The 51%/98% GPU usage dilemma

[sirmo]

Someone better tell Nvidia then, as we will probably see a 4th gen Titan within a year. I mean how else do you explain the success of the GTX 960 without accepting that it lived in the 980, Titan X and 980 ti's halo? The 280x was faster, the 380x is faster. I can buy the power savings argument for the 750 ti that doesn't need a 6 pin, but the 960 was a clear example of Nvidia's marketing winning when their technology lost.



Even AMD admits that is a VR card. Crossfire might as well not exist if you are a zero day gamer, which is the kind of gamer that keeps the whole wheel spinning.

Titan is more aimed to woo investors, and besides it's not like they don't make a really nice markup on those types of cards.

In either case. AMD's decision to go for the mainstream market first I think is a brilliant move. For multiple reasons:

- why fight it out with Nvidia, when they can have the far larger market to themselves. Nvidia still enjoys a pretty big mindshare in enthusiast circles, so the question for AMD was, 30% of the tiny enthusiast market or 100% of the mainstream market? No brainer really.

- AMD has plenty of fab capacity they need to use up. They have a wafer supply agreement with Global Fundries they have to honor. And Zen isn't ready yet. So might as well put those wafers to good use.

- Apple is a pretty important OEM for them, and Apple is doing product refreshes on their most popular line of computers this year. Securing another design win with Polaris was pretty important.

- Make small dies until the yields improve. They did the same thing with their Cypress 5870 launch. When they first moved to 40nm they first ran a die shrunk run of 4770 on it until the yields improved for the 5870. Both of which were killer products.

- Mainstream market is pretty saturated. If you look at all the recent products in the $300-$600 market you will find tons of competitive products there. The mainstream has been largely left unserviced. In fact AMD's best selling card has been the r9 380 which is exactly in the $200 range.

- And also they might be right about their "increasing the TAM" line. I see tons of folks building their first PCs. rx480 is god sent for those intimidated and on a budget, trying their luck at building their own PC and getting their feet wet in PC gaming.

This is the absolute best move AMD could have made with the 14nm transition imo. It will generate most impact for their bottom line in the future. It will also increase their install base which also can't hurt. Especially since they already dominate consoles, their goal is to take away Nvidia's grip on game developers.

 

[HurleyBird]

This is a myth, I wish people would stop repeating it. Halo product is pointless. 4870 and 4850 killed and there were no halo products. Also right now we do have the single fastest card on the planet in AMD's court the Radeon Pro Duo.

Sorry, but the terrascale generation is the clearest indication of AMD's uncanny ability to shoot itself in the foot that exists.

Terrascale had such a huge architectural advantage over Tesla, it was literally like comparing Core to Netburst. Had AMD actually made a 400mm2+ die they would have wiped the floor with Nvidia the same way they did with the 9700 pro, and they desperately needed a big win after the whole R600 fiasco.

Instead they produced tiny little dies that, while obviously able to hit way above their weight class, gave Nvidia plenty of room to capture the performance crown through sheer brute force. With RV770 they cemented their place as the economy brand of GPUs, and quite literally snatched defeat from the jaws of victory.

 

[sirmo]

Sorry, but the terrascale generation is the clearest indication of AMD's uncanny ability to shoot itself in the foot that exists.

Terrascale had such a huge architectural advantage over Tesla, it was literally like comparing Core to Netburst. Had AMD actually made a 400mm2+ die they would have wiped the floor with Nvidia the same way they did with the 9700 pro, and they desperately needed a big win after the whole R600 fiasco.

Instead they produced tiny little dies that, while obviously able to hit way above their weight class, gave Nvidia plenty of room to capture the performance crown through sheer brute force. With RV770 they cemented their place as the economy brand of GPUs, and quite literally snatched defeat from the jaws of victory.

They ditched Terascale because they wanted to use RISC instead of VLIW. Which when you look at their HSA initiative makes absolute sense. VLIW makes for complex compilers and it has some inherent instruction bandwidth issues. Instruction caching is also a challenge.

I think VLIW is just harder to scale and scaling RISC is much better understood particularly when you consider the brain trust in the CPU side of the AMD house.

GCN is a good architecture for consoles, because console developers actually target it for maximum utilization. While on the dGPU side of things AMD has paid a price with a lot of those features going unused.

This is changing now however. Mantle influence on the APIs has been successful. And we're already seeing significant performance gains on DX12/Vulkan titles. It was just one of those bandages AMD had to pull at some point. And now that they've been through that transition I think the worst is behind them.

 

[HurleyBird]

Uh, that's not relevant to what I was saying.

 

[GodisanAtheist]

Sorry, but the terrascale generation is the clearest indication of AMD's uncanny ability to shoot itself in the foot that exists.

Terrascale had such a huge architectural advantage over Tesla, it was literally like comparing Core to Netburst. Had AMD actually made a 400mm2+ die they would have wiped the floor with Nvidia the same way they did with the 9700 pro, and they desperately needed a big win after the whole R600 fiasco.

Instead they produced tiny little dies that, while obviously able to hit way above their weight class, gave Nvidia plenty of room to capture the performance crown through sheer brute force. With RV770 they cemented their place as the economy brand of GPUs, and quite literally snatched defeat from the jaws of victory.

Thank you. I've been saying this for years. Imagine if AMD had released something close to the 5870 (1600 SPs) on the 55nm process with a die size of 500mm^2+. The GTX 280 would have been a bloody corpse left by the wayside weighed down by its CUDA transistors.

Same applies to the 40nm generation, with a theoretical 3200sp 500mm^2+ VLIW chip laying waste to the Fermi equivalent. NV's putting around with CUDA left them in a vulnerable position, but AMD was either too timid or too CPU centric to capitalize on the weakness.

 

[Element115]

Clarification from AMD themselves:

https://www.reddit.com/r/Amd/comments/4m692q/concerning_the_aots_image_quality_controversy/d3sw31g

 

[BlitzWulf]

Scaling is 151% of a single card

https://www.reddit.com/r/Amd/commen...s_image_quality_controversy/d3sw31g?context=3


edit: beat me :biggrin:

 

[khon]

Clarification from AMD themselves:

https://www.reddit.com/r/Amd/comments/4m692q/concerning_the_aots_image_quality_controversy/d3sw31g

Two comments

1. That's quite bad for CF scaling

2. If 151% of RX 480 beats a GTX 1080, that's incredibly value for $200

I know this title is good for AMD, which is why they picked it, but I don't think an R9 390X gets that close to a GTX 1080, so this looks very promising.

 

[Leadbox]

Two comments

1. That's quite bad for CF scaling

2. If 151% of RX 480 beats a GTX 1080, that's incredibly value for $200

I know this title is good for AMD, which is why they picked it, but I don't think an R9 390X gets that close to a GTX 1080, so this looks very promising.

It's not you typical AFR, this is SFR crossfire.

 

[DarkKnightDude]

Two comments

1. That's quite bad for CF scaling

2. If 151% of RX 480 beats a GTX 1080, that's incredibly value for $200

I know this title is good for AMD, which is why they picked it, but I don't think an R9 390X gets that close to a GTX 1080, so this looks very promising.

I'd imagine the crossfire driver side needs to be optimized a bit. I have seen some ridiculous scaling on some of their xfire setups.

 

[flash-gordon]

Scaling is 151% of a single card

https://www.reddit.com/r/Amd/commen...s_image_quality_controversy/d3sw31g?context=3

edit: beat me :biggrin:

Probably SFR, if it was AFR it would be getting a higher scaling.

But why say GPU usage if it was scaling? AMD mixed everything they could to shadow the results.

1-Show the 4GB cheaper RX480.
2-Bench the 8GB and more expensive RX480.
3-Say it was using just 51% of the GPUs but it wasn't, it was scaling.

If AMD really care for PC gamers the way Koduri like to say, then just be objective and stop trying to confuse us.

 

[hawtdawg]

Just my take on the situation.

It's gonna be waaay slower. I'd be surprised if it's much, if any faster than a 980, and it will probably also have trademark AMD overclocking, IE top out at like 1400mhz.

 

[Noctifer616]

I'd imagine the crossfire driver side needs to be optimized a bit. I have seen some ridiculous scaling on some of their xfire setups.

This isn't crossfire. There is no crossfire or SLI support in DX12, custom mGPU options are used instead. AoTS doesn't use AFR for mGPU in DX12 but SFR which doesn't scale so well with performance but has other benefits like smoothness and no latency increased compared to AFR.

Also, drivers shouldn't improve performance unless there are bugs in the drivers. Performance is dependent on game optimizations.

 

[Erenhardt]

it will probably also have trademark AMD overclocking, IE top out at like 1400mhz.

I don't think so. AMD best overclocker hd7950 could be overclocked 50%+ on the core. Some samples went from 800mhz to almost 1300mhz.

Asking for 1400Mhz overclock on 480 which will have its core clock around 1300mhz is silly. That would mean additional 100% core frequency, almost double of what 7950 offered before amd boosted stock clocks.

It would leave out too much on the table. I guess they would rather claim first 2GHz GPU then leave that much OC headroom.

 

[Glo.]

Guys, think for a moment. Polaris 10 Doom Presentation was showing that in 1440p they got minimum 60 FPS. This is Fury Nano level of performance.

Also, the score from 3dMark 11 benchmark shows Polaris 10 on par with AMD Fury.

 

[hawtdawg]

I don't think so. AMD best overclocker hd7950 could be overclocked 50%+ on the core. Some samples went from 800mhz to almost 1300mhz.

Asking for 1400Mhz overclock on 480 which will have its core clock around 1300mhz is silly. That would mean additional 100% core frequency, almost double of what 7950 offered before amd boosted stock clocks.

It would leave out too much on the table. I guess they would rather claim first 2GHz GPU then leave that much OC headroom.

I meant max out at 1400mhz.

 

[OatisCampbell]

The game was CPU bound at that point so it couldn't send the GPU more work, so it was going unutilized.

They didn't want to show single 480 vs 1080 since it would be (slightly?) slower, and people would ignore the huge $200-250 vs $700 price difference.

Kinda like 290 vs 780 Ti all over again, except even bigger price difference.

This is far beyond impossible.

AMD gets good scaling with Crossfire. Their slide that shows two 480s as slightly faster than a 1080 means one is a lot slower than a 1080, probably more like a 970/290.

I don't think even AMD would claim Polaris is 1070 level or close to it. They're pricing it at $199, would be at least $300 if it could trade blows with 1070s.

The cards you're talking about will be launched later this year.

 

[OatisCampbell]

SFR ???

Pay attention to the last two quotes.

- Frames no longer need to be queued, time between frame completion and user viewing reduced to 2-3x.
- Using the GPUs in parallel to work on one frame allows to multiple GPUs to behave like one much more powerful GPU.

The dual RX480 with only 51% utilization was faster than GTX1080 and perhaps at the same or lower energy consumption.

SFR has been around forever, SFR and tiling were the first multi GPU techniques.

SFR was abandoned because scaling is highly variable, load balancing is difficult.

I also don't get your math with "Polaris plus 50% slightly faster than 1080" means Polaris is 25% slower than 1080.

 

[RussianSensation]

I meant max out at 1400mhz.

This is a lower end card @ $199-249, aimed for 1080p 60Hz. If June 29th comes and there is something better between $150-300, then this point could have some merit. Otherwise, since 1070 also barely overclocks 11-14%, but the cheapest one costs $379, the RX480 just obsoleted every single AMD and NV card from $150-300.

Not directed at you, I hate to bring this up now, but the anti-AMD brigade who has now desperately moved on to crap on the RX 480 has ignored HD7000 series (esp. 7950 800mhz -> 1150-1300mhz, 7970 925mhz -> 1150-1250mhz) overclocking throughout the entire duration of that generation against Kepler. Many of the same people have also ignored lower VRAM for 1080p 60Hz gaming cards while pushing/recommending 950/960 2GB over AMD's R9 series. Finally, by far the predominant majority of NV supports always ignored how after-market $250-280 290 4GB slaughtered 960 2-4GB by 50-70%, with even a thermal throttling R9 290 actually matching or even beating 960 SLI.

In contrast, today, the same people are trying to do everything possible to try to hype up a $379 1070 for 1080p 60Hz, a NV card that will have a smaller lead over $199-229 RX 480 than a $250-280 R9 290 had over the 960?!!!

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_960_SLI/23.html

Do you realize how that actually looks from this perspective? In other words, if they were being consistent, the 1070 would need to cost $250-280, be 45-56% faster at 1080p 60Hz/1440p, and actually be as fast as RX 480 CF. Even if somehow 1070 can match the performance gain, it'll still never pass this litmus test because 1070 costs $379.

Plain and simple: R9 280X/380X and R9 290 destroyed 950 and 960 in price/performance. On the contrary, 1070 looks to offer far inferior price/performance for 1080p 60Hz gamers over the RX 480. Take your pick:

$379 1070 is 90% more expensive than a $199 RX 480 4GB for XX% more performance

$379 1070 is 52% more expensive than a $249 RX 480 8GB for XX% more performance

$449 reference blower FE 1070 is 126% and 80% more expensive than reference blower RX 480 4GB and 8GB cards, while offering XX% more performance.

Now this XX% I left empty since we don't have reviews of RX 480 yet. But do you actually believe 1070 will be 52%-90% faster than RX480 in games?

For 1080p 60Hz gaming, for the 84% majority who buy in the $100-300 mainstream/performance segment, AMD is about to launch a card that has no viable competition from NV until 1060/Ti shows up.

 

[Leadbox]

This is far beyond impossible.

AMD gets good scaling with Crossfire. Their slide that shows two 480s as slightly faster than a 1080 means one is a lot slower than a 1080, probably more like a 970/290.

I don't think even AMD would claim Polaris is 1070 level or close to it. They're pricing it at $199, would be at least $300 if it could trade blows with 1070s.

The cards you're talking about will be launched later this year.

The good scaling you're talking about happens with AFR, AOTS does not use AFR for Xfire.

 

[Dresdenboy]

Some more info:

Benchmark results:

2x Radeon RX 480 - 62.5 fps | Single Batch GPU Util: 51% | Med Batch GPU Util: 71.9 | Heavy Batch GPU Util: 92.3% GTX 1080 – 58.7 fps | Single Batch GPU Util: 98.7%| Med Batch GPU Util: 97.9% | Heavy Batch GPU Util: 98.7%

https://www.reddit.com/r/Amd/comments/4m692q/concerning_the_aots_image_quality_controversy/

 

[96Firebird]

Ah, good ol' marketing. I still remember the slide that showed Fury X faster than the 980 Ti... Of course it was at select settings, presumably to show Fury X in the best light.

I wouldn't put too much thought into these results.

 

[antihelten]

The good scaling you're talking about happens with AFR, AOTS does not use AFR for Xfire.

First of all AotS doesn't use Crossfire (or SLI for that matter), it uses DX12's explicit multi-adapter (EMA) feature. Secondly, unless it was changed recently, AotS uses AFR not SFR.

Just because DX12 makes it easier to use SFR, doesn't mean that every DX12 game out there will suddenly use it.

 

[AtenRa]

I also don't get your math with "Polaris plus 50% slightly faster than 1080" means Polaris is 25% slower than 1080.

??

I dont know what you are talking about.

 

[flash-gordon]

The good scaling you're talking about happens with AFR, AOTS does not use AFR for Xfire.

Some more info:

https://www.reddit.com/r/Amd/comments/4m692q/concerning_the_aots_image_quality_controversy/

WTH, so it isn't scaling? I guess this low GPU usage on reduced draw calls (single and medium batch) must be the effect of some power gating ability.