Vega Frontier Edition for DC?

[crashtech]

I'm wondering how the upcoming Vega Frontier Edition will do at DC. Typically the pro cards have the specs, with the drivers seeming to be the problematic part. Since it's not likely to see any benchmarks that would indicate Vega's true DC capabilities, the only way to know might be to take a chance and buy one. Curious to know what you guys think.

 

[StefanR5R]

AnandTech's graphics card reviews used to include FAHBench, which doesn't tell absolute F@H performance but at least relative F@H performance. Latest GPU reviews dropped FAHBench for an unknown reason.

With Vega FE having same shader count as Fury X, ~1.5 times the clock, and architectural improvements with the goal of better utilization/ more instructions per clock, maybe a rough estimate of Vega FE's DC throughput would be 1.5 times that of Fury X. Could be more if those architectural improvements are highly effective on existing OpenCL applications. Could be less if not.

A Fury X entry exists in OCN's F@H PPD database, but it is highly unreliable due to too few samples (11 currently).

As far as I recall, Vega's FP64:FP32 ratio is not yet announced but is said to be just as weak as Fiji's, i.e. 1:16.

 

[crashtech]

Fury times 1.5 would be around 1070 performance in DC, which would be rather disappointing. The other, perhaps more important question is if there will be any way to know if DC apps can make proper use of the new hardware when a gaming driver is expected and not found.

 

[StefanR5R]

Hmm, aren't Nvidia : AMD performance ratios across the board between the various DC applications?

Re driver: Not having much experience with GPU computing, my theoretical guess is that you merely need a working OpenCL driver, regardless if gaming or otherwise. (I never had issues with the now somewhat old W7000 = GCN1.0. I used it with two different driver versions since I got into DC. But I accept from what I have read here and in other forums that driver compatibility can be a huge issue.)

 

[StefanR5R]

PS,
Tom's Hardware had workstation GPU reviews occasionally, with roughly half of the benchmarks targeted towards graphics performance, half towards computing performance. What I took from those reviews was that there are very wide differences between applications WRT how they perform on Quadros vs. FirePros, and WRT how they scale from small to large cards.

 

[iwajabitw]

I think, unfortunately so, that the DC world is way behind the gaming world. Top card in gaming doesn't translate to DC that well enough for me to justify the costs. Based on some of the results for tasks we have talked about through the Pentathlon and FB I don't see the results of even a 1080Ti being worth the costs for some gpu projects that simply cant take advantage of it. DC is kind of like an old hot rod. Needs gas and good spark, but superchargers and nitros, are just eye candy that it cant benefit from. I think Mark had posted in the Pentathlon thread his 1080ti was crunching Einstein tasks, 2 at a time in 22 minutes, if I remember correctly. My 980's were doing it in 26minutes, the 280x's, 19 minutes. So that translates to me, 2 more tasks per hour on older hardware. Now, not to start a flame war, but for budget sake, I don't see the worth in using the latest hardware for projects that upgrade there software, hardware, at a much slower rate than the gaming world. And that's because the projects are mostly underfunded, and rely on donations for there costs.

 

[StefanR5R]

Re 1080Ti against 980, Einstein task times are

1080Ti ... 11.5 - 13 minutes ............. @1.85 GHz, average 175 W
1080 ...... 17.5 - 18.5 minutes .......... @1.86 GHz, average 120 W​

when these cards are loaded with 2 tasks at a time (i.e. mean time between task completions is half of those durations), using a 2.4 GHz feeder CPU. The cards could run Einstein at 2.0 GHz too for higher throughput at less energy efficiency. And task times were better and less variable when I had a faster CPU. --- Throughput of 10x0 cards (Pascal) can be very well compared with 9x0 cards (Maxwell) based on shader count and clock, since they are architecturally very similar. Energy efficiency of Pascal is better than Maxwell of course.

 

[crashtech]

Re: SP/DP ratio, if 16:1 is seen as weak, then what do we say about Nvidia's current 32:1 ratio for consumer cards? Even the Titan has this ratio now; more favorable DP performance now being limited to the highest-end Quadro and Tesla cards. It was my speculation that the Vega Frontier edition would perhaps feature better DP performance, since AMD doesn't have the ability to produce as wide a range of products for specific purposes.

 

[StefanR5R]

Re: SP/DP ratio, if 16:1 is seen as weak, then what do we say about Nvidia's current 32:1 ratio for consumer cards?

"Even weaker"? This reminds me that whenever DC performance or efficiency of AMD and NVIDIA cards are compared, we need to keep in mind that the NVIDIA versions of most DC GPU applications have higher CPU requirements than respective AMD versions.

It was my speculation that the Vega Frontier edition would perhaps feature better DP performance, since AMD doesn't have the ability to produce as wide a range of products for specific purposes.

Let's see when official specs arrive. However, either the FP64 market is not that attractive, or its prime time has not arrived yet e.g. due to VRAM limits.

 

[StefanR5R]

With Vega FE having same shader count as Fury X, ~1.5 times the clock, and architectural improvements with the goal of better utilization/ more instructions per clock, maybe a rough estimate of Vega FE's DC throughput would be 1.5 times that of Fury X. Could be more if those architectural improvements are highly effective on existing OpenCL applications. Could be less if not.

@crashtech, computing benchmarks of Vega RX (not FE) directly versus Fury X show that the performance factor of 1.5 is more of an upper bound, with real factors anywhere between 1.0 and 1.5:

http://www.anandtech.com/show/11717/the-amd-radeon-rx-vega-64-and-56-review/17
https://www.computerbase.de/2017-08/radeon-rx-vega-64-56-test/4/#diagramm-compubench​

Computing performance per Watt seems to be slightly improved relative to Fury X if you don't raise the power target of Vega RX.

 

[crashtech]

There's apparently no reason to buy Vega for DC. GPU prices are so inflated right now, I may just wait until there is a big enough cryptocurrency dip/crash to bring prices back in line. I did want to add to the fleet before this winter's races, though.

Edit: I'll probably just add another lowly 1060, since those have not been affected by the mining craze.

 

[ao_ika_red]

A bit off topic, even if 7970 is regarded as an old tech, its FP64 perf is still among the best in the consumer card category. I prefer to look for a decent second hand of it rather than buy an inflated-price card.

 

[crashtech]

I have 3 7970's and 1 R9 280x (same thing), they do well for their age. My problem is that I have hit my power budget, so any increase in computing power must be by replacing cards with higher performing, more efficient ones.

 

[ao_ika_red]

Sad thing is, even with 14nm tech, Vega power consumption is almost the same with 7970. But, its FP32 performance is very compelling. I think, in a few months, there will be a story about undervolting / underclocking of Vega that we've witnessed with early Polaris card.