• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

VGLeaks - Orbis GPU Detailed - compute, queues and pipelines

codhand

Member
gpu_queues.jpg
.
 

mrklaw

MrArseFace
It looks like a pretty interesting customisation, would be interested to hear comments from those that can better understand it.

What is VShell, is that like the OS/system? Is it reserving GPU resources, or is the ring just allowing it to hold onto things ready to flip across when you hit the OS button for instance
 

Perkel

Banned
Seriously that RAM talk everywhere but not in ram thread is super annoying. Here we have some info about PS4 GPU and 3 out of 6 post are some shitty GDDR5 mame.

Looks like Cerny was right saying GPU was customized for compute. Exciting stuff if we will have proper physic acceleration in GPU.

Hair cloth particles "here we go !"
 

Kaako

Felium Defensor
Hey developers, how do you feel about this new diagram detailing the pipelines of the Orbis GPU?
 

Biggzy

Member
Seriously that RAM talk everywhere but not in ram thread is super annoying. Here we have some info about PS4 GPU and 3 out of 6 post are some shitty GDDR5 mame.

Looks like Cerny was right saying GPU was customized for compute. Exciting stuff if we will have proper physic acceleration in GPU.

Hair cloth particles "here we go !"

You can quite clearly see it in Knack.
 

Kinitari

Black Canada Mafia
I should really just dedicate a day to understanding 1/10th of what all this stuff means, you know really sit down and learn. Well that's not fair, I understand a lot of what it means, but not enough to really get a full picture.

One day.
 

demolitio

Member
I'm just excited to see the evolution of physics. I'm going to have a hard time not destroying shit in games that have good physics. :D
 

gofreak

GAF's Bob Woodward
This is so low level... it is a much more detailed view of the ACE stuff rumoured before though. I think 'better/easier utilisation during mixed compute/graphics work' is still the 20,000ft version here.
 

Biggzy

Member
you can see it too in the Havok demo

I used Knack as an example because it is an actual game.

With all due respect to Knack it didn't push that much particles on screen with physic in trailer not it had cloth or hair physic.

No. But the way the character dissembled and resembled all his pieces was definitely influenced by the GPU. Of course we can expect more down the line, but for a taster it isn't a bad start.
 

Osiris

I permanently banned my 6 year old daughter from using the PS4 for mistakenly sending grief reports as it's too hard to watch or talk to her
So more Queues means better load balancing? Am I reading this right?

Pretty much, these don't add performance directly, they make your use of resources more efficient by reducing stalls, keeping the CU's fed and speeding up context switching between general compute / gfx workloads.
 
- High Priority Graphics (HP3D) ring and pipeline

New for Liverpool
Same as GFX pipeline except no compute capabilities
For exclusive use by VShell

Is this AMD's hardware solution for 3D/stereoscopic gaming?
 

sangreal

Member
Is this AMD's hardware solution for 3D/stereoscopic gaming?

It sounded to me like something for the OS, etc to render without having to wait for the game but I don't know shit. So basically if there was something in the HP DRAW and the GFX DRAW queues, the "Arbitration" would give priority to the HP DRAW
 

benny_a

extra source of jiggaflops
Pretty much, these don't add performance directly, they make your use of resources more efficient by reducing stalls, keeping the CU's fed and speeding up context switching between general compute / gfx workloads.
There was talk about 100% efficiency GPU. Can anyone describe that this leak means in that context and what 100% efficiency means?

I assume it has something to do with memory bandwidth and FLOPs and having these things be adjusted so that neither part has an obvious bottle neck?
 

artist

Banned
In simpler terms, the CPU is dividing the calls based on the type (graphics vs compute) and feeding the GPU accordingly. The compute tasks get further funneled to the units that are idle and get better utilization. 8x8 granularity is a lot better versus 2x2 approach of GCN.

The graphics pipeline is obviously higher priority and doesnt need the scheduling of ACEs to distribute the load.

edit: It also looks like each ACE can also interface with the DMA directly.
 

AgentP

Thinks mods influence posters politics. Promoted to QAnon Editor.
It is kind of hard to do a 1:1 with the two flow charts. I hope someone can make heads or tails from these.
 

Biggzy

Member
There was talk about 100% efficiency GPU. Can anyone describe that this leak means in that context and what 100% efficiency means?

I assume it has something to do with memory bandwidth and FLOPs and having these things be adjusted so that neither part has an obvious bottle neck?

Think of it as the GPU is able to reach it's full potential.
 

demolitio

Member
Destruction in Battlefield 4 think about it.

Yea, that's actually one of the reasons I want BF5 to return to WWII. True destruction in a war that was pretty much all about destruction would be epic, especially if they brought naval battles back. Imagine a destroyer shelling a village along the coast with all of its weapons...MMMmmmm.

I hope more MP games are willing to take the risk on destruction and realize the potential instead of just thinking it would ruin the flow of a map.
 

tipoo

Banned
Do we really have to make the RAM thing the next Cell meme?

So it doesn't look exactly like the GCN architecture, a few enhancements in there. I wonder if we're looking at a preview of the GCN 2.0? Like how the 360 GPU had some HD2000 series features.
 

Elios83

Member
These things are really advanced and specific, but basically it confirms that there are specific customizations to the standard GCN cores aimed at balancing the workload so that everything is more efficient.
 

gofreak

GAF's Bob Woodward
There was talk about 100% efficiency GPU. Can anyone describe that this leak means in that context and what 100% efficiency means?

There's no such thing.

Or no such one thing. And you can verge into semantics here. For example, an algorithm might be considered by some to be '100% efficient' when it's as optimised as it can be, though if you were looking at compute resources its executing on, it may not be firing all cylinders all of the time.

So it depends what you're measuring by.

What 'improved efficiency' means in relation to this is, that by having more threads ready, the gpu can switch faster to another thread if the current one stalls, and switch more readily between graphics and compute threads. Apparently with regular GCN there was a certain non-negligible penalty in doing that kind of switching between different workloads. So much so that AMD recommended doing compute and graphics on separate chips inside a system where available.

Or at least that's one takeaway, I think. It doesn't mean '100% efficiency' though. It's 'just' another tool for potentially improving utilisation and thus helping code thrown at the GPU finish faster.
 

benny_a

extra source of jiggaflops
Think of it as the GPU is able to reach it's full potential.
I was hoping for more insight in how that is calculated because I think 100% efficiency is obvious, but how one gets to make that statement is not.

There's no such thing.
It came up with the Durango GPU. It's at ~1.2Tflops but 100% efficiency and the 360 was rated at 60% efficiency.
 

Durante

Member
This is really interesting, probably the most interesting thing on VGLeaks so far.
It should allow better GPU utilization, especially with mixed graphics/compute workloads. What I don't understand yet is how the load balancing between the individual queues/pipelines will be performed or controlled.

Is this AMD's hardware solution for 3D/stereoscopic gaming?
Nah, it has nothing to do with that. It appears to allow the OS higher priority access to the GPU.
 

Biggzy

Member
I was hoping for more insight in how that is calculated because I think 100% efficiency is obvious, but how one gets to make that statement is not.

I am no tech expert, but making sure the GPU is not stalling and doing nothing for a number of cycles is probably one way of increasing efficiency. I am sure someone will be able to give you a far better answer than that.
 
Top Bottom