Rx Vega hits SiSoft Sandra, 35% faster than gtx 1080.

[vanished]

So Vega is 35% faster than a 1080, and the 1080 Ti is 35% faster than a 1080, but the 1080 Ti is 25% faster than Vega?  Am I missing something...?

[juri-han]

personally id rather see vega on the budget end 1080 at 480 prices i know this wont happen with HBM2 tho 

[Liltrekkie]

Just now, Coaxialgamer said:

Hbm memory chips are not counted when it comes to gpu die size, and the memory controllers are smaller tgan traditional gddr controllers. 

I thought as much, though I wasn't expecting the controllers to be smaller, any idea why that is?

[Valentyn]

Just now, Coaxialgamer said:

My point was that if can't engineer a chip that's faster than the titan xp despite being considerably larger and on a more advanced node, it's going to be bad. Especially considering vega will be competing against volta until navi comes out, and there are no node shrinks in sight. 

We have to remember that unlike NVIDIA, they're not cutting out all their compute functionality.
AMD's consumer cards have always had very similar compute performance to their Pro line.
Since Maxwell NVIDIA has sliced as much out as possible and only really kept it on Tesla, and a lesser Extent Quadro.

 

All that is very unnecessary for gaming, and cutting it out could mean higher clocks like Pascal and Maxwell got. 

[Coaxialgamer]

Just now, Liltrekkie said:

I thought as much, though I wasn't expecting the controllers to be smaller, any idea why that is?

Don't know why tbh. Might have to do with clock speeds.

I know for a fact the 4096bit hbm controllers from fiji are smaller than the 512 bit gddr5 controllers from hawaii, and by quite a bit. 

[juri-han]

1 minute ago, Ryan_Vickers said:

So Vega is 35% faster than a 1080, and the 1080 Ti is 35% faster than a 1080, but the 1080 Ti is 25% faster than Vega?  Am I missing something...?

isn't the 1080ti actually only 35% faster if its OC and the 1080 isnt? lol

 

 

 

 

 

 

 

 

4

[Liltrekkie]

Just now, Coaxialgamer said:

Don't know why tbh. Might have to do with clock speeds.

I know for a fact the 4096bit hbm controllers from fiji are smaller than the 512 bit gddr5 controllers from hawaii, and by quite a bit. 

Weird, one would think the extra bus width would need a a bigger controller, but maybe this isn't the case. If your theory on clock speeds is right, maybe that's why HBM1 didn't have great memory overclocking. or very great speeds at all. But, that could have just been a limitation on HBM1 as well. 

[vanished]

4 minutes ago, juri-han said:

isn't the 1080ti actually only 35% faster if its OC and the 1080 isnt? lol

idk, I just remember that being the number that was thrown around.  But if that is true, then the current numbers make even less sense.

[Coaxialgamer]

Just now, Valentyn said:

We have to remember that unlike NVIDIA, they're not cutting out all their compute functionality.
AMD's consumer cards have always had very similar compute performance to their Pro line.
Since Maxwell NVIDIA has sliced as much out as possible and only really kept it on Tesla, and a lesser Extent Quadro.

 

All that is very unnecessary for gaming, and cutting it out could mean higher clocks like Pascal and Maxwell got. 

Nvidia didn't "cut out all the compute functionality". 

They just designed the architecture with a heavy focus on graphics performance. 

They did remove the hardware scheduler, which is neither good nor bad depending on what you need to do, but they did heavily increase the geometry and texture throughput of their chips ( one of the reasons maxwell is so good at tessellation), among other things. Note flops are not a good measure of performance, because it's just a function of clock speeds * alu count* 2.

 

The fact is nvidia is selling quite a few tesla's and quadro's, so if they were that bad for compute they'd be doing quite poorly in those markets. 

[Coaxialgamer]

26 minutes ago, jasonc_01 said:

The 1080ti is still 25% ahead of the this new Vega test in Sandra

Source? 

[jasonc_01]

8 minutes ago, Ryan_Vickers said:

So Vega is 35% faster than a 1080, and the 1080 Ti is 35% faster than a 1080, but the 1080 Ti is 25% faster than Vega?  Am I missing something...?

No, in SiSoft Sandra vega is 35% fast than the 1080 but still 25% behind the 1080ti. I dont think it translates the same to graphical performance

[jasonc_01]

2 minutes ago, Coaxialgamer said:

Source? 

https://www.pcgamesn.com/amd/amd-rx-vega-sisoft-benchmarks

Quote

The overall general purpose compute power of what we expect to be the flagship AMD Radeon RX Vega card comes in at 2756.69 Mpix/sec, while the GTX 1080 can be found with scores of 2,050.72 Mpix/sec. That’s a lead just shy of 35% for the AMD card when it comes to the computational side of the equation. Running the same test on our GTX 1080 Ti though has the latest Nvidia card running at 3,451.05 Mpix/sec, or some 25% quicker again than the RX Vega.

 

[vanished]

2 minutes ago, jasonc_01 said:

No, in SiSoft Sandra vega is 35% fast than the 1080 but still 25% behind the 1080ti. I dont think it translates the same to graphical performance

So what test we're talking about really really matters.  So blanket statements like "X is y% faster than Z" don't really have any meaning since you have to also say "in test W"

[Coaxialgamer]

1 minute ago, jasonc_01 said:

https://www.pcgamesn.com/amd/amd-rx-vega-sisoft-benchmarks

 

Thanks, didn't catch that when i read the article. Guess i need some sleep lol. 

[jasonc_01]

1 minute ago, Ryan_Vickers said:

So what test we're talking about really really matters.  So blanket statements like "X is y% faster than Z" don't really have any meaning since you have to also say "in test W"

its all sabre rattling, still interesting though

[Valentyn]

5 minutes ago, Coaxialgamer said:

Nvidia didn't "cut out all the compute functionality". 

They just designed the architecture with a heavy focus on graphics performance. 

They did remove the hardware scheduler, which is neither good nor bad depending on what you need to do, but they did heavily increase the geometry and texture throughput of their chips ( one of the reasons maxwell is so good at tessellation), among other things. Note flops are not a good measure of performance, because it's just a function of clock speeds * alu count* 2.

 

The fact is nvidia is selling quite a few tesla's and quadro's, so if they were that bad for compute they'd be doing quite poorly in those markets. 

None of the current NV consumer cards have Double Precision though, that's all been cut down. Where as Tesla is a monster for compute and deep learning. 

The GeForce cards were "cut down" in that regard.

 

AMD did the same with Fiji and Polaris, but they're still stronger than than their NV counter parts.
Hawaii is still a compute beast, both consumer and FirePro.

 

[Coaxialgamer]

8 minutes ago, Valentyn said:

None of the current NV consumer cards have Double Precision though, that's all been cut down. Where as Tesla is a monster for compute and deep learning. 

The GeForce cards were "cut down" in that regard.

 

AMD did the same with Fiji and Polaris, but they're still stronger than than their NV counter parts.
Hawaii is still a compute beast, both consumer and FirePro.

 

Well, that's not entirely true. All gp100 chips support full rate fp64 (1/2)

 

The other cards still support it, but only at 1/32 the rate of fp32, making it essentially useless.  

Amd has also been cutting down fp64 the last few gens to conserve die space, but to a lesser extent ( hawaii had 1/8, both gcn 1.2 and polaris have 1/16). 

Ironically, the hd7970/280x still has the best fp64 performance in amd's entire consumer lineup. 

 

But then again, amd's current lineup can't touch nvidia when it comes to geometry performance. 

[zMeul]

why are we still comparing compute perf?!

haven't "we" learned anything 

[Demonking]

I'll wait for official reviews before i judge this product,don't want anaother Ryzen or Fury x hype again lead by disappointments.

[Coaxialgamer]

Just now, zMeul said:

why are we still comparing compute perf?!

haven't "we" learned anything 

To be fair, this isn't just floating point performance. 

[Okjoek]

Wait isn't the 1080 ti "35%" faster than the 1080 too?

[vanished]

Hm... just recently we had leaks about the TFLOPs of Vega.  I did some calculations at that time that give results that re in the ballpark of what we're seeing here.  Interesting.

actually I just realized that might not be true... what clockspeed was Vega running during this test!?  We kinda need to know that  

Edited March 13, 2017 by Ryan_Vickers

[vanished]

Just now, Okjoek said:

Wait isn't the 1080 ti "35%" faster than the 1080 too?

That's a different benchmark/test I guess

[zMeul]

5 minutes ago, Coaxialgamer said:

To be fair, this isn't just floating point performance. 

it's compute perf via OpenCL .. and we all know how good AMD is with OpenCL and how crappy nVidia is

so ........

 

compare for example RX480 and GTX1060

[Okjoek]

2 minutes ago, Ryan_Vickers said:

That's a different benchmark/test I guess

The fact that leaked benchmarks are even placing it above the standard GTX 1080 is a pleasant sound to me