AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[swilli89]

That was all on Ryzen, not Vega.

Yeah and as an investor.. it wasn't even totally Ryzen. Look at nVidia from Feb 2016 onward. And Tesla. Tech stocks as a whole skyrocketed. AMD's gains were all speculation based which is why in early March it started to sag.. and has now basically sat firm at $14 or below the past several months. Only a positive EPS on their Q2 earnings later this month can send it back north of $15 again. If RX Vega comes out looking better than expected that may help, but honestly if its worst case scenario as bad as everyone fears I doubt stock price will be effected by much.

 

[Elixer]

If they want to get into the enterprise GPU market, they need to support CUDA. Period.

Not really.

For PyTorch, we're seriously looking into AMD's MIOpen/ROCm software stack to enable users who want to use AMD GPUs.

We have ports of PyTorch ready and we're already running and testing full networks (with some kinks that'll be resolved). I'll give an update when things are in good shape.

Thanks to AMD for doing ports of cutorch and cunn to ROCm to make our work easier.

(Facebook's AI guy)
https://www.reddit.com/r/MachineLea...leased_by_amd_deep_learning_software/djpfmu1/

They just need their openCL/ROCm stuff to be faster/better than CUDA.

 

[darkswordsman17]

AMD's biggest mistake is that it is developing two separate architectures with less resources than Nvidia. GCN 4 for RX 400 and 500 and a different engine for Vega.

Nvidia develops one big product and then cuts it off into smaller pieces constitution all the cards we see now from the titan xp to the 1050. Its one architecture so they can develop and optimize it to the max and then just cut it and provide the cheaper versions of it.

I think if they worked on a big RX 480/580, increased the die area, added more shaders, put in optimizations that they've seen in the RX 400/500 series, like better memory compression, tile rasterization and more pixel processing power they could have had a true power house, especially since custom OC'd RX 580's to 1470+MHz compete against the GTX 1070 in some DX12 titles, it even comes very close to the 1070 in Doom, a vulkan based game.

Polaris was just a stopgap. In many ways you could argue a poor one as they could have just shrunk Hawaii on 14nm and gotten probably better performance. I think they were trying to see what efficiency gains could be had by adjusting certain parts though (and felt it was fairly safe due to where they were positioning it; they could make up for losing SPs with higher clockspeeds). I feel like AMD really didn't have a good grasp on GCN, like where it was really bottlenecked or how its performance would change depending on different things they would change (hence the very unbalanced Fiji). Also with Microsoft talking up the way they chose the design for Scorpio based on their simulated performance stuff, that leaves me under the impression that AMD didn't have anything that robust/sophisticated to gauge things themselves.

It's been said several times that AMD should have long ago realized Hawaii was the optimal GCN configuration.

2816:64:512-bit in Polaris would provide 15-20% IPC improvements (GCN 4 vs 1 or 2) and 20%+ higher clocks and potentially even more gains from a more favourable bandwidth to FLOP ratio (512-bit 6Gbps in 390X, vs nearly double that if they went GDDR5X).

I don't think they have the money to have multiple chips that much anymore. But AMD still seems to think they can defeat their bottlenecks but still create unbalanced chips. Polaris 10 is an improvement, but 32 ROPs and 256-bit GDDR5 still hold it back. And it remains to be seen if Vega fixed Fiji's problems. Maybe AMD doesn't realize Hawaii's ratio is the ideal GCN configuration actually...

But Xbox One X (Scorpio) is 2560SP and 384-bit, so even if it is still 32 ROPs it could provide a decisive 1060 killer, and possibly a 1070 competitor if it is 64 ROPs. Why this isn't a dGPU is odd.

Polaris was just a stopgap between Vega which I'm sure they knew would take longer, hence why 2016 was Polaris, 2017 is Vega. Polaris was not a big change, and I think it was more of a trial and error (see how tweaks to the architecture impact performance) and making small changes (wasn't there something about how Polaris handles wavefronts and some other stuff, maybe cache?) towards what Vega brings. And then we'll see cutdown Vega replace Polaris probably next year (with probably mildly tweaked Vega chips like the 580 was to the 480 this year). Even the year before the lower cards weren't really a new architecture (just upping to 8GB mostly I thought), so most of the real work likely went into Fiji (which a lot of that was probably just dealing with HBM).

Definitely AMD has been hurting for resources. Actually I think that's exactly why it isn't a dGPU, as they don't have the resources to support another chip that differs from all of their PC ones in some manner. Microsoft is likely handling most of the low level software support for Scorpio. Plus I think it has some stuff that are probably Microsoft only (there's some hardware thing that manages calls for DX12, although if I'm AMD I would be asking to get to implement that on PC if it helps a lot, if for APUs alone).

Vega is stepping stone for future generations. We would not see any improvement per clock, like we see with Vega, compared to previous generations of GCN, if what you describe be the case.

I agree, which is why I'm not surprised that it seems to be underwhelming so far. We saw GCN age well, and I could see similar happen with Vega. While disappointing, I certainly won't be shocked to find that Vega takes a while to offer what it is capable of. I think it is bad tidings for what to expect from AMD though, as they have other major changes they're making soon as well (Navi which allegedly pushes the multi-GPUs operating as one, and then the mention of next gen memory), and so there could be a lot of teething problems while they sort things out. But maybe the work put into Vega will then streamline Navi (so the focus can be on the multi-GPU setup). I have a hunch more that both will be needing to be worked on. But I also think that might mean there's a good chance of Vega being a dual chip card at some point. It would be an early start on the interGPU communication, but not going to the level that Navi has been talked of being (mention of 4+). It will be interesting to see for sure.

 

[EXCellR8]

The good news is that this puts RX Vega above overclocked GTX 1070, bad news, it might still be slower than overclocked GTX 1080. I guess time will verify those results

RX Vega 3DMark Performance?

 

[JDG1980]

Not really.
(Facebook's AI guy)
https://www.reddit.com/r/MachineLea...leased_by_amd_deep_learning_software/djpfmu1/

They just need their openCL/ROCm stuff to be faster/better than CUDA.

That's just a negotiating tactic for Facebook to get better prices on Nvidia hardware.

 

[JDG1980]

Polaris was just a stopgap. In many ways you could argue a poor one as they could have just shrunk Hawaii on 14nm and gotten probably better performance. I think they were trying to see what efficiency gains could be had by adjusting certain parts though (and felt it was fairly safe due to where they were positioning it; they could make up for losing SPs with higher clockspeeds). I feel like AMD really didn't have a good grasp on GCN, like where it was really bottlenecked or how its performance would change depending on different things they would change (hence the very unbalanced Fiji). Also with Microsoft talking up the way they chose the design for Scorpio based on their simulated performance stuff, that leaves me under the impression that AMD didn't have anything that robust/sophisticated to gauge things themselves.

If random enthusiasts like you and I really know more about the bottlenecks of GCN than RTG's own engineers, that's an incredibly damning indictment of RTG.

I know that shortly after the development of Hawaii, AMD laid off the old ATi engineers. (Rory Read, the git who keeps on giving, thought that dGPUs were going away.) GPU development was then delegated to a new, cut-rate team in Shanghai. It's pretty obvious at this point that the Shanghai team, which didn't develop GCN, doesn't really understand it all that well, or have a clear idea on how to effectively improve it in terms of performance and efficiency.

And then we'll see cutdown Vega replace Polaris probably next year (with probably mildly tweaked Vega chips like the 580 was to the 480 this year).

Assuming RX Vega doesn't improve substantially beyond GTX 1080 performance, it's not clear that a cut-down Vega 11 chip would even beat Polaris 10 in performance or efficiency. We've heard essentially nothing about Vega 11; cancellation is a distinct possibility. RTG's latest roadmap still has Polaris parts listed for 2018.

I think it is bad tidings for what to expect from AMD though, as they have other major changes they're making soon as well (Navi which allegedly pushes the multi-GPUs operating as one, and then the mention of next gen memory), and so there could be a lot of teething problems while they sort things out. But maybe the work put into Vega will then streamline Navi (so the focus can be on the multi-GPU setup). I have a hunch more that both will be needing to be worked on. But I also think that might mean there's a good chance of Vega being a dual chip card at some point. It would be an early start on the interGPU communication, but not going to the level that Navi has been talked of being (mention of 4+). It will be interesting to see for sure.

Here's the problem: for this to be at all workable for gamers, it absolutely has to be transparent to the game software. Crossfire/SLI is practically dead at this point, and explicit multi-GPU never took off and there's no reason to think it ever will. Unfortunately, everything we've seen recently from RTG indicates that they really, really want to get out of the business of writing good DX11 gaming drivers, with their need for optimization and game-specific hacks. This is why they've been pushing DX12 and Vulkan so hard - it takes the optimization work off AMD's shoulders and puts it back on the game devs. But game devs hate it for the same reason, and most will just keep writing DX11, let Nvidia optimize for them, and ignore the relatively small Radeon market share. Note how Vega's DX11 drivers don't support any of the new features on existing titles - developers have to optimize for them explicitly. If AMD tries this stunt with Navi, they're doomed.

 

[Paratus]

If random enthusiasts like you and I really know more about the bottlenecks of GCN than RTG's own engineers, that's an incredibly damning indictment of RTG.

I know that shortly after the development of Hawaii, AMD laid off the old ATi engineers. (Rory Read, the git who keeps on giving, thought that dGPUs were going away.) GPU development was then delegated to a new, cut-rate team in Shanghai. It's pretty obvious at this point that the Shanghai team, which didn't develop GCN, doesn't really understand it all that well, or have a clear idea on how to effectively improve it in terms of performance and efficiency.



Assuming RX Vega doesn't improve substantially beyond GTX 1080 performance, it's not clear that a cut-down Vega 11 chip would even beat Polaris 10 in performance or efficiency. We've heard essentially nothing about Vega 11; cancellation is a distinct possibility. RTG's latest roadmap still has Polaris parts listed for 2018.



Here's the problem: for this to be at all workable for gamers, it absolutely has to be transparent to the game software. Crossfire/SLI is practically dead at this point, and explicit multi-GPU never took off and there's no reason to think it ever will. Unfortunately, everything we've seen recently from RTG indicates that they really, really want to get out of the business of writing good DX11 gaming drivers, with their need for optimization and game-specific hacks. This is why they've been pushing DX12 and Vulkan so hard - it takes the optimization work off AMD's shoulders and puts it back on the game devs. But game devs hate it for the same reason, and most will just keep writing DX11, let Nvidia optimize for them, and ignore the relatively small Radeon market share. Note how Vega's DX11 drivers don't support any of the new features on existing titles - developers have to optimize for them explicitly. If AMD tries this stunt with Navi, they're doomed.

So for Navi, HBM is essentially silicon on silicon right? So what if instead of only laying traces on the interposer they threw in some transistors and pulled some of the front end of the traditional GPU and moved it to the interposer.

It would schedule the 1, 2, or 4 Navi modules behind it and hide the whole thing from the rest of the system. Add a few more transistors to aggregate the output from the ROPs and viola it looks like a single GPU to the system.

My guess is latency between chips on an interposer is an order of magnitude better than through PCIe or a crossfire dongle.

Anyways, just a thought.

 

[Wall Street]

So for Navi, HBM is essentially silicon on silicon right? So what if instead of only laying traces on the interposer they threw in some transistors and pulled some of the front end of the traditional GPU and moved it to the interposer.

It would schedule the 1, 2, or 4 Navi modules behind it and hide the whole thing from the rest of the system. Add a few more transistors to aggregate the output from the ROPs and viola it looks like a single GPU to the system.

My guess is latency between chips on an interposer is an order of magnitude better than through PCIe or a crossfire dongle.

Anyways, just a thought.

Not a great idea. Then they would need to use a recent high performance process to create the interposer in that case. The idea is that the interposer can currently be made on a process that is useless for modern logic chips to save cost (there is little other demand for that equipment). Making an 800 mm2 interposer on a modern process will double the costs. For example, stacking a 500 mm2 14 nm die on an 800 mm2 28 nm die would be too costly.

 

[GodisanAtheist]

AMD is a bit of a wild card in the launch department. I'm *positive* that they will find some way to screw up the Polaris launch and subsequent launches (Terrible pricing, clocked too low, crap stock cooler, horribly delayed, etc etc etc). The last clean sweep launch they had was way back when with the HD5xxx series .

Fact is AMD is strapped for cash, has suffered some amount of brain drain and is trying to launch an overhauled microarch at a foundry that's untested for the task (Has GF actually made any chips of note?).

AMD demoed working Polaris silicon before we even knew what the 10xx series from Nvidia would be called. Nvidia has presented, demoed and will even "launch" their Pascal silicon before we have any concrete details about Polaris.

I think AMD fans should temper their expectations a bit; AMD just doesn't have the resources at the moment to execute like Nvidia does.

- Despite seeing what happened with the Polaris launch, I really wanted to believe AMD was going to pull a rabbit out of its hat and all the fun engineering terms and "secret sauce" were going to culminate in something that would really light a fire under the Titan XP and 1080ti. Its becoming more and more apparent that this is not going to be the case and even performance splitting the 1080ti and 1080 might be a bit much to hope for.

AMD needs to do some serious soul searching and figure out what the hell it actually wants to be because it clearly cannot be everything. They still have something left since they managed to pull off a decent launch with Ryzen, but what then? I don't see AMD letting their CPU business decay to put resources into RTG to pull off a similar feat.

Oh well just another couple dozen days to go. I would love to be proven wrong, in fact I hope I am wrong, but I wasn't wrong before and I don't think I'll be wrong now.

 

[Elixer]

Thing is, they knew they had a tough battle designing Ryzen, they brought on some more external help, and they either met or exceeded their goals if you go by what they say.
However, it hits a brick wall when trying to get higher speeds than what the 1800 is shipping.

With Polaris, they seem to have hit a brick wall, they can't get much higher clocks without causing massive upswing to TDP.

With Vega, I am sure they had a performance target, and they seem to be hitting that brick wall yet again, too much voltage, and the heat goes out of control. I highly doubt they are going to say they have met or exceeded their goals on this one--at least so far.

Only thing all these have in common is GloFlo's fabs.
Either the process is giving them fits, or the design is just way off for what GloFlo can handle.

 

[Paratus]

Not a great idea. Then they would need to use a recent high performance process to create the interposer in that case. The idea is that the interposer can currently be made on a process that is useless for modern logic chips to save cost (there is little other demand for that equipment). Making an 800 mm2 interposer on a modern process will double the costs. For example, stacking a 500 mm2 14 nm die on an 800 mm2 28 nm die would be too costly.

Maybe, maybe not. The trick would depend on whether an interface between the system and the bleeding edge 14nm dies could be designed to hand off work fast enough to keep the "GPU" fed with 65nm speeds. Something similar to but smarter and faster than the PCIe bridge chips the multi GPU cards currently use. (All the geometry engines, TMUs, ROPs and shaders would still be on 14nm)

At any rate for a seamless multi GPU solution I think there has to be some dedicated hardware to interface between the system and the multi GPUs behind it. Otherwise it will be no better than SLI/Xfire.

Anyway this has gone OT. So I'll drop it.

 

[beginner99]

If they want to get into the enterprise GPU market, they need to support CUDA. Period.

They will or already do with HIP

 

[QualityTime]

I'm not sure I agree. The 7970 had 32 ROPs and competed very well against the GTX 680, 770, and these days even does alright against the 780.

Fury X was basically twice an 7970Ghz. So they doubled the ROPs. Vega adds another 550mhz and architectural changes to the mix.

64 ROPs doesn't seem to me to be this obvious design stumble.

When you consider GCN designs usually have 4 ROPs per 32b memory channel, a 96/128 ROPs design has some fun implications: 32 ROPs - 256b(8 channels), 64 - 512b(16), 96 - 768b(24), 128 - 1024b(32).

There would be slight issues with power and board cost.

You maybe right. However Fury X competed well against the 980Ti despite the difference in ROPs.


I'm not even sure what bottlenecks the ROPS. My old 9600XT could do a theoretical 2+ gigapixels per second which is 70% more than 4K @ 144hz requires but even a 1080Ti struggles at that rez.

Memory reads and writes.

 

[Veradun]

I see we are at "HBM2 is really expensive, GDDR5X is almost free argument" again

Care to give us all some proof as we are looking for it for ages now?

 

[lobz]

If they want to get into the enterprise GPU market, they need to support CUDA. Period.

This is plain, short-sighted BS. Period.

Sent from my VTR-L09 using Tapatalk

 

[lobz]

That's just a negotiating tactic for Facebook to get better prices on Nvidia hardware.

And this is just the same tiresome argument we're hearing from nightmare troubled intel supporters in the past 6 months, whenever a major server player does so much as even mention AMD...

Sent from my VTR-L09 using Tapatalk

 

[3DVagabond]

It's been said several times that AMD should have long ago realized Hawaii was the optimal GCN configuration.

2816:64:512-bit in Polaris would provide 15-20% IPC improvements (GCN 4 vs 1 or 2) and 20%+ higher clocks and potentially even more gains from a more favourable bandwidth to FLOP ratio (512-bit 6Gbps in 390X, vs nearly double that if they went GDDR5X).

I don't think they have the money to have multiple chips that much anymore. But AMD still seems to think they can defeat their bottlenecks but still create unbalanced chips. Polaris 10 is an improvement, but 32 ROPs and 256-bit GDDR5 still hold it back. And it remains to be seen if Vega fixed Fiji's problems. Maybe AMD doesn't realize Hawaii's ratio is the ideal GCN configuration actually...

But Xbox One X (Scorpio) is 2560SP and 384-bit, so even if it is still 32 ROPs it could provide a decisive 1060 killer, and possibly a 1070 competitor if it is 64 ROPs. Why this isn't a dGPU is odd.

They can't put GDDR RAM on an interposer for bandwidth starved APU's. That's where they are heading and they need to do the research and have test bed products to get there.

 

[3DVagabond]

Sure, it is always a business decision, but, that don't mean it is the correct decision.
There is no possible way AMD was thinking that HBM2 was gonna be cheaper than GDDR5(X), yet, they seem to have bet the farm on HBM2. Sure, it is supposed to be cheaper than HBM, but I doubt it is a substantial savings. I also do realize that they most likely couldn't afford to do a dual memory design Vega chip.
That also seems to be one of the reason why the chip is so massive, to support the HBM2 tech, but, that isn't 100% clear yet. We need to see a die shot to see what is eating up all the space.

What is your source for this? Has it been posted by someone who would have a clue? Or, are you educated about HBM2 controllers?

Ryzen gives life to AMD but Vega only gives Hope, that's why AMD doesn't care so much about Perf of Vega.Ryzen is everywhere : Tablet , Mobile , Desktop , Server , Datacenter and etc....

I don't think it's they don't care but they do have to prioritize their resources. Ryzen was make or break for the company and Vega isn't to that level of importance for sure. But they are counting on this tech in the future. The sooner that future comes the better.

 

[CatMerc]

I believe that AMD and Nvidia have a cross-licensing agreement. Otherwise they'd both constantly be infringing on each others' patents.

They do not. They just both know that suing the other would be a disaster for both sides, as both have incredibly crucial patents. It would be MAD.

Polaris was just a stopgap. In many ways you could argue a poor one as they could have just shrunk Hawaii on 14nm and gotten probably better performance. I think they were trying to see what efficiency gains could be had by adjusting certain parts though (and felt it was fairly safe due to where they were positioning it; they could make up for losing SPs with higher clockspeeds). I feel like AMD really didn't have a good grasp on GCN, like where it was really bottlenecked or how its performance would change depending on different things they would change (hence the very unbalanced Fiji). Also with Microsoft talking up the way they chose the design for Scorpio based on their simulated performance stuff, that leaves me under the impression that AMD didn't have anything that robust/sophisticated to gauge things themselves.

The chances of that are nill. They have far more knowledge of the internal workings of GCN than any one of us, or even our entire collective knowledge. They know where the bottlenecks are before the chip is even out.

It's a game of trade offs, and why one chooses certain tradeoffs is sometimes not readily apparent. In the case of Fiji for example, it's because they were at the reticle limit and literally couldn't add anything else without it being unmanufacturable. They wanted the extreme TFLOPS number for pure raw shader performance, and that came at the cost of balance with the rest of the chip.

 

[Head1985]

It's been said several times that AMD should have long ago realized Hawaii was the optimal GCN configuration.

2816:64:512-bit in Polaris would provide 15-20% IPC improvements (GCN 4 vs 1 or 2) and 20%+ higher clocks and potentially even more gains from a more favourable bandwidth to FLOP ratio (512-bit 6Gbps in 390X, vs nearly double that if they went GDDR5X).

I don't think they have the money to have multiple chips that much anymore. But AMD still seems to think they can defeat their bottlenecks but still create unbalanced chips. Polaris 10 is an improvement, but 32 ROPs and 256-bit GDDR5 still hold it back. And it remains to be seen if Vega fixed Fiji's problems. Maybe AMD doesn't realize Hawaii's ratio is the ideal GCN configuration actually...

But Xbox One X (Scorpio) is 2560SP and 384-bit, so even if it is still 32 ROPs it could provide a decisive 1060 killer, and possibly a 1070 competitor if it is 64 ROPs. Why this isn't a dGPU is odd.

hawaii was last good GPU from ATI/AMD.And should be also last GCN GPU.After hawaii they should just create brand new architecture and forget about GCN.Amd need their own maxwell.Just burry super power-hungry GCN and create something new.Because GCN will never be power efficient.

Vega is still GCN.They just add some new stuff like primitive shaders and tile rasterizer, but its still 6years old GCN from fermi days.

 

[CatMerc]

hawaii was last good GPU from ATI/AMD.And should be also last GCN GPU.After hawaii they should just create brand new architecture and forget about GCN.Amd need their own maxwell.Just burry super power-hungry GCN and create something new.Because GCN will never be power efficient.

Vega is still GCN.They just add some new stuff like primitive shaders and tile rasterizer, but its still 6years old GCN from fermi days.

That is an irrelevant non statement.
An architecture could still be called GCN and be so different it might as well be called something else. Or it could be called something else but still be GCN underneath.

There are two times where it makes sense to completely go at something from scratch:
First is when the architecture is fundamentally unworkable, example being Bulldozer. But GCN is not Bulldozer, it's actually a fairly elegant architecture that has some weaknesses that need to be addressed. How one addresses some of them will send a ripple effect throughout the architecture, changing some fundemental parts of it, which we did see in Linux drivers.

The second is when the workloads in the landscape change and you want to address that. VLIW to GCN was that. VLIW could have been iterated upon to remain fairly competent at graphics, but it would fall flat on its face with compute. And AMD wanted to address compute, so we got GCN. The landscape only changed in GCN's favor, so this reason doesn't make sense either.

 

[tential]

I can't see how I'll get acceptable mini ITX performance out of Vega at all. Seems like it loses that whole market.

 

[tg2708]

AMD will deliver. Have a great Day to you all.

 

[Head1985]

That is an irrelevant non statement.
An architecture could still be called GCN and be so different it might as well be called something else. Or it could be called something else but still be GCN underneath.

There are two times where it makes sense to completely go at something from scratch:
First is when the architecture is fundamentally unworkable, example being Bulldozer. But GCN is not Bulldozer, it's actually a fairly elegant architecture that has some weaknesses that need to be addressed. How one addresses some of them will send a ripple effect throughout the architecture, changing some fundemental parts of it, which we did see in Linux drivers.

The second is when the workloads in the landscape change and you want to address that. VLIW to GCN was that. VLIW could have been iterated upon to remain fairly competent at graphics, but it would fall flat on its face with compute. And AMD wanted to address compute, so we got GCN. The landscape only changed in GCN's favor, so this reason doesn't make sense either.

i dont agree GCN is good architekture.It WAS good in fermi days and kepler days.But it is clearly inferior to maxwell/pascal by large margin.Its like NV would stay on fermi 6 years and only add some new stuff like delta color compression.Thats what AMD doing- staying on their fermi for 6 years now.
They need change it because GCN is old power hungry architecture like fermi was.

 

[tamz_msc]

i dont agree GCN is good architekture.It WAS good in fermi days and kepler days.But it is clearly inferior to maxwell/pascal by large margin.Its like NV would stay on fermi 6 years and only add some new stuff like delta color compression.Thats what AMD doing- staying on their fermi for 6 years now.
They need change it because GCN is old power hungry architecture like fermi was.

GCN is good precisely because AMD can use it to compete with Tesla, Quadro and Geforce - all with one architecture. Just like on the CPU front with Zen.