[VR-Z]AMD R7970 Preferred over NVIDIA Kepler in Real GPGPU Deployments?

[Skurge]

Once again AMD dropping the ball by not pressing their advantage while big kepler is a few months away.

http://vr-zone.com/articles/from-gt...-kepler-in-real-gpgpu-deployments-/15903.html

 

[ShintaiDK]

Thats a quite silly article. Real GPGPU deployments uses Tesla or MIC primary. And the first Tesla based Kepler just released didnt it? I dont even think it shipped yet. Not to mention that you would use CUDA or custom x86 code for MIC for GPGPU.

Arh Theo as the writer.....

A little additional hint. Check the power consumption + TDP for Tesla based cards vs the equal Geforce or Quadro. Then you know why Geforce/Quadro or Radeon cards in "Real GPGPU" is an utter joke.

 

[sontin]

People who are buying consumer cards for their "professional" applications are not the market for nVidia. So they don't care for them.

 

[Skurge]

If AMD was moving towards a more GPGPU friendly acrh. You would think they would have the 1st FirePRO and FireStream cards out by now to rival Quadro and Tesla.

 

[RussianSensation]

That article is a joke. It uses synthetic SP and DP benchmarks in SiSoftware Sandra to insinuate that HD7970 would be more suitable than Kepler architecture in real world compute applications.

Of course on paper HD7970 has unbelievable SP and esp. DP performance but that doesn't always translate into real world advantages in professional or even consumer compute applications.

Remember how many people were speculating that GK104 was going to be underwhelming for Folding@Home because every reviewer under the sun concluded that Kepler has been "neutered" for compute?

People clump Compute performance under one umbrella but it's much more complex. One has to understand what a specific program calls for exactly:

1) OpenCL performance
2) Single Precision
3) Double precision
4) How is that software optimized for a specific architecture
5) Developer support/driver updates

It's evident that NV has a clear strategy in how it segments its products in the compute market. It works to provide solutions for professionals that help them for their specific computational needs. What's AMD's compute strategy? No one knows really. They just launched an excellent GCN compute architecture but didn't market it effectively and have shown lack of support. They aren't doing any of the work to sway professionals to consider the products as viable alternatives.

Even if on paper HD7970 is superior to Kepler, unless it translates to real world advantage for professionals, it's meaningless (and I am not talking about Bitcoin mining).

Also, economics are a factor and Kepler is leading in performance/watt.

While the next 3 benches are not compute related, I just wanted to point out why drivers/and optimizations are critical for professional applications in Quadro and Tesla markets. Quadro 5000 has inferior hardware to GTX470 but look at the performance in professional apps:

In professional and compute applications, you cannot just compare videocards based on theoretical/"on-paper" performance. And that's exactly what VR-Zone has done.

 

[Skurge]

http://pcper.com/reviews/Graphics-C...-Professional-Graphics-Review/SPECviewperf-11

AMD's performance here is pretty inconsistent, but that's with the VLIW4, You would have to think that GCN would do much much better. It would keep the advantages that VLIW4 had and improve in areas where it was weaker. In any case I do agree that AMD hasn't marketed and supported GCN very well

 

[ShintaiDK]

Bingo RussianSensation

 

[Pottuvoi]

Also the current Kepler is not aimed for heavy scientific processing due to the lack of ECC.
GK110 will address pretty much all shortcomings of GK104 and will add some very interesting new abilities like launching new jobs within GPU. (GK110 whitepaper)

 

[3DVagabond]

Also the current Kepler is not aimed for heavy scientific processing due to the lack of ECC.
GK110 will address pretty much all shortcomings of GK104 and will add some very interesting new abilities like launching new jobs within GPU. (GK110 whitepaper)

Any prediction on when we're likely to see GK110?

 

[f1sherman]

By the look of these benchmarks, even 6970 should have been
preferred over GTX 580.

Sandra 2012 Real GPGPU Deployments

Theo hahaha Valich - Passionate about life... Enthusiast, Journalist...Entrepreneur in Media and Tech. Publisher of Bright Side of News* and HDTelevizija.

 

[f1sherman]

Real GPGPU Deployments

Take a wild guess what Green color stands for.

 

[HurleyBird]

By the look of these benchmarks, even 6970 should have been
preferred over GTX 580.

As it should.

6970 has a ton more theoretical SP, despite this the 580 is neck and neck thanks to it's superior compute architecture.

For DP, consumer Fermi variants are complete neutered. A Tesla Fermi variant would have slaughtered the 6970 in the DP test. Since AMD doesn't cripple DP on their consumer cards they're always going to win this battle against Nvidia's consumer cards.