Asynchronous Compute

There's always intel XD

TBH, skylake has decent performance for igp. 5x more performance and it would run most games fine for 60fps FHD.

 

I gave you the link, it's gamegpu, front page. And yeah, here we are with you trying to spin the fact that all you've said about how badly Kepler is doing has no merit basically.

Where the hell is Fermi performing better than Kepler "all this time"? And what "this time" are we talking about?

Is it this the time where Fermi is doing better than Kepler?

That's 680 launch, 680 is 33% faster than 580.

Or maybe it's this time where Fermi is keeping up with Kepler?

That's Titan launch. 680 is 40% faster than 580.

Or is it here?

That's 980 launch. 680 is 22% faster than 580.

Or here?

Fury X launch last year. 680 is 26% faster than 580.

You should really check your data before saying anything because from what I see 680 was actually getting even further from 580 after launch with driver updates and DX11 being used more extensively.

So you just totally ignored what I've posted about the amount of games where 680 owners are missing out compared to 7970 owners, how this amount compares to the amount of games which are unplayable on both cards and continued with your holy crusade? Well done.

To get this train back on track - I don't know if you've noticed but the performance differences from async compute on GCN are marginal at best in all real world titles we have at the moment.

The choice between 7970 and 680 always was a choice between a higher performance headroom (3GBs alone made it apparent right on the start) and a cooler, quieter card with NV's software ecosystem. This choice hasn't changed now and I don't think that anyone who consciously chosen 680 back then are regretting this choice now. I certainly wouldn't.

 

Yes they can in dx11. There is always some room left by dx api/driver <>cpu overhead otherwise they wouldnt be able to do that what they did with R337 in the first place.

Although I was talking about dx12 boost there, and there is still room for improvement too, they're just waiting to get a whole package right async + extra api boosts, its what my gut is telling me and it didnt left me down yet :nerd:.

And well Ive been following nv driver for a loong time (since 6600gt era), so I think I know a thing or two by now and how they usually play the game if its getting a bit tight.


Anyway what are you trying to accomplish in this thread, all I see is you dissing nv async and babling how its SW while in fact it isnt. Also how Maxwell now sux because furyx got close to it / "winning" by one frickin beta benchmark that's build in Amd favor by default and one ultra buggy Glacier2 engine - Hitman...
*sarcasm* wow, yeah that is something special and should be taken as a fact how it will be in the future.*sarcasm*

 

I know it's not the same, but, up until the last few months of last year I was using GTX680m 4GB version and I had a great time with it. Only GTA V, Metal Gear and Project Cars made me upgrade (all still playable), but, I can say I really enjoyed gaming with that card and I do still have it. I bought this laptop back in March 2013 so lasted me a good 2 1/2 years.

In the laptop world, the laptop equivalent 7970m got wasted by the GTX680m.

As far as I'm concerned, for my interests Nvidia is the best option for laptop gamers. For this AMD has no answer...so went external gpu solution through thunderbolt instead. XD

 

The 370/7870 approaching the 770 in almost all latest titles is just normal. Unless all the latest Guru3D benches are wrong that is.
Another normal day in Kepler land.

Spoiler

Fermi has been performing much more consistently against the equivalent AMD cards of the time it was out. I would take the GTX 480 ove the 6970 ANY time. You can't seriously say that for ANY Kepler card and the AMD equivalent. You can't even say it for Maxwell any longer, except maybe the 980Ti.

Tj, I'm sure there might be a possibility for that man, but the cards seem to be used 100% almost all the time now (which is excellent). Overhead is an issue as long as the GPU is underutilized because it's not "fed" fast enough. From the few DX11 vs DX12 results we have it seems like NVIDIA cards don't really have any overhead/utilization issues at all. I'm not so sure than NVIDIA might squeeze more out of them.

As for the whole topic, this is my last post in it because we are all going in circles:

dr_rus, you are one of the best word twisters I've ever encountered. I mean that in the best and worst possible senses. You should be a layer sir, salutations.

Async compute won't provide almost anything for current NVIDIA GPUs. NVIDIA themselves in their developer guidelines that I have posted at least twice in here, say that the cost of task switching is too high. The whole point of Async Compute is that your hardware stays busy all the time by using the pipeline for various tasks simultaneously. Maxwell is a shorter pipeline design that has very heavy penalties when task switching in the middle of another task (which is what Async compute is). It won't happen for NVIDIA, best case scenario is that it doesn't have any performance penalty.

For the sake of people shoving $500-700 for Maxwell cards, I hope I'm wrong.

 

But there is API overhead still, even in dx11.. I can give you 3-4 games out of my head atm.

Dying light with view distance 3.60
GTA5 by grass and shadows / mostly foliage.
new TR grass/ but mostly world shadows
Hitman Absolution
Thief

to name a few, I dont see full gpu utilization when there is some extreme cpu limit (was the same on 780GTX), unless I raise more gfx settings which cripples min fps further, so I don't see a point in that in the first place.

So more drawcalls, index, vertex buffers, shaders and all that shizle is still welcome and it can be done in dx12 too.



Speaking of dx12 driver overhead/async it should still help where its cpu dx api bound or like by this over async abuse SoA or new Hitman. Its what async is here in the first place to eliminate some extra cpu driver overhead limits..

Or that starswarm, mantle / "dx12" helped AMD and how they now harvest more gpu power., But when nv optimized dx11 with 337.50 so much it actually beaten Mantle overhead even though Oxide crippled dx11 on purpose further with another benchmark update.. I've seen it my self on a highly OC'ed 780GTX, 92fps then after I posted this at their forums, 1-2weeks later another update and dx11 now just ~ 70ish fps, which made mantle now tiny bit faster at AMD equivalent gpu (290x).

So that being said, NV still has a trick or two up their sleeve, you just wait and see.. Im sure it won't take that loong. R364 is a test branch for minor AoS speedups by cpu part, but yeah its just a test atm still missing vsync parts (not enabled), and they already managed up to 15-20fps boost by cpu bound scenarios (according to one guru3d users with this benchmark @ NV hw).

Async can be demanding if its used wrong, if applied properly it can have drastic boosts.

Look at how DX12 do/s and don'ts work by Nvidia. Its how it is, since now devs need to optimize their games not the other way around.
https://developer.nvidia.com/dx12-dos-and-donts

https://developer.nvidia.com/dx12-dos-donts-updated

 

So you've again just ignored all data we have on the topic and continued with your personal fantasy. I see no point in arguing anymore on this as you're obviously incapable of seeing the facts.

 

:cheers:

 

Wow, I didn't know about that. That's some real shady stuff going on.

Oxide isn't to be trusted at all tbh. There is no feather in anyone's cap where AoTS is concerned. Don't buy this crap. Starswarm + AoTS = Don't buy!

 

Most of all, I have always felt that AoS is more or less tailored to be a benchmark, never minding the game itself (haven't read a review about the game itself up till now, all the talk's about the benchmark). Is it even anything special in the RTS genre?

 

Lol everyone arguing about Async Compute, meanwhile I went AMD because I simply wanted OpenCL 2.0 to make better physics for my waifu's skirt in Skyrim.

 

This is the key to understanding AMD's Asynch Compute-tailored marketing program. All we get are ****ty broken games (Hitman) and benchmarks like AOTS, and perhaps some gameplay-irrelevant cosmetics like hair in Deus Ex. The good games remain vendor-agnostic and probably run better on Nvidia hardware anyway.

 

And what cloth physics engine is running on OpenCL 2.0 exactly?

 

The push IMO has a lot to do with marketing , the move to Win10/Dx12 .
MS and of course AMD want to show biggest improvement so that users will move to Win10, using Dx12 as a motivator .
Same with the exclusive MS store Dx12 games .

 

Well Hilbert updated the Hitman review with the DX11 vs 12 comparison of the Fury. 5% at 1080p, looks like maybe 2% at QHD and almost nothing at 4K. AMD claimed "PC gamers may have heard of asynchronous compute already, and Hitman demonstrates the best implementation of this exciting technology yet." So the best implementation nets an average of 3.5% performance on the Fury and maybe 8% on the 390x. Not bad, but not incredible either. Need more data, QB needs to come out.

 

What people do not understand of "low overhead API" stands for?

 

HDT SMP uses OpenCL2.0 in Skyrim.

 

It seems like there is a limit to how much can actually be achieved in real games. I suspect engine limitations and perhaps having to support dx11 doesn't help. We should hope that AC does offer these reported big improvements because it will eventually benefit Nvidia users aswell.

On the other hand, I notice on Nvidia's site that there is almost nothing left referring to dx12.

 

i found this image on Geforce.com

 

Tiled resource wasn't something on dx11.2 or something ? Back in windows 8 days ?