result of all the gameplay videos we've seen from all the games.
- Status
