Technical issues holding back custom RX Vega cards

[Doobeedoo]

Well that sucks, though was said they're on the way rather soon anyway. From some manufacturers sooner some later. 

[goodtofufriday]

47 minutes ago, Sakkura said:

I doubt that very much, unless you have evidence. In the past, it's been shown that there's no particular affinity between Intel and Nvidia, and in fact AMD cards benefitted slightly more from Intel CPUs than Nvidia cards did.

 

http://www.tomshardware.co.uk/crossfire-sli-scaling-bottleneck,review-32668.html

I think he's misquoting. The general consensus was that things have been optimized in a learn towards intel and nvidia hardware. Which is why you saw some marginal gains when using them in a comno.

[LAwLz]

7 hours ago, Taf the Ghost said:

There's 4 CPU + GPU combos: Intel + Nvidia; Intel + RTG; AMD + Nvidia; AMD + RTG.  What I'm saying is in the market dominant Intel + Nvidia combo, Nvidia's driver team was able to find some very clever tricks to bump up their performance in DX9 and DX11. The issues in a lot of DX12 titles is that bump seems to go away. It's also, likely, part of the reason that the results are so sensitive to inter-core latency and is, thus, something Nvidia is working to lessen in their drivers.

I see a lot of talking but no evidence.

Sorry but I don't trust observations like that without any actual evidence. Humans are very good at finding patterns which aren't there, or go looking for evidence for predefined conclusions, unconsciously ignoring evidence that does not fit their theory and highlighting evidence that does.

[Taf the Ghost]

4 hours ago, goodtofufriday said:

I think he's misquoting. The general consensus was that things have been optimized in a learn towards intel and nvidia hardware. Which is why you saw some marginal gains when using them in a comno.

What I was actually doing was praising Nvidia's driver development team for finding some interesting tricks that have let the Sandy Bridge & Haswell uArchs have an impressive life span for top-end gaming GPUs. It's also part of the reason the G4560 is also such a good performer. Inter-connected computer parts will never run at 100% efficiency, and I was praising the Nvidia team for finding some specific ones (that I can only point to the likely area that causes it)

50 minutes ago, LAwLz said:

I see a lot of talking but no evidence.

Sorry but I don't trust observations like that without any actual evidence. Humans are very good at finding patterns which aren't there, or go looking for evidence for predefined conclusions, unconsciously ignoring evidence that does not fit their theory and highlighting evidence that does.

Nearly all public testing in the computer space is "bad". Sometimes just "bad", but most of the time "woefully bad" to the point I've been tempted to write a basic guide on "how not to look like an idiot in reporting benchmarking results" or something to that effect. I have little desire to do a mass data collection and run analysis on it, just to have to point out that I've introduced so many variables that the data set is now corrupted beyond all but the most stupidly divergent outcomes within. (Something about half of Science needs to be beaten over the head about.)

 

But it's also the reason I've said the 8700k, with a high OC, is likely to be the best gaming CPU on the market and it should last for a very long time. That inclusive L3, low core-to-core latency (ring bus) and those benefits from the tricks Nvidia has used since Sandy Bridge (which don't work on the X299 platform) should make the CPU an utter, utter beast.

[themctipers]

Poor Volta.

[PCGuy_5960]

1 minute ago, themctipers said:

Poor Vega

FTFY

[themctipers]

Just now, PCGuy_5960 said:

FTFY

i love my cut down gpus!

 

1030 -> 1050ti -> 1060 3g -> 1060 6g -> 1070 -> 1070ti? -> 1080 -> 1080ti -> Titan xp -> Titan Xp -> some Quadro or something that's 'full GP102'

 

[DoctorWho1975]

Damn, I really dont want to dump 1500 dollars switching over to NVidia.

[mr moose]

1 hour ago, Taf the Ghost said:

What I was actually doing was praising Nvidia's driver development team for finding some interesting tricks that have let the Sandy Bridge & Haswell uArchs have an impressive life span for top-end gaming GPUs. It's also part of the reason the is also such a good performer. Inter-connected computer parts will never run at 100% efficiency, and I was praising the Nvidia team for finding some specific ones (that I can only point to the likely area that causes it)

 

Yeah, I still don't think it's quite like that.  If most games only need 2 cores to function, then a chip with 2 good cores with HT is going to perform admirably beside a chip with 8 not so good cores.     I think the nvidia architecture is just superior to GCN, it doesn't take as much for them to make cut down GPU's (XX50 XX60) perform really well.  Unfortunately I feel we are going to have to wait until after Navi to find out whats left in the GCN tank for gaming.  

 

 

[Agosto]

I'll have to wait for Vega 11... again.

[BuckGup]

AMD can just stick to CPUs now

[Jito463]

2 hours ago, mr moose said:

I think the nvidia architecture is just superior to GCN

He was talking about Nvidia cards on Intel processors as opposed to AMD processors, not Nvidia cards versus AMD cards.  On an Intel system, Nvidia gains a performance boost due to some _^demonic trickery in the drivers that manages to eke out additional FPS.

[mr moose]

2 hours ago, Jito463 said:

He was talking about Nvidia cards on Intel processors as opposed to AMD processors, not Nvidia cards versus AMD cards.  On an Intel system, Nvidia gains a performance boost due to some _^demonic trickery in the drivers that manages to eke out additional FPS.

I know.  I just disagree that the drivers are that good, I think the nature of the GCN architecture next to Nvidias is the driving factor between the differnce.

[NvidiaIntelAMDLoveTriangle]

16 hours ago, Taf the Ghost said:

There's 4 CPU + GPU combos: Intel + Nvidia; Intel + RTG; AMD + Nvidia; AMD + RTG.  What I'm saying is in the market dominant Intel + Nvidia combo, Nvidia's driver team was able to find some very clever tricks to bump up their performance in DX9 and DX11. The issues in a lot of DX12 titles is that bump seems to go away. It's also, likely, part of the reason that the results are so sensitive to inter-core latency and is, thus, something Nvidia is working to lessen in their drivers.

That's because DX12 is garbage. It has nothing to do with AMD or Nvidia knowing or not knowing how to better optimize for it and or give better results on certain configurations.

DX12 in itself is pure trash. It was promised to be this great thing and just like DX10, only promises but no deliveries.

[leadeater]

45 minutes ago, mr moose said:

I know.  I just disagree that the drivers are that good, I think the nature of the GCN architecture next to Nvidias is the driving factor between the differnce.

I know in games that have DX11 and DX12 modes the differences between the same cards from AMD and Nvidia lessens but the exact reasons for why aren't that easy to point to. Could be hardware related, could drivers, could be API, could game engine, could be a heck of a lot of things. Sure the trend of the gap closing, or increasing if AMD card was already ahead, is there for almost every DX11 + DX12 combo game but without some kind of full diag trace of the entire rendering output process from start to finish we'll never truly know.

 

There as been some evidence that the DX11 difference is wider on Intel CPUs than AMD Ryzen yet the difference is equivalent for DX12 which is where the theory of Nvidia DX11 driver magic comes from, again without the diags who knows why.

[Mira Yurizaki]

1 hour ago, NvidiaIntelAMDLoveTriangle said:

That's because DX12 is garbage. It has nothing to do with AMD or Nvidia knowing or not knowing how to better optimize for it and or give better results on certain configurations.

DX12 in itself is pure trash. It was promised to be this great thing and just like DX10, only promises but no deliveries.

It's only "garbage" because DX12 works best in a CPU constrained environment, which a vast majority of games are not CPU constrained. NVIDIA and AMD don't have to optimize anything because the optimization relies heavily on the developer.

 

DX12 didn't promise anything more than a programming model designed to better resemble how the GPU wants instructions and data to minimize overhead, nothing more. If your CPU was already more than capable of handling this overhead, minimizing the overhead isn't going to magically improve your performance.

 

51 minutes ago, leadeater said:

I know in games that have DX11 and DX12 modes the differences between the same cards from AMD and Nvidia lessens but the exact reasons for why aren't that easy to point to. Could be hardware related, could drivers, could be API, could game engine, could be a heck of a lot of things. Sure the trend of the gap closing, or increasing if AMD card was already ahead, is there for almost every DX11 + DX12 combo game but without some kind of full diag trace of the entire rendering output process from start to finish we'll never truly know.

 

There as been some evidence that the DX11 difference is wider on Intel CPUs than AMD Ryzen yet the difference is equivalent for DX12 which is where the theory of Nvidia DX11 driver magic comes from, again without the diags who knows why.

NVIDIA's DirectX 11 lead has to do with when NVIDIA experimented with deferred contexts (a great write up can be found at AnandTech. Some more reading can be found at NVIDIA's website). When they saw how much of a performance improvement this could bring, they decided to implement on the driver level, regardless if the application actually implemented it or not. AMD didn't bother with supporting this feature for some reason other than DirectX 11 doesn't make it mandatory to support it.

 

As far as the performance gap between the Core series and Ryzen in DX11, if there is one and assuming this is using NVIDIA, may be due to forced deferred contexts. Even though NVIDIA's draw call compiler is multithreaded, it still has to send data back to the immediate context who submits the draw calls. Because Ryzen cross core communication can have much higher latency in some cases, this will cause performance hits.

 

Moving to GPU architectures, the thing with GCN is how it wants its instructions. You really need to feed it as much as possible, otherwise you'll have execution bubbles. Each CU accepts a wavefront, which consists of 64 work items (as there are 64 ALUs in a CU). So basically, if you're not submitting 64 work items per wavefront in each draw call, you're wasting execution resources when the GPU goes to run them. NVIDIA's SMs on the other hand accept a warp, which consists of 32 CUDA Threads. This means you don't have to fill up a warp as much as you have to fill up a wavefront on NVIDIA's GPUs to get the most out of them.

 

How this all ties back to the DX12 performance gap, and this is speculation, is that if NVIDIA didn't really change the basic idea of how they submit draw calls to the GPU, because it was already mulithreaded to begin with, then DX12 won't really make a difference. It may reduce the overhead of processing the instructions since the data formats being submitted are easy to translate, but that's about it. However, DX12 opens up more draw call queues that can be submitted to the GPU at once. This opens up more channels to submit work for AMD GPUs since they never implemented deferred contexts. Meaning now AMD has to support some form of mulithreaded draw call compiling. And since for a given price range AMD's GPUs contain more execution units than NVIDIA's, this could help explain the gap.

 

But take that as you will.

 

EDIT: Wanted to expand on the GCN work submission part. You might be tempted to think that if you don't submit a full wavefront, the CU can work on something else when its done. But the CU works on the entire wavefront at once. And it can't mix in work items from other wavefronts to fill in the gaps because a CU is a single instruction, multiple data processor. Meaning every ALU in the CU is doing the exact same thing.

 

Note this is more or less true with NVIDIA's GPUs.

[leadeater]

52 minutes ago, M.Yurizaki said:

NVIDIA's DirectX 11 lead has to do with when NVIDIA experimented with deferred contexts (a great write up can be found at AnandTech. Some more reading can be found at NVIDIA's website). When they saw how much of a performance improvement this could bring, they decided to implement on the driver level, regardless if the application actually implemented it or not. AMD didn't bother with supporting this feature for some reason other than DirectX 11 doesn't make it mandatory to support it.

 

Quote

Coincidentally, last month's interview with AMD's Richard Huddy at Bit-Tech also has a lot in common with this. AMD says DX11 multi-threaded rendering can double object/draw-call throughput, and they want to go well beyond that by bypassing the DX11 API.

March 16, 2011

 

So instead of AMD focusing on DX11 improvements they went down the Mantle path, in some ways it did work out but it seems to be a rather heavy price to pay for a long term goal that is still yet to be realized but at least we can actually see it now.

[Mira Yurizaki]

4 minutes ago, leadeater said:

So instead of AMD focusing on DX11 improvements they went do then Mantle path, in some ways it did work out but it seems to be a rather heavy price to play for a long term goal that is still yet to be realized but at least we can actually see it now.

It makes me wonder why AMD didn't want to implement deferred contexts at all. Like what did they see in Mantle that apparently nobody else could?

[leadeater]

2 minutes ago, M.Yurizaki said:

It makes me wonder why AMD didn't want to implement deferred contexts at all. Like what did they see in Mantle that apparently nobody else could?

I think they expected it to catch on a lot faster than it did. I think they were victims of listening to developers that say they want something that don't really actually want what they are asking for i.e. low level APIs. It's easy to say yea I want that and be able to point to a bunch on benefits but you actually have to follow through and support it, develop the tools, be in it for the long haul and really help. Seems to me most wanted all those benefits but were only prepared to make the switch when it was as mature as DX11 was. DX12 and Vulkan are suffering the same issues.

 

Not sure why AMD thought something as disruptive as Mantle would catch on any faster than DX11 did over DX9, or even as fast as.

[Mira Yurizaki]

2 minutes ago, leadeater said:

Not sure why AMD thought something as disruptive as Mantle would catch on any faster than DX11 did over DX9, or even as fast as.

They were willing to license it so others could use it. But the fact that a GPU company is making the API probably didn't settle well with anyone else who made GPUs.

 

But hey, AMD got their wish in the end since Vulkan contains a lot of Mantle.

[LAwLz]

5 hours ago, Jito463 said:

He was talking about Nvidia cards on Intel processors as opposed to AMD processors, not Nvidia cards versus AMD cards.  On an Intel system, Nvidia gains a performance boost due to some _^demonic trickery in the drivers that manages to eke out additional FPS.

Again, lots of talking but no actual evidence of this supposed "demonic trickery" in Nvidia's drivers.

 

I remember someone testing how much CPU the drivers used before. The result was that AMD's drivers required a lot more CPU cycles than Nvidia's in general. But that was not specific to Intel or AMD. It was just that a better CPU helped AMD more because their drivers were less efficiency. Are you sure people aren't getting that confused with this magical "Nvidia's drivers are optimized for Intel!"?

[mr moose]

3 hours ago, leadeater said:

 

March 16, 2011

 

So instead of AMD focusing on DX11 improvements they went down the Mantle path, in some ways it did work out but it seems to be a rather heavy price to pay for a long term goal that is still yet to be realized but at least we can actually see it now.

I was going to raise Mantle earlier when talking about AMD spending resources on drivers to the same degree that nvidia have,  I still think they are definitely putting in as much effort on the driver side, just it seems they are trying to force a square peg onto a round hole with GCN and gaming.

 

 

 

 

 

[Trixanity]

6 hours ago, LAwLz said:

Again, lots of talking but no actual evidence of this supposed "demonic trickery" in Nvidia's drivers.

 

I remember someone testing how much CPU the drivers used before. The result was that AMD's drivers required a lot more CPU cycles than Nvidia's in general. But that was not specific to Intel or AMD. It was just that a better CPU helped AMD more because their drivers were less efficiency. Are you sure people aren't getting that confused with this magical "Nvidia's drivers are optimized for Intel!"?

I think it can be evidenced by the fact that X299 CPUs with the new mesh interconnect has the same issue Ryzen has. Nvidia's driver simply works best on whatever 'trickery' they have done with the ring bus or something related to it. Well, that or the mesh adds latency which Nvidia is incredibly sensitive to. In either case there is a performance discrepancy there.

RTG does not share that problem but has been prone to be bottlenecked by weak CPUs on DX11 due to needing brute force to keep the beast fed.

[Taf the Ghost]

6 hours ago, M.Yurizaki said:

They were willing to license it so others could use it. But the fact that a GPU company is making the API probably didn't settle well with anyone else who made GPUs.

 

But hey, AMD got their wish in the end since Vulkan contains a lot of Mantle.

 

2 hours ago, mr moose said:

I was going to raise Mantle earlier when talking about AMD spending resources on drivers to the same degree that nvidia have,  I still think they are definitely putting in as much effort on the driver side, just it seems they are trying to force a square peg onto a round hole with GCN and gaming.

3Dfx and Glide come to mind. And AMD making a lot of fairly costly design choices in the late 2000s early 2010s that they are still paying for.

[Jito463]

6 hours ago, LAwLz said:

Again, lots of talking but no actual evidence of this supposed "demonic trickery" in Nvidia's drivers.

 

I remember someone testing how much CPU the drivers used before. The result was that AMD's drivers required a lot more CPU cycles than Nvidia's in general. But that was not specific to Intel or AMD. It was just that a better CPU helped AMD more because their drivers were less efficiency. Are you sure people aren't getting that confused with this magical "Nvidia's drivers are optimized for Intel!"?

I wasn't trying to prove anything, simply attempting to help clarify the conversation.

 

As for the "demonic" comment, that was intended to be a joke.