I don't think that's going to happen anytime soon.
There seem to be a lot of hyperbole surrounding one alpha test of single game engine. Doubt Nvidia with their mountain of cash would sit on their laurels and let AMD have an edge over them.
-
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
I don't think that's going to happen anytime soon.
There seem to be a lot of hyperbole surrounding one alpha test of single game engine. Doubt Nvidia with their mountain of cash would sit on their laurels and let AMD have an edge over them.
There is this fantastic ELI5 explanation on a Reddit thread
https://www.reddit.com/r/pcgaming/comments/3j1916/get_your_popcorn_ready_nv_gpus_do_not_support/
https://www.reddit.com/r/pcgaming/comments/3j1916/get_your_popcorn_ready_nv_gpus_do_not_support/
R10Neymarfan
Member
Yep they'll just gameworks every popular DX12 game with their blood money
dr. apocalipsis
Banned
So? That doesn't changes the fact that both GPU and CPU have to battle with a single memory controller for memory access. That alone will leave real world memory bandwidth far away from theoretical figures, not even talking about how queues will hurt CPU performance every time you need to take some data out of the cache.
Also, 'physically', memory chips takes their time to be ready to write or read, losing several cycles each time they need to change state. So shared memory has it's own advantages, but a discrete buffer will be always better for raw performance.
It's not like industry is dumb enough to overlook this until Sony invented shared memory for a cheap performer console.
Well, if you run out of local RAM you will have to page, that will hurt performance alone just with the interface overhead. You shouldn't be using PCIe like that to start with.
With single top end cards setups, the difference between PCIe2 x8 (4GB/s) and PCIe3 x16 (15.8~) is barely a 5%, most of it coming from the improvements made to the interface efficiency, reducing overhead from 20% to a measly 2%. Sure, some BIG SLI/Cross configurations might get better figures on some benchmarks, since the traffic between GPUs increase wildly the bandwidth usage, but those are minor extreme situations.
serversurfer
Member
I don't understand why the advantage isn't immediately apparent to you. You know the thing that shows you what the next block is in Tetris? It helps you plan ahead and fill the space more efficiently, right? Now imagine you can peek eight blocks ahead, and choose any block you like to use next. That would probably make the game a lot easier, right? How exactly would you go about proving that? It's practically a different game now. It's not possible to play the old game in the new way, and if you play the new game but restrict yourself to the old rules, obviously you won't see any improvement. Hell, maybe you're dumb and the extra information and options just serve to confuse you.
So how are we to "prove" that having more queues is better? If a game is designed to use 16 queues managed by two schedulers, then no, running it on a 64-queue system with eight schedulers won't make much difference. If the game is designed to use 17 or more queues, then it won't run on the 16-queue system at all. Does that mean it's not a valid test, or does it prove the point?
No one ever claimed it was required. The claim is that it increases the cases where you're able to leverage GPGPU because it eliminates the copying. Your argument is a strawman.
You understand this isn't like a clock bump that has some repeatable and easily measured effect, right? Async compute is a new and better tool for developers that allows them to make new and better games. But yes, like any tool, it is what the craftsman makes of it, and it doesn't help you much if you leave it rusting in the corner of your garage. People are using the technique, and they're both improving utilization and trimming execution time with it. What more do you want?
Actually, it may be a wiring issue. Chafing can cause that sometimes.
Yes, let's criticize AMD because it took years for anyone to bother using their very useful tech. =/
SneakyStephan
Banned
Buys 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?
Yep
I suggest you throw it in a ditch somewhere.
Where the fuck did that come from? Why did you deliberately try to put words into my mouth? Weird.
The context of this thread is Maxwell 2 and its lack of async support in one early test from what I understood. Generally PC games' async future is somewhat dependent on DX12 uptake speed. Are there any games supporting it out on the market? So is DX12 practically relevant for consumers right now? Yes or no?
no, let's criticize AMD for having shitty DX10/11 performance in their drivers and focusing almost entirely on features that nobody was even able to utilize. it's good that mantle helped drive the industry toward the low-level API future we're slowly realizing but it's pretty much the only thing they've done well that i can remember from the last few years. in practical, real-world terms they've been vastly inferior to the CPU and GPU competition in almost every way.
Lalalandia
Member
It's like I've travelled back in time to the RivaTnT debates ('who even uses 32 bit colour?'), we have one test point and the debate has somehow spun off into AMD/Nvidia willy waving (with consoles as a side dish, da fuq?). Right now we have one dev who is using async compute reporting anomalies in how NV uses async compute that has led them to conclude it actually works in serial rather than parrallel. This could be a driver fault, it could be silicon, given how aggressive NV's PR is I can't imagine they'll be letting this lie.
If all is as reported your 970/980/Titan is still an amazing card, you may suffer form a performance deficit vis a vis GCN if the title you want to play uses async in the fashion that AoS does. Given that you'll probably have swapped to a newer NV card by the time the AAA get around to rewriting their stuff for DX12 I wouldn't sweat it.
If this is confirmed to be a thing via multiple other tests then we can say NV have sacrificed the future for the present (and vice versa for AMD) until then this just an interesting data point.
If all is as reported your 970/980/Titan is still an amazing card, you may suffer form a performance deficit vis a vis GCN if the title you want to play uses async in the fashion that AoS does. Given that you'll probably have swapped to a newer NV card by the time the AAA get around to rewriting their stuff for DX12 I wouldn't sweat it.
If this is confirmed to be a thing via multiple other tests then we can say NV have sacrificed the future for the present (and vice versa for AMD) until then this just an interesting data point.
You have to remember that consoles have AMD APUs and don't need to wait for DX12 to be relevant anywhere.
Devs have the possibility right now to take advantage of this. Will they?
But what if asynchronous compute isn't that groundbreaking? There you would be right by saying it's not really important. Maybe you should argue about that and not about when DX12 will be relevant (will it?).
It just sucks for people that upgraded to a 980 Ti expecting it to last a few years into the future. (Since it was marked as DX12 compatible and with an higher DX feature level than AMD)
I'm waiting for Pascal myself, but I'm afraid Pascal might have the same uArch design as Maxwell but with HBM.
I would love for Nvidia to start investing in a single uArch instead of designing a new one every 2 years. GCN investment by AMD is starting to pay off. Whenever there is a gain in a GCN card, 2-3 years of cards reap the benefits.
But lets be real: 980 Ti destroys everything right now. Next year forward with DX12 games coming up, that might not be the case anymore.
SpankmasterC
Member
What? AMD has always been competitive with nVidia. Sure there have been performance differences, but most of the times all cards were priced competitively. Let's not over low this and call every AMD card from the 5870 on shitty.
Some users at Beyond3D have tested the Async performance of both the 980 Ti and the 7870XT
980 Ti
7870XT
980 TI can't run both tasks in parallel, they are ran in a sequence order.
Apparently, this is a huge blow for VR performance (Not much into VR yet). It seems Nvidia have lag issues that create nausea due this latency of having to run both tasks in sequence.
I'd like to hear more about this. I have heard before that AMD is preferred for VR. But either a) this is suppressed knowledge or b) it is bogus information. Here on GAF, there are many Nvidia PCs built with VR in mind and no one ever says anything.
See the posts by zlatan:
http://forums.anandtech.com/showthread.php?t=2437630
Dictator93
Member
NV has not said a thing adn the dev seems uncertain. So... who knows.
TheAdmiester
Member
Not to mention that the 980Ti and its relatives are powerful enough to blow through games even without these benefits. Also, I'm sure it'll be a long time before DX12 games stop offering DX11 in their settings (hell, even GTA 5 still supports DX10), so it's not like there's a huge problem here even IF it's true.
Sebbbi says it gives them huge gains.
https://forum.beyond3d.com/posts/1866842/
https://forum.beyond3d.com/threads/dx12-performance-thread.57188/page-7#post-1868894
Seeing the investments AMD made towards HSA, it is likely that current cards implement Async Compute in a better way.
Having said that: It will be a generation or two, before this feature is required for stable framerates, so don't worry everybody. Your top of the line cards will play current games fine.
Nice.
serversurfer
Member
That was a pretty good explanation.
"Battle"? You've some evidence the memory controller can't handle the traffic from both chips? Seems like a lot of overhead is going to be reduced just by eliminating all the copy commands.
Sorry, what does this mean?
But again, shared memory will reduce the need for reads and writes, because it eliminates the shuffling.
Err, shared memory has been around for a while, and using it like this is sorta AMD's thing. So, you're just opposed to this because you hate Sony? =/
Sorry, I was just making a joke, because it seemed like you were misplacing the blame a bit.
I'm gathering that it isn't. What I don't understand is why AMD are to blame for that. People ignoring a technology doesn't serve as proof that it sucks. If anything, I would say that Mantle and DX12 serve as proof more people should've been paying attention to AMD for a while now.
I don't follow PC stuff, but it seems like AMD created an advantageous tech, and it was never leveraged in DirectX, which is what everybody uses for everything. So eventually they came up with their own substitute, at which point MS said, "Oh, yeah, we do that too."
What should AMD have done differently? Go it alone from day one instead of hoping/expecting support for their new tech in DX? I realize DX12 is new, but GCN launched in 2011. Four years later, the rest of the industry finally starts to catch up and they're laughing at the "slackers" who wasted their time inventing stuff? =/
More likely Microsoft and Sony both worked with nVidia and both companies fell out with them. Also Nintendo has been working with AMD since they bought ArtX and for reasons of compatibility - as well as a long-term relationship in good standing - they have no reason to switch to nVidia. And their APU offerings, of course.
Coulomb_Barrier
Member
It does make sense that AMD bet heavily on asynchrous computing and are only now reaping the benefits. Heck looking at it historically it seems like they created and pushed Mantle as a way of encouraging developers to use coding techniques that would show their GPUs in a better light. Didn't really gain any traction but Microsoft's DX12 has certainly put the whole thing into the spotlight.
Muppet of a Man
Member
So an AMD APU was the right choice for PS4 and X1. Sony and Microsoft redeemed for their CPU/GPU decisions.
diffusionx
Gold Member
I was planning to build a new PC this winter, with a 980ti. I had to push back the build until Spring for reasons, but after reading this I think it's the right decision. Actually, after the 970 deception, it'd be nice to give Nvidia the finger next time around.
Obviously it's not like tomorrow all games will go DX12 but I think it will be adopted much much faster than DX10/11 were. I am sure Nvidia will respond but it won't be until their next round of GPUs at the earliest.
Obviously it's not like tomorrow all games will go DX12 but I think it will be adopted much much faster than DX10/11 were. I am sure Nvidia will respond but it won't be until their next round of GPUs at the earliest.
My comment was simply about consoles' lacking horsepower compared to PCs with their much faster CPU+GPU combos. Those are bound to be a limitation in a game engine faster than having separate RAM and VRAM on PC (which was the argument in the post I replied to).
I am not saying that unified memory architecture is in some way worse, it would be nice to see it on PCs but that's not going to happen anytime soon.
DouglasteR
Member
I just bought 10 sacks of corn.
Lets see how Nvidia will respond to this.
Lets see how Nvidia will respond to this.
frontieruk
Member
Damn you, how can they fix this without that corn????
Ah crap, no problem. I'm really tired and perhaps in a bad mood so I kind of bluntly missed your joke
Why not, perhaps even their TDPs would be healthier if so..
Yes, we did, back on the first page...
Doubt the Developer is lying about them trading emails with Nvidia and Nvidia saying for them to disable the feature all together since it introduces delay with their uArch.
Nvidia silence on this matter is also pretty telling to be honest. Last time Nvidia did this was when the 3.5 GB debate came up.
frontieruk
Member
Could be driver, but in previous articles the Dev has going on record as saying their async code has passed AMD, Microsoft's and NVidias dx12 compilation tests so not the code, it could just be leveraging AMD strengths as it was a mantle test case as the starswarm demo.
Have you seen the results from the benchmark made by a dev on beyond3d?
Dictator93
Member
Wait, we confirmed that Maxwell 2 does not have it at all?
"We are focusing our efforts in bring you the best possible experience for the 99.999999999% of the games in the market also we are making progress to bring that support for the 0.00000001% games with DX12 and future releases.Thank you."
But PS4 does not use DX12. Xbox One on the other hand may just get that secret sauce.
serversurfer
Member
Edit: ^^^ notsureifserious
Yes, and you implied that would make up for the lack of hUMA, but it won't.
/comfortAh crap, no problem. I'm really tired and perhaps in a bad mood so I kind of bluntly missed your joke
We entertained notions. Nothing is confirmed by anyone at all.
Is this a serious post? I can't really tell.
It does not need one. PS4 has its own low level API that can do everything from day1 [while Xbone is fumbling with slower API, slower hardware and more complicated hardware architecture].
Nah, some console games in development are already seeing big gains.
Psyrgery ES
Member
I'm far from being a tech savvy guy, but this is not the first time we've heard that DX12 was boosting AMD's gpus performance whereas it didn't do much on NVidias.
Not trying to sound like a fanboy here, I love everysingle NV card I've got.
As I said, I believe this is not the first time this has been reported.
Not trying to sound like a fanboy here, I love everysingle NV card I've got.
As I said, I believe this is not the first time this has been reported.
dr. apocalipsis
Banned
I'm talking about an obvious performance penalty from a single controller managing heterogeneus workloads.
Thanks for enlightening me about shared memory.
Let's act as it hasen't been massively used on low end iGPUs for decades or phones instead of high performance stuff.
You can't be taken seriously when you talk about hUMA as an alternative to 68 GB/s for CPU and 336.5 GB/s for GPU instead of what it is, a solution for lower end products.
All AMD did was take the research was being done on DX12, release a proof of concept driver too buggy and lackluster for being used as an everyday driver and then stop all investment on it, cause it had no future since the beginning, extent they knew when they released all their PR bullshit about it.
MS introduced DX12 later as a fully solid and useable API for a wide spectrum of hardware, that doesn't mean they coded it after Mantle. Just NOT at all.
Cause no one else invent useful stuff. Every advantage you can get from using Intel or Nvidia hardware is made of money, nothing else.
The good and evil narrative, so tiresome.
icecold1983
Member
https://www.reddit.com/r/AdvancedMi...ide_games_made_a_post_discussing_dx12/cul9auq
has this been posted yet? amd employee says maxwell 2 is incapable of async compute
has this been posted yet? amd employee says maxwell 2 is incapable of async compute
Thats not how its working. Its automatic scheduling and more queues is just a resource. A resource that can be completely unnecesery in 90% cases and give slight advantage in 10% of cases. I dont know how You cant get that. Its the same as You would increase the wifth of memory bus, without changing the bandwith and expected the faster transfer.
Or putting 128 ROPs into PS4. It wont help. Its not a simple system, but everything depends of everything.
No performance comparisons in any tech paper for more than 1,5 years is not putting it in a good light either.
Thats not true. hUMA was pushed by many as something that give devs a tool to achieve things with compute that was been able before or have been done in very basic, like things i've mentioned.
I'm not going to bother quoting more, because for whole discussion You have not provided any real example for any of Your point or that would counter my points.
The only thing You can do, is to talk about generic theoretical advantages that were claimed by AMD and discussed here and beyond3d at large 1.5 years ago.
If You want to quote and engage with discussion with me, first dont call me stupid, second provide anything to support Your point. Something that is based on a research.
Because now discussion is similar to the 'coding to metal' theory.
Buys 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?
High five, I'm in same boat....