-
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.
Oxide: Nvidia GPU's do not support DX12 Asynchronous Compute/Shaders.
- Thread starter Arkanius
- Start date
Spot on. And while I get the enthusiasm of the PC AMD cards owners and the hope that finally the drivers won't hold the cards back anymore, I'm amazed about the PS4 owners gloating over this. I mean I'm sure the DirectX 12 will greatly improve the performance of the AMD APU in ... Oh, wait.
You know the PS4 uses an API equal to DX12 and Vulkan right?
Dictator93
Member
It has had access to a low level API since its inception. I do not see your point as well.
JeffG
Member
Regardless of what API a machine has, the more developers that use async compute, the more it will show up everywhere. (on machines that support it.)
async compute is not dependent on DX12
serversurfer
Member
Well, it's not obvious to me, so again, what performance penalty? Do you have any evidence the memory controllers in the APUs are being overloaded? Why would the be, assuming they're designed to handle those requests?
What are you even talking about? Nobody said any of that stuff.
Indeed, but I'm starting to wonder if it's all you have
No, that's not really what it's like. Like I said, it gives more flexibility to improve scheduling. Will that benefit every workload? Obviously not, but it will improve some, particularly the larger and/or more complex loads. But like I said, it's just a tool, so the benefit gained will vary from developer to developer and from project to project.
We've seen the benefits. Why do you continue to pretend there's no evidence for it? ><
Sorry, I'm not sure what you're trying to say here. Are you saying that even without hUMA, you can sometimes perform some of the same techniques to a more limited degree? Again, that was never disputed. The argument is that hUMA makes these techniques easier to leverage by reducing their overhead, meaning you can use them in more places, and to greater benefit.
I have, but you continue to pretend I haven't, while ignoring my requests that you explain what it is you're looking for. It's starting to seem that the only thing that will satisfy you is a hen's tooth.
What? I never called you stupid. =/
That's not the point. Is that API just launching now?
JeffG
Member
The PS4 api has been there since day 1.
Developers tend to code to the lowest common denominator (usually due to budget/time constraints)
The easier it is for them to adopt "best practices", the more you will see them implemented.
The API is the details of the implementation, the first step is for the developers to re-structure the code to use the new features appropriately.
There were examples given in this thread of games that already use async compute, so it's not like if the devs wanted to use that didn't. I agree though that more games using it might mean more games optimised for consoles too.
But 1st party developers had access to it from the start. So again. There's no reason for gloating, the performance is what it is there already.
mario_O
Member
Interesting read from the Nvidia reddit.
https://www.reddit.com/r/nvidia/com...loper_says_nvidia_was_pressuring_them/culu5uu
https://www.reddit.com/r/nvidia/com...loper_says_nvidia_was_pressuring_them/culu5uu
Dictator93
Member
If a game made tons of async compute stuff and completely ignored NV specific stuff, then yes, an AMD card could do better in a game.
That is... assuming now other performance /graphical features favour NV hardware.
The plot thickens..
DouglasteR
Member
Nope, mostly early ps4 games didn't use it or barely used it at all.
Not even the 1st party games?
As far as I know, yes.
Why would that be true? It's their API. Why wouldn't they be using it? I can't buy that, sorry.
Edit: if the post above is true, what would push the 1st party developers to use it in the future? It seems to be not such an easy win if only Infamous used it.
icecold1983
Member
i dunno, to me he basically tried his best to positively spin that nvidia hardware sucks at async
Because it takes time, and lots of those games started development before the specs were finalized.
Battlefield 4 and Infamous Second Son started also around the same time.
Not sure what your point is.
Optimizing games for pc and xbone using dx12 will push using async, which will translate to async being used in more ps4 ports as well. As for why 1st party devs haven't been doing it in the past, probably because it's still relatively new, and their experience in building pc games under dx11 never used async.
So it just isn't something devs are used to, but it will become commonplace as dx12 takes over on pc and xbone.
That the excuse of early start of development doesn't hold against the examples. We are already one year and a half later compared to Infamous and according to your statement most of the games don't use the new revolutionary thing. On top, Infamous SS and Battlefield 4 are not some top models in terms of graphics and performance. So the question still remain, what will push the developers to use it more?
So considering the development cycles, with some luck it will become mainstream in the last year of life of the current consoles.
Above.
What?
As async becomes industry standard, 1st party ps4 devs will follow suite to achieve more performance. Why is it a bad thing if we haven't seen everything the ps4 can do?
Edit oh I didn't see your edit. Yeah that might be the case. But i think we will more and more games using it starting next year.
Well those games probably use it lightly, like ashes does.
I expect the next trials game, for example, to be a good showcase. According to the senior graphics programmer they're getting more than 30% more performance with it.
dr. apocalipsis
Banned
MDolenc at Beyond3D did a small benchmark.
It throws some interesting figures:
- As you can see, Maxwell is up to 5 times faster if you dont fill all its 31 queues, each time you do that you incur on a time penalty.
- GCN 1.1 can eats up to 128 chunks without penalty, but starting at a much slower speed that will slow down rendering pipeline even at the lightest compute load.
So, at the end of the day, Nvidia will still be faster rendering beatiful graphics with not too stressful compute loads, and GCN will be better at not that hot graphics but with massive compute loads.
Faster but narrower vs slower but wider. Fancy ports car vs ugly truck, same story as ever.
tl:dr
Light async load:
Maxwell>GCN
Heavy async load:
GCN>Maxwell
This thing might be useful just on PS4 thanks to it having a shitty CPU and spare ACEs, but not that much on PC (cause of more resources) or XOne (cause of an already exceeded GPU)
It throws some interesting figures:
- As you can see, Maxwell is up to 5 times faster if you dont fill all its 31 queues, each time you do that you incur on a time penalty.
- GCN 1.1 can eats up to 128 chunks without penalty, but starting at a much slower speed that will slow down rendering pipeline even at the lightest compute load.
So, at the end of the day, Nvidia will still be faster rendering beatiful graphics with not too stressful compute loads, and GCN will be better at not that hot graphics but with massive compute loads.
Faster but narrower vs slower but wider. Fancy ports car vs ugly truck, same story as ever.
tl:dr
Light async load:
Maxwell>GCN
Heavy async load:
GCN>Maxwell
This thing might be useful just on PS4 thanks to it having a shitty CPU and spare ACEs, but not that much on PC (cause of more resources) or XOne (cause of an already exceeded GPU)
MDolenc at Beyond3D did a small benchmark.
It throws some interesting figures:
- As you can see, Maxwell is up to 5 times faster if you dont fill all its 31 queues, each time you do that you incur on a time penalty.
- GCN 1.1 can eats up to 128 chunks without penalty, but starting at a much slower speed that will slow down rendering pipeline even at the lightest compute load.
So, at the end of the day, Nvidia will still be faster rendering beatiful graphics with not too stressful compute loads, and GCN will be better at not that hot graphics but with massive compute loads.
Faster but narrower vs slower but wider. Fancy ports car vs ugly truck, same story as ever.
tl:dr
Light async load:
Maxwell>GCN
Heavy async load:
GCN>Maxwell
This thing might be useful just on PS4 thanks to it having a shitty CPU and spare ACEs, but not that much on PC (cause of more resources) or XOne (cause of an already exceeded GPU)
Why does AMD seem to do better at higher resolutions and nvidia better at 1080p? What are beautiful graphics? Gameworks effects that hit performance so hard you need an sli set up to run them at high resolutions?
MDolenc at Beyond3D did a small benchmark.
It throws some interesting figures:
- As you can see, Maxwell is up to 5 times faster if you dont fill all its 31 queues, each time you do that you incur on a time penalty.
- GCN 1.1 can eats up to 128 chunks without penalty, but starting at a much slower speed that will slow down rendering pipeline even at the lightest compute load.
So, at the end of the day, Nvidia will still be faster rendering beatiful graphics with not too stressful compute loads, and GCN will be better at not that hot graphics but with massive compute loads.
Faster but narrower vs slower but wider. Fancy ports car vs ugly truck, same story as ever.
tl:dr
Light async load:
Maxwell>GCN
Heavy async load:
GCN>Maxwell
This thing might be useful just on PS4 thanks to it having a shitty CPU and spare ACEs, but not that much on PC (cause of more resources) or XOne (cause of an already exceeded GPU)
Following the conversation, we may want to wait for an updated version of the benchmark.
[Fake edit] What dogen said. Looks like seppi is providing feedback on what to do and what to look for to get good data out of it. [/Fake edit]
serversurfer
Member
Well, not really. Again, this is just a tool, and it takes skill to use a tool well. Skill typically comes through practice, so really, Sony 1st-party just has more practice than anyone else. Dating back to the PS3, actually, because similar job packaging and scheduling was done for the SPUs.
So, improving throughout this generation and commonplace in the next generation. And that's a problem because ?
Oh, you're just trolling? heh You got me.
Who's that?
SunBroDave
Member
Disappointing, but not surprising. This is Nvidia efficiently investing in having the best DX11 cards, while also pressuring consumers to upgrade for DX12. It's a win-win for them.
So in the end, its all about:
That post is wrong about Async Timewarp.
Go back a few pages and see the Nvidia presentation on GameworksVR posted by Locuza. They're preempting other operations in the graphics pipe in order to do Async Timewarp, they're not using compute at all for this.
DonasaurusRex
Online Ho Champ
no both cards will do well with DX12 that much is clear its just the asyc compute feature that Maxwell doesn't have cannot take advantage of something in DX12. I have no idea if Pascal has the feature, and furthermore DX12 has more performance features than just async. From what i hear pascal is a monster anyway nothing to worry about here, but the good news is GPU's being used to a fuller extent and better performance for customers new and old. The gain by the 290X is wonderful news, that is the bump you're looking for great news.
Look, after reading what the EVGA guy wrote on Reddit and all that back and forth the only thing that I can say is time will tell. If AMD one ups NVIDIA then good on them and I guess that will give Nvidia the incentive to do better.
In the meantime, people who are looking to get rid of their 980ti's, how much?
In the meantime, people who are looking to get rid of their 980ti's, how much?
Lalalandia
Member
Yup async timewarp != async compute but the other points in that reddit post seemed reasonable enough, the games that show the biggest improvements from DX12 thus far are RTS titles that have huge numbers of drawcalls which most other genres don't really need. Whether the apparent weakness of NV at 'true' async compute with DX12 is seen in titles over the next 2-3 years is still an open to question to my mind.
dr_rus
Member
So I'll have to say it again it seems: ANY DX12 code support Asynchronous Compute IF it has at least one compute job queue in addition to the obvious graphics job queue. You don't have to do anything more that have two shaders in your code - one for graphics and one for compute. API+driver will detect this and will launch them either in parallel or serially depending on the h/w capabilities.
There seems to be a lot of misinformation around the topic, partly because a lot of info coming from PS4 which has a different API and a completely different h/w profile and optimization guidelines.
Except it's also completely misleading. All execution on modern vec1 SIMDs is done in a serial fashion, so there is no "8 roads with 8 lanes for trucks which can be used to move freely" but there is 8 roads with 8 lanes which are waiting to be picked from to the execution pipeline. The more you have - the higher efficiency of the execution you may achieve.
I've heard about Maxwell's queue granularity being worse than that of GCN before. The question is is it so bad that it's completely unusable or are we looking at cases in which GCN will be better and cases where Maxwell will be better? Another question is does Maxwell even need async compute to keep its utilization in DX12 at peak?
frontieruk
Member
My take on it ATM, keep your 980ti's, async is probably going to be used for improved lighting and AA which aren't going to over load draw calls, so you'll see nvidia ahead in most games as AMD will suffer from latency of getting workloads into the queues where NV is quicker at low levels of async compute. If only we could get benches from the fable beta
to confirm thisBuys 980Ti Friday, this news breaks today....
So it's time to bin it already. So the card will not be able to run anything on DX12 above 30FPS?
Yeah you and everyone else in this thread who are ready to throw out a 980Ti and replace it with a R9 290(because apparently now a R9 290 is equivalent to a 980Ti)
Go ahead and do it, you're making the right decision, because all benchmarks show that the 980Ti is just inferior to every AMD GCN 1.2 ever made.
http://www.extremetech.com/gaming/2...he-singularity-amd-and-nvidia-go-head-to-head
Foutrologue
Banned
My take on it ATM, keep your 980ti's, async is probably going to be used for improved lighting and AA which aren't going to over load draw calls, so you'll see nvidia ahead in most games as AMD will suffer from latency of getting workloads into the queues where NV is quicker at low levels of async compute. If only we could get benches from the fable beta
Dictator93
Member
My take on it ATM, keep your 980ti's, async is probably going to be used for improved lighting and AA which aren't going to over load draw calls, so you'll see nvidia ahead in most games as AMD will suffer from latency of getting workloads into the queues where NV is quicker at low levels of async compute. If only we could get benches from the fable beta
Fable beta would definiely be interesting. A shame it is locked behind NDA atm.
BTW, earlier you .gif posted my mentioning of Crysis 2 concerning tesselation but failed to respond. The tessellation myth in Crysis 2 was disproved years ago just so you know. By Cryengine devs themselves among many others. A shame that silly wccftech report taking pictures and videos in debug mode came out... they just id not understand how cryengine works...
I should point out that a 290 isn't even GCN1.2
Who cares? The benchmarks show that it CLEARLY blows that GreedTX 980 Ti out of the water in every category.
/s
I don't think so. Fable Legends uses UE4 which seemingly isn't doing anything interesting in this area atm.
It is only interesting as a test of a different dx12 workload.
I'm pointing out that those wanting to switch from a 980ti to a 290 aren't even looking at the best GCN gpus.