PlayStation 4 (codename Orbis) technical hardware investigation (news and rumours)
- Thread starter Love_In_Rio
- Start date
- Status
No, it's showing the combined total bandwidth of the CPU+GPU decreases disproportionally as you use more CPU bandwidth.
Actually it says exactly that right on the slide, heh.
Repost of the slide since I just saw it posted on GAF anyway
No where does it say it's 140GB/s & not 176GB/s it's just a graph used to show that the more bandwidth the CPU use the less the GPU will have.
There were 3 different virtual paths in the leaked docs:
- CPU to "CPU memory" that the GPU doesn't know about.
- CPU to shared memory which is also buffered by the GPU.
- CPU to shared memory with some other variant on buffering.
So, it's possible that they're talking about shared memory access (basically how sharing the caches hurts performance).
It's also possible that this is simply an overhead of 2 different clients trying to use the same bus.
As it's an AMD chip, that would probably be similar to a graph for any AMD APU?
Yes! and nowhere on that graph does it say the memory bandwidth is now 140GB/s & not 176GB/s.
even if the GPU use only 80GB/s it doesn't change the fact that the GDDR5 is 176GB/s.
AMD HD7870 GPU's have a max memory bandwidth of 153.6GB/s even with the 2.56TFLOPS GHZ Edition. What exactly make people think that because the PS4 GPU is shown using 140GB/s that the GDDR5 has somehow changed from 176GB/s to 140GB/s?
http://www.eurogamer.net/articles/digitalfoundry-the-complete-xbox-one-interview
On XB1 the effective bandwidth of DDR3 is 50-55 GB/s while the theoretical bandwidth is 68 GB/s. For GDDR5 the theoretical bandwidth is 176 GB/s and the effective bandwidth is 135-140 GB/s (according to the graph it's around 135 GB/s). GDDR5 efficiency is around 75-80 percent.
No, the PS4 APU has an internal bus dedicated to the CPU, which is hooked up to the internal memory controller crossbar. That doesn't make some memory bandwidth dedicated for anything per se. It may be that the memory controller prioritizes CPU memory accesses though (due to latency concerns and whatnot), which in reality would have pretty much the same effect.
It would be pretty silly to set aside N gigabytes per second of bandwidth for CPU use if the CPU never even uses that much...
This is just bizarre pedantry. Look at the fking graph, what does it show you? Seriously, just look at it. Tell us what you see.
I'm not sure if you're being obtuse on purpose or if it simply comes natural to you. That graph shows an effective bandwidth representation, alright? A real-world situation, not a paper spec number. 176 = paper spec, graph = approximation of real world performance.
MetalSpirit
Newcomer
Heres my take on it...
176 GB/s is maximum theoretical Bandwith, something you can only achieve on peaks, but not on a sustained way.
if you read the Digital Foundry interview with the Xbox architects, they claim that the DDR3 effectiveness is somewhere from 70 to 80%. That means a sustained speed of 68*0,7=47,4 GB/s minimum, and a maximum of 68*0,8=54,5 GB/s.
Even the allmighty ESRAM with 204 GB/s theoretical Bandwith is referred by Microsoft on the same interview as reaching a maximum between 140 and 150 GB/s, and these under optimal conditions.
Usually DDR3 is a CPU memory, and GDDR5 a GFX memory. But on the PS4, the GDDR5 is also a CPU memory.
So, what if the effectivess is the same as DDR3?
That would mean 80% effectiveness on GPU only usage, and 70% when CPU is at it´s peak.
That would translate to the following real world speeds of:
176*0,8=140,8 GB/s with GPU only usage, and 176*0,7=123,2 GB/s with CPU and GPU usage at its peek (if you remove the 20 GB/s of the CPU this would leave about 100 GB/s to the GPU as also seen on the chart).
This would explain the values, but i´m just theorizing here!
Any saying on this? I´m no expert on this matter!
176 GB/s is maximum theoretical Bandwith, something you can only achieve on peaks, but not on a sustained way.
if you read the Digital Foundry interview with the Xbox architects, they claim that the DDR3 effectiveness is somewhere from 70 to 80%. That means a sustained speed of 68*0,7=47,4 GB/s minimum, and a maximum of 68*0,8=54,5 GB/s.
Even the allmighty ESRAM with 204 GB/s theoretical Bandwith is referred by Microsoft on the same interview as reaching a maximum between 140 and 150 GB/s, and these under optimal conditions.
Usually DDR3 is a CPU memory, and GDDR5 a GFX memory. But on the PS4, the GDDR5 is also a CPU memory.
So, what if the effectivess is the same as DDR3?
That would mean 80% effectiveness on GPU only usage, and 70% when CPU is at it´s peak.
That would translate to the following real world speeds of:
176*0,8=140,8 GB/s with GPU only usage, and 176*0,7=123,2 GB/s with CPU and GPU usage at its peek (if you remove the 20 GB/s of the CPU this would leave about 100 GB/s to the GPU as also seen on the chart).
This would explain the values, but i´m just theorizing here!
Any saying on this? I´m no expert on this matter!
Last edited by a moderator:
No, the PS4 APU has an internal bus dedicated to the CPU, which is hooked up to the internal memory controller crossbar. That doesn't make some memory bandwidth dedicated for anything per se. It may be that the memory controller prioritizes CPU memory accesses though (due to latency concerns and whatnot), which in reality would have pretty much the same effect.
It would be pretty silly to set aside N gigabytes per second of bandwidth for CPU use if the CPU never even uses that much...
This is just bizarre pedantry. Look at the fking graph, what does it show you? Seriously, just look at it. Tell us what you see.
I'm not sure if you're being obtuse on purpose or if it simply comes natural to you. That graph shows an effective bandwidth representation, alright? A real-world situation, not a paper spec number. 176 = paper spec, graph = approximation of real world performance.
That graph isn't changing anything the 176GB/s bandwidth is still the same as it was before. how much they are able to get out of it doesn't change the specs that we already knew.
It's still the same 176GB/s peak memory bandwidth. just because the CPU/GPU isn't using 100% of that bandwidth doesn't change it from being 176GB/s peak memory.
MetalSpirit
Newcomer
That's normal on all memories and on all systems. Theoretical are what is spoken as reference, but real life sustainable speeds are diferente. And smaller.
The graph doesn't provide information on the methodology used for the test, or whether it was just some kind of average over a range of workloads.
One can find all the reviews for CPUs on the net and check out the Sandra bandwidth numbers and see that there's a significant shortfall from peak to the benchmark results, and benchmarks aren't realistic workloads.
Losing 20% isn't out of the ordinary for memory controllers designed by Intel, going by some Ivy Bridge tests.
AMD is typically significantly worse, so the PS4's utilization being in the range of Intel would have been a feather in AMD's cap.*
*However, the Sandra results are for the CPU and don't involve a high-contention scenario like this. When involving the CPU, the percentage is much worse, which is difficult to compare since nobody benchmarks Sandra while running a graphics test. However, utilization being below average once an AMD CPU has to make use of bandwidth isn't unusual.
The 80% in the GPU only PS4 scenario isn't equivalent. In certain scenarios, parts of the GPU can hit 90% or more memory utilization. In more complex scenarios it is likely to be lower because at some point we expect to GPU to do more than run its ROPs.
D
Deleted member 11852
Guest
True but this was the message that Sony was sending to PlayStation 4 developers. Unless Sony are in the habit of trying to mislead developers for kicks, I think we can safely assume that the point of this slide is to convey 'average' real world cases.
- Status
