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

[IntelUser2000]

Even assuming that Vega 10 can sustain boost clocks only around 1450 Mhz we are looking at 37.5% higher clocks and 40% more transistors. Vega 10 looks to be around 35% faster than Fiji. This is a regression in perf/sqmm given the 14LPP process brings close to 2x (close to 100%) the transistor density. Vega regresses in terms of perf/watt in a big way given that 14LPP brings a 50% power reduction over 28nm.

14LPP brings 50% power reduction with iso-perf settings. Since they used process to improve frequency, the 50% doesn't apply. Claims with process are always one or the other. The word "AND" commonly claimed regarding next process should properly be worded "OR". As in "14nm process brings 20% performance improvement OR 37% power reduction.

And since process gains are marketing numbers, the 50% likely only applies to super low power and frequency chips.

We can also look at it this way.

14nm(With Vega) allowed a 484mm2 chip to beat a previous 596mm2 chip(Fiji) by 35%. Because even Nvidia claimed they needed redesign to allow higher frequencies with Pascal. AMD said that too with Vega. That's why traditional Moore's Law scaling is dead. You need a architectural redesign to get whatever the process claims it'll bring.

But using prior estimations, two Polaris would be 180% of GTX 1080 in DX12 and still cost less!

I'm pretty sure they'd have went that way if it was that much better. But there's just no guarantee. They designed Vega with the belief they'd achieve whatever goals they'd set. The 2x Polaris could have easily screwed up just as with Vega.

 

[AtenRa]

And since process gains are marketing numbers, the 50% likely only applies to super low power and frequency chips.

According to official document, its not at low power or clocks

https://www.globalfoundries.com/sites/default/files/product-briefs/product-brief-14lpp.pdf

We can also look at it this way.

14nm(With Vega) allowed a 484mm2 chip to beat a previous 596mm2 chip(Fiji) by 35%. Because even Nvidia claimed they needed redesign to allow higher frequencies with Pascal. AMD said that too with Vega. That's why traditional Moore's Law scaling is dead. You need a architectural redesign to get whatever the process claims it'll bring.

AMD managed to almost have the same performance with less transistors at 100W lower power using 14nm LPP one year earlier with RX 480 vs R9 390X.

Now one year later, with 40% more transistors than Fury X and with HBM 2 they only got 30-35% higher performance with even higher power (275W vs 345W) and nobody find it a little bit peculiar ??

 

[IntelUser2000]

According to official document, its not at low power or clocks

https://www.globalfoundries.com/sites/default/files/product-briefs/product-brief-14lpp.pdf

That doesn't tell me anything about the performance at the voltages high end CPUs and GPUs run at. At 0.85V, most GPUs would crash. An extremely high proportion of them would crash in the numbers that AMD will ship to consumers. Also, they could have measured that using much simpler circuits, like SRAM(caches). Things like much reduced leakage in FinFET processes have an unrealistically big impact on those, unlike real products.

Now one year later, with 40% more transistors than Fury X and with HBM 2 they only got 30-35% higher performance with even higher power (275W vs 345W) and nobody find it a little bit peculiar ??

Not saying it could have been better, but Fury to Vega comparison itself suggests an improvement. Like someone said, there's no way to guarantee double Polaris would have been better. Plus, Vega has additional features.

 

[AtenRa]

Not saying it could have been better, but Fury to Vega comparison itself suggests an improvement. Like someone said, there's no way to guarantee double Polaris would have been better. Plus, Vega has additional features.

Fury to what we know today of VEGA is a huge regression both in IPC, perf/watt, perf/mm2 etc etc. Just compare Polaris 10 RX480 to Hawaii R9 290X/390X and you will see that there is something completely strange with VEGA if the performance is close to GP104 (GTX1080) at 345W TDP.

 

[.vodka]

Yeah.

RX Vega reviews either are going to be glorious and AMD has been trolling us all this year -at expense of their image- and they've had VegaFE crippled everywhere all this time save for pro apps... or they're going to be a disaster. I'm going with the latter.

 

[raghu78]

That doesn't tell me anything about the performance at the voltages high end CPUs and GPUs run at. At 0.85V, most GPUs would crash. An extremely high proportion of them would crash in the numbers that AMD will ship to consumers. Also, they could have measured that using much simpler circuits, like SRAM(caches). Things like much reduced leakage in FinFET processes have an unrealistically big impact on those, unlike real products.

Not saying it could have been better, but Fury to Vega comparison itself suggests an improvement. Like someone said, there's no way to guarantee double Polaris would have been better. Plus, Vega has additional features.

Nvidia has been able to improve perf by 80% from Titan Maxwell to Titan Pascal. This is in line with what we are used to with a new process node. AMD has failed to do this . Simple as that.

GM102 - 8 billion transistors TSMC 28nm
GP102 - 12 billion transistors (80% perf improvement) TSMC 16FF+

Fiji - 8.9 billion transistors TSMC 28nm
Vega - 12.5 billion transistors (35% perf improvement) GF 14LPP

Whatever problems AMD has with improving perf are unique to them as Nvidia is executing like clockwork. Volta GV102 on TSMC 12FFN is going to bring another 50% perf improvement over GP102. There will be a die size increase from GP102 to GV102. But the key metric - perf/watt is going to increase in a big way.This is why AMD are failing hard with GCN and Vega.

btw 14LPP brings 2x improvement in transistor density and almost 60% power reduction at iso-perf. The or is only when looking at perf improvement or power reduction - either 50% perf improvement or 60% power reduction.

 

[Rasterizer]

HBM rises to 95C and levels off... is the MEMORY thermally limited?! Could this be making real bandwidth much lower than design goals? We know they had to overvolt HBM and still couldn't meet their clock rate target.

Now THIS is an interesting question, because it may help explain why Vega FE appears to have 20% less memory bandwidth than Fiji (!!!), and an even worse (!!!!) regression in effective texture bandwidth vs. Fiji:

I would think that it is not reasonable to suggest that RTG would have deliberately allowed such regressions at the design stage, which means either they discovered the problem in hardware too late too fix it (which seems unlikely given all the delays), or the drivers have been seriously borked in some way (which might explain the repeated delays of the RX Vega launch and NDA drop). Testing has already shown that Vega FE is memory bandwidth bound in gaming and in ETH hashrate, so any improvement in memory bandwidth would likely directly improve performance in those areas.

IF the issue is thermal, even partially, then it will be interesting to see if AVFS (which was not enabled on Vega FE) makes any difference on that front.

 

[Paratus]

You heard it here first. Volta is 50% faster. Not 60%, not 40% but 50%.

 

[HurleyBird]

RX Vega reviews either are going to be glorious and AMD has been trolling us all this year -at expense of their image- and they've had VegaFE crippled everywhere all this time save for pro apps... or they're going to be a disaster. I'm going with the latter.

AMD isn't trolling us. If the RX Vega reviews show substantially better performance than what we've been led to expect, it will be down to last minute improvements in the drivers.

 

[Glo.]

AMD isn't trolling us. If the RX Vega reviews show substantially better performance than what we've been led to expect, it will be down to last minute improvements in the drivers.

Drivers will make the hardware at best 20% faster than 1.6 GHz Fiji, with properly working load balancing, BIOS control over it(Load balancing requires overclocking, and declocking different parts of shader engine to balance the long and short running shaders in each part of the display engine), and tiled rasterization. So I wouldn't expect miracles on this front.

What will make Vega faster, per core, and per clock than Fiji, are Primitive Shaders, and Programmable Geometry Pipeline.

 

[swilli89]

Anyone still holding out for surprise RX Vega performance at this point is fooling themselves. We've already had AMD pit the Vega 64 vs 1080, and that's in cherry picked, internal benchmarks. Not only that but they are already putting out the qualifier statements like "the very best card... under $499" which is a fine thing to do.. but it tells us we shouldn't expect more performance.

Either something very wrong happened with Vega to see such poor performance, or we are indeed facing down the reality that it takes billions of dollars to design a cutting edge chip which AMD simply did not have available to fund Vega.

 

[Glo.]

Anyone still holding out for surprise RX Vega performance at this point is fooling themselves. We've already had AMD pit the Vega 64 vs 1080, and that's in cherry picked, internal benchmarks. Not only that but they are already putting out the qualifier statements like "the very best card... under $499" which is a fine thing to do.. but it tells us we shouldn't expect more performance.

Either something very wrong happened with Vega to see such poor performance, or we are indeed facing down the reality that it takes billions of dollars to design a cutting edge chip which AMD simply did not have available to fund Vega.

If 768 core GCN4 GPU is on par with 768 CUDA core GPU, with similar memory bandwidth. clock for clock, and core for core in gaming performance, as yourself, why is 4096 GCN core chip, clocked at 1.5 GHz not able to get to level of performance delivered by 3584 CUDA core GPU?

Judging by performance of previous generations of GCN, and how it compared clock for clock, core for core, vs Pascal, it is very strange that Vega is not able to achieve GTX 1080 Ti levels.

That comparison I have not pulled out of my ass.

We have bought Macbook Pro 15 inch 2017, with Radeon Pro 555. It has 855 MHz, and 81 GB/s Bandwidth.
Then I have provided my personal GTX 1050 Ti KalmX for tests. We declocked the GTX to 900 MHz, because that was all we were able to set. The difference in Nvidia optimized games was 3 FPS for Nvidia GPU.

Similar pattern we see with GTX 1070 declocked to RX 470 levels of performance.

 

[IntelUser2000]

If 768 core GCN4 GPU is on par with 768 CUDA core GPU, with similar memory bandwidth. clock for clock, and core for core in gaming performance, as yourself, why is 4096 GCN core chip, clocked at 1.5 GHz not able to get to level of performance delivered by 3584 CUDA core GPU?

Maybe after years of designing small die GPUs they are not used to making a large die one. For years they would take a small die one and double it up with X2 versions. On top of that, they went absolutely bleeding edge with HBM memory on a consumer variant nonetheless. Having HBM only with HPC-class GPUs as with Quadro means low yield and volume isn't that big of a problem, neither is cost. HBM's problem is not just costs in interposer and packaging but the difficulty in doing so.

Nehalem architect Glenn Hinton said while they could have architected the core for even more performance they chose not to do so, because it would be risky for the project. Risk, if it doesn't work out ends up in a product that misses its target date and performance. So proper risk management is just as important as pure specs.

 

[lobz]

AMD managed to almost have the same performance with less transistors at 100W lower power using 14nm LPP one year earlier with RX 480 vs R9 390X.

Now one year later, with 40% more transistors than Fury X and with HBM 2 they only got 30-35% higher performance with even higher power (275W vs 345W) and nobody find it a little bit peculiar ??

Vega has a lot more features than Furx X which do not translate to (immediate) performance gains in gaming. In the professional space Vega is threatening the Pascal cards, the FE has already twice the performance/$ at around 90% the performance in lots of pro software with launch drivers versus the P6000. I'll be very curious about the WX9100 reviews.

 

[Konan]

You heard it here first. Volta is 50% faster. Not 60%, not 40% but 50%.

Remember: NVIDIA said in its white paper describing the Volta architecture that Volta "features a major new redesign of the [streaming multiprocessor] architecture that is at the center of the GPU" and that this new design "is 50% more energy efficient than the previous generation Pascal design."

Based on that I'm expecting 50% energy efficiency, 50% performance we'll have to see but totally plausible.

 

[lobz]

delete

 

[lobz]

You heard it here first. Volta is 50% faster. Not 60%, not 40% but 50%.

alright... I wanna hear it here first. Which Volta GPU is 50% faster than which NVIDIA/AMD GPU?

 

[Paratus]

Nvidia has been able to improve perf by 80% from Titan Maxwell to Titan Pascal. This is in line with what we are used to with a new process node. AMD has failed to do this . Simple as that.

GM102 - 8 billion transistors TSMC 28nm
GP102 - 12 billion transistors (80% perf improvement) TSMC 16FF+

Fiji - 8.9 billion transistors TSMC 28nm
Vega - 12.5 billion transistors (35% perf improvement) GF 14LPP

Whatever problems AMD has with improving perf are unique to them as Nvidia is executing like clockwork. Volta GV102 on TSMC 12FFN is going to bring another 50% perf improvement over GP102. There will be a die size increase from GP102 to GV102. But the key metric - perf/watt is going to increase in a big way.This is why AMD are failing hard with GCN and Vega.

btw 14LPP brings 2x improvement in transistor density and almost 60% power reduction at iso-perf. The or is only when looking at perf improvement or power reduction - either 50% perf improvement or 60% power reduction.

alright... I wanna hear it here first. Which Volta GPU is 50% faster than which NVIDIA/AMD GPU?

Need to ask Raghu. He said it

 

[lobz]

Need to ask Raghu. He said it

Ahh sorry then

 

[raghu78]

Need to ask Raghu. He said it

I am just extrapolating GV102 perf from GV100 perf 15 tflops vs GP100 10.6 tflops. Thats a 41.5% perf increase

http://www.anandtech.com/show/11367...v100-gpu-and-tesla-v100-accelerator-announced

We can surely expect 40-50% perf increase over GP102 from GV102.

 

[leoneazzurro]

Do you understand that teraflops performance <> gaming performance, and this is valid for Nvidia architectures, too?

 

[PeterScott]

I am just extrapolating GV102 perf from GV100 perf 15 tflops vs GP100 10.6 tflops. Thats a 41.5% perf increase

They also have ~43% more stream processors, so it really depends on how much they want to increase the stream processor count and die size. 16nm->12nm FFN is not much of smaller process, if that is the process they are going with.

 

[Malogeek]

There's lots of reasons why such an extrapolation of expected performance is just silly.

 

[JDG1980]

If 768 core GCN4 GPU is on par with 768 CUDA core GPU, with similar memory bandwidth. clock for clock, and core for core in gaming performance, as yourself, why is 4096 GCN core chip, clocked at 1.5 GHz not able to get to level of performance delivered by 3584 CUDA core GPU?

Because the 4096-core GCN chips are bottlenecked by low ROP counts (64) and a narrow front-end (4 triangles/clock). In contrast, both GM200 (980 Ti) and GP102 (1080 Ti) have 96 ROPs and can do 6 triangles/clock.
Also, as noted previously, Vega isn't delivering expected memory bandwidth. They wanted to double bandwidth per pin over Fury X (this is even touted in the release day slides) but only got ~1.8x per pin, and that only by overvolting and overclocking. And it seems to be thermally throttled as well (liquid cooling or better air cooling on AIB cards might help with that, at least).

 

[Head1985]

Anyone still holding out for surprise RX Vega performance at this point is fooling themselves. We've already had AMD pit the Vega 64 vs 1080, and that's in cherry picked, internal benchmarks. Not only that but they are already putting out the qualifier statements like "the very best card... under $499" which is a fine thing to do.. but it tells us we shouldn't expect more performance.

Either something very wrong happened with Vega to see such poor performance, or we are indeed facing down the reality that it takes billions of dollars to design a cutting edge chip which AMD simply did not have available to fund Vega.

I dont know why you guys still speculating about gaming performance.Vega cards will be sell out 5min after launch or at 1000+ USD because miners.And it will continue few years.You wont be abble buy vega for gaming unless pying 1000usd so dont bother.
Bitcoin is up 700USD from 2days ago and all time maximum and other coins going up as well.There will be even bigger mining craze in few days/weeks than current one.