PlayStation 4 (codename Orbis) technical hardware investigation (news and rumours)
- Thread starter Love_In_Rio
- Start date
- Status
How come when I tried to talk about this last month you & other members tried to make it seem as if I was crazy & it wasn't possible?
You'll have to specify which posts you are referencing.
The ones that I can recall did not say the same thing as your quote.
The ones that I can recall did not say the same thing as your quote.
I was talking about the PS4 being able to run Graphics & compute at the same time without taking away from the graphics & I was posting links to prove that's what the people at Sony was saying & I even got a warning for talking about it because everyone else said I was wrong .
.
You’re looking at the 30,000 feet level. At that level, you couldn’t give any details.
They were looking at the 3,000 feet level. Someone posted a specific scenario where it could happen.
Note that the magic is still with the developer although the architecture and tools may make it easier to take advantage of these “gaps”.
Clarifying this will require going over the specific posts and the wording used.
The point I remember being debated was the claim that the full floating point throughput of the GPU could be used for graphics even if compute was ongoing.
Shifty Geezer's post is not saying the same thing.
No that's the conclusion that other people jumped to when I was talking about the quotes from Sony & I even pointed out what I was talking about.
http://forum.beyond3d.com/showthread.php?p=1723390#post1723390
Last edited by a moderator:
There seems to be some confusion here. I recall Shifty saying that a CU is unable to work on graphics and computer work at the same time. I recall patsu correcting him on that and providing a lot of us here with some education as to the way GCN works when concerning wavefronts.
http://forum.beyond3d.com/showpost.php?p=1724729&postcount=1259
http://forum.beyond3d.com/showpost.php?p=1724729&postcount=1259
They were talking about maintaining or exceeding the 1.84 peak FLOP count at that point. Cerny is more talking about a specific case of achieving/improving efficiency (by giving developers stats and control). It also entails how they see the CPU and GPU work together.
If we only look at the GPU attributes, we may miss the CPU story.
If we only look at the GPU attributes, we may miss the CPU story.
The post I thought of was earlier in the thread:
There were also a number of awkwardly bolded press lines that were later used as argumentation.
Ignoring that run of posts, the Gamasutra article quotes Cerny making clear the goals of the various tweak so reducing overhead and allowing compute and graphics loads to run well concurrently without excessively interfering with one another.
In any complex system with varying workloads, the impact each type of kernel has on the other is non-zero. The goal, as Cerny points out, is to make the interference as low as possible so that the concurrency is a net gain.
It's pretty straightforward to generate a scenario where either the compute kernel or graphics pipeline could have shaved a few cycles off of their execution times if they could use all of the GPU's resources. In the concurrent execution case, this is almost always not going to be the case, and any scenario where they do not impact each other at all is likely to be trivial or contrived. The PS4's stated goal is to make the sacrifice as small as possible, not zero.
rockaman
Regular
I thought he explained it well also. It seems when a graphics task is occupying a certain subset of X pipeline and where no other graphics tasks can fit in very well, he is giving a way to access that remainder for compute.
So it's like a complicated way of saying we're can use some of the excess unused portion of the GPU to get some other work done without disrupting anything too much.
So yea if you interpret it one way, the compute doesn't take away from the GPU rendering process in X situation. But at the same time the max throughput for the GPU as a whole still is 1.84 tflops no matter what you do. So compute still isn't necessarily being done free, only that there are some neat ways to fit it around the graphics processing sometimes. Sharing is caring and all that.
Sounds like a resolution of this whole discussion to me.
So it's like a complicated way of saying we're can use some of the excess unused portion of the GPU to get some other work done without disrupting anything too much.
So yea if you interpret it one way, the compute doesn't take away from the GPU rendering process in X situation. But at the same time the max throughput for the GPU as a whole still is 1.84 tflops no matter what you do. So compute still isn't necessarily being done free, only that there are some neat ways to fit it around the graphics processing sometimes. Sharing is caring and all that.
Sounds like a resolution of this whole discussion to me.
That whole thread was a nonsense rather than a technical discussion. If that wasn't your intention then you have only your posting style to blame. Your conversational technique didn't include specific technical questions, but just spammed repetitious quotes and muddled one-liners and a constant reference to rumours with a disregard for their relevance when the hardware was explained to us. Rather than trying to understand what was happening and connect the dots, you were just pushing a view and not knowing when to let it drop as you weren't making any headway.
Technically, what Cerny describes is no different to GCN. The existing GCN platform can still find holes in the graphics work to fit in some compute workloads as I understand it. Sony have just added a lot more threads to allow for a more flexible pool of compute resources, with perhaps a few improvements to the existing GCN memory systems to save cache pollution.
Ignoring the efficiency of utilising these spaces (how much compute can gain use these GPU stalls), can you put an estimate figure (or rather, range depending on game) on how much idle time there can be on the ALUs in a frame? Like 1%, or 10%, or 50%? That's important to understand how much compute can be obtained from the GPU without impacting graphics, and indeed, how much wastage there is on a GPU not running compute.
itsmydamnation
Veteran
sounds good to me and fits other AMD designs pretty well.
According to a leak, I think that's all the CPU has available anyway.
If you're rendering a bunch of simple triangles into shadow maps it's possible 50% of compute is idle, but if there's heavy pixel shading the number will be much less. I can't give a number for that situation that would be any more than a guess. Note one of the customizations Cerny mentioned addresses compute + geometry heavy workloads.
- Status
