tyan + k20

Discuss anything you like on this forum.
Post Reply
User avatar
gabrielefx
Licensed Customer
Posts: 1701
Joined: Wed Sep 28, 2011 2:00 pm

Tyan released the all new barebones with Tesla K20!

http://www.isycorp.com/Home_i.ASP?pagina=7
quad Titan Kepler 6GB + quad Titan X Pascal 12GB + quad GTX1080 8GB + dual GTX1080Ti 11GB
Timmaigh
Licensed Customer
Posts: 168
Joined: Mon Nov 01, 2010 9:52 pm

Nice. The price for these Teslas is however just :lol: :lol: :lol:

Anyway, i would still be interested in their Octane performance, especially given those promoted new features like thread spawning or whatever it was... so i am fairly disappointed with the Otoy/Refractive approach to this kind of stuff - i am not aware of a single official word on this topic around here.
Intel Core i7 980x @ 3,78GHz - Gigabyte X58A UD7 rev 1.0 - 24GB DDR3 RAM - Gainward GTX590 3GB @ 700/1400/3900 Mhz- 2x Intel X25-M G2 80GB SSD - WD Caviar Black 2TB - WD Caviar Green 2TB - Fractal Design Define R2 - Win7 64bit - Octane 2.57
User avatar
abstrax
OctaneRender Team
Posts: 5510
Joined: Tue May 18, 2010 11:01 am
Location: Auckland, New Zealand

Timmaigh wrote:Nice. The price for these Teslas is however just :lol: :lol: :lol:

Anyway, i would still be interested in their Octane performance, especially given those promoted new features like thread spawning or whatever it was... so i am fairly disappointed with the Otoy/Refractive approach to this kind of stuff - i am not aware of a single official word on this topic around here.
Huh? What are you talking about??
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra
Timmaigh
Licensed Customer
Posts: 168
Joined: Mon Nov 01, 2010 9:52 pm

abstrax wrote:
Timmaigh wrote:Nice. The price for these Teslas is however just :lol: :lol: :lol:

Anyway, i would still be interested in their Octane performance, especially given those promoted new features like thread spawning or whatever it was... so i am fairly disappointed with the Otoy/Refractive approach to this kind of stuff - i am not aware of a single official word on this topic around here.
Huh? What are you talking about??
Lets say hypothetically i have money for this Tesla and i want to buy it for Octane. I have no idea about its performance though, and going with the past experience (other kepler cards) i could do a terrible mistake buying one of these. Clearly, there are no hardware review sites ever going to test the performance of this piece of HW within Octane, so what are my options now? How do i find out, whether to buy or not?
I am not saying you are in any way obliged to promote Nvidia products, but few words here and there about the performance of their HW would be nice. Obviously, if you know it, i assume you kinda cooperate with Nvidia and have some info we dont, but i could be wrong. So perhaps now would be the right time to clear this out :)
Intel Core i7 980x @ 3,78GHz - Gigabyte X58A UD7 rev 1.0 - 24GB DDR3 RAM - Gainward GTX590 3GB @ 700/1400/3900 Mhz- 2x Intel X25-M G2 80GB SSD - WD Caviar Black 2TB - WD Caviar Green 2TB - Fractal Design Define R2 - Win7 64bit - Octane 2.57
User avatar
abstrax
OctaneRender Team
Posts: 5510
Joined: Tue May 18, 2010 11:01 am
Location: Auckland, New Zealand

Well, the reason why we don't write anything about it is, because we don't know. We don't have any Tesla GPGPU boards in our office. OTOY LA did some tests with the GK104 based K10 and it runs slower than a GTX 690, because it has a lower clock rate, but has more memory and is passively cooled, which allows it to be used in a server environment.

I think we said it many times before: Octane is designed for consumer GPUs. You can run it on Tesla boards, but the performance per dollar is probably not worth it, because advanced HPC features like high double precision performance or ECC RAM are not used or have no benefit in Octane. And personally I don't think that will change in the near future.

Or see it the other way around: We have been putting a lot of effort into making Octane run fast on consumer graphics hardware, which has the highest performance per dollar due to market pressure - as long as you stay in single precision land. And we intend to continue doing that.

Cheers,
Marcus
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra
Timmaigh
Licensed Customer
Posts: 168
Joined: Mon Nov 01, 2010 9:52 pm

abstrax wrote:Well, the reason why we don't write anything about it is, because we don't know. We don't have any Tesla GPGPU boards in our office. OTOY LA did some tests with the GK104 based K10 and it runs slower than a GTX 690, because it has a lower clock rate, but has more memory and is passively cooled, which allows it to be used in a server environment.

I think we said it many times before: Octane is designed for consumer GPUs. You can run it on Tesla boards, but the performance per dollar is probably not worth it, because advanced HPC features like high double precision performance or ECC RAM are not used or have no benefit in Octane. And personally I don't think that will change in the near future.

Or see it the other way around: We have been putting a lot of effort into making Octane run fast on consumer graphics hardware, which has the highest performance per dollar due to market pressure - as long as you stay in single precision land. And we intend to continue doing that.

Cheers,
Marcus
Thank you for the answer.
I understand your emphasis on consumer cards and i like it, given the price of the Teslas. But currently we are in the situation, where the latest generation of consumer cards is slower than the previous one and i am not holding my breath that this is going to be any better. Or if you want, i am not sure, that what yesterday appeared as Tesla K20 is going ever to hit the market as Geforce.

We were recently buying new computer into office and the choice of components was up to me, as i am the only one to have any idea about this stuff...and i did not know, which GPU to buy, i tried to explain the boss that even if we buy lets say gtx670, its not really going to be any faster than current card we have (470)...he said "so lets buy 4 of those 670/680"... but that would be pretty stupid thing to do IMHO, given all the heat and noise those 4 cards would produce, especially if one of these new Teslas is perhaps going to be as fast as these 4 cards combined...thats i would like to know, i realize i do not need dual precision or ECC, but how about that Hyper-Q stuff and the other so much promoted new feature? What if these actually make some significant difference in case of Octane, would we not be better then paying 3 thousands EUROs for one Tesla instead of 2 thousands for 4 680s?

EDIT: What i wanted to say, even though you say Octane is designed for consumer cards, i do not think this is your decision. Its up to Nvidia. Their next generation of consumer cards may not even have CUDA functionality anymore, what is going to happen in such case?
Intel Core i7 980x @ 3,78GHz - Gigabyte X58A UD7 rev 1.0 - 24GB DDR3 RAM - Gainward GTX590 3GB @ 700/1400/3900 Mhz- 2x Intel X25-M G2 80GB SSD - WD Caviar Black 2TB - WD Caviar Green 2TB - Fractal Design Define R2 - Win7 64bit - Octane 2.57
User avatar
abstrax
OctaneRender Team
Posts: 5510
Joined: Tue May 18, 2010 11:01 am
Location: Auckland, New Zealand

Well the math is pretty simple: A K20 has 13 multi-processors with 192 cores each and a clock rate of 705 MHz and will probably cost 3000-3100 Euros excl. GST/VAT. A GTX 690 has in total 16 multi-processors with 192 cores each and a clock rate of 900 - 1000 MHz and costs in Germany ~900 Euros incl. VAT (19%).

So, where do you think Octane runs faster and which one is the better deal ;)

All the new features every new CUDA release brings are usually aimed at HPC applications and special needs, so there is usually no real advantage for Octane. And please don't forget that we are currently supporting all CUDA GPUs including those with compute model 1.x...

Cheers,
Marcus
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra
User avatar
abstrax
OctaneRender Team
Posts: 5510
Joined: Tue May 18, 2010 11:01 am
Location: Auckland, New Zealand

And one more thing: The Kepler GPUs really run very silent. You really can't compare them with the first generation Fermi GPUs (GTX 400 series). And a GTX 690 runs roughly as fast as a GTX 590. A GTX 680 is slower than a GTX 580, but should be at least as fast as a GTX 480 (which had one multi-processor fused off due to thermal issues).

Cheers,
Marcus
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra
Timmaigh
Licensed Customer
Posts: 168
Joined: Mon Nov 01, 2010 9:52 pm

abstrax wrote:Well the math is pretty simple: A K20 has 13 multi-processors with 192 cores each and a clock rate of 705 MHz and will probably cost 3000-3100 Euros excl. GST/VAT. A GTX 690 has in total 16 multi-processors with 192 cores each and a clock rate of 900 - 1000 MHz and costs in Germany ~900 Euros incl. VAT (19%).

So, where do you think Octane runs faster and which one is the better deal ;)
Yeah, i suppose that is true, if only things that matter are number of CUDA cores and clockrate. But is that true? Why is gtx680 slower than 580 then, if it has more cores/higher clocks? From what i could gather, its due to the architectural changes within Kepler chip, which lacks some HPC features, that Fermi cards (even the gaming ones) used to have. With this big Kepler chip, these capabilities may be back, alongside the new stuff i talked about. Based on this, we could perhaps expect this Tesla to be faster than gtx690, even though it has less cores/slower clockrate.

abstrax wrote: All the new features every new CUDA release brings are usually aimed at HPC applications and special needs, so there is usually no real advantage for Octane. And please don't forget that we are currently supporting all CUDA GPUs including those with compute model 1.x...

Cheers,
Marcus
The bolded part is what i wanted to know, basically. Still, this Hyper-Q stuff is a hardware feature, not a software one (just CUDA release related), i thought.

Anyway, thanks again for your answers. I did not want be a dick with my first post :) , i just remember days 3 years ago, when Radiance himself would specifically post over here, that buying gtx460 with 2GBs of RAM is the best option for most...and recently i had a subjective feeling, you dont get this kind of advices over here anymore and have to rely just on yourself or other users experience and their good will to share that experience to find these things out.

EDIT> Just found this on local HW site:

Image

seems Nvidia recommends Quadro for Octane :)
Intel Core i7 980x @ 3,78GHz - Gigabyte X58A UD7 rev 1.0 - 24GB DDR3 RAM - Gainward GTX590 3GB @ 700/1400/3900 Mhz- 2x Intel X25-M G2 80GB SSD - WD Caviar Black 2TB - WD Caviar Green 2TB - Fractal Design Define R2 - Win7 64bit - Octane 2.57
User avatar
abstrax
OctaneRender Team
Posts: 5510
Joined: Tue May 18, 2010 11:01 am
Location: Auckland, New Zealand

Timmaigh wrote:Yeah, i suppose that is true, if only things that matter are number of CUDA cores and clockrate. But is that true? Why is gtx680 slower than 580 then, if it has more cores/higher clocks? From what i could gather, its due to the architectural changes within Kepler chip, which lacks some HPC features, that Fermi cards (even the gaming ones) used to have. With this big Kepler chip, these capabilities may be back, alongside the new stuff i talked about. Based on this, we could perhaps expect this Tesla to be faster than gtx690, even though it has less cores/slower clockrate.
No, as long as you are comparing the same architecture, you can stick to multi-processor count, shader core count and clock rate. The reason why Keplers are slower than Fermis if you compare the core count alone, is because they are a different architecture and your parallel code needs to run more coherently than on Fermis to feed all cores, which is hard to achieve in any real world GPGPU application. And I don't know, but I don't expect huge difference between GK104 and GK110 in Octane.
Timmaigh wrote:The bolded part is what i wanted to know, basically. Still, this Hyper-Q stuff is a hardware feature, not a software one (just CUDA release related), i thought.

Anyway, thanks again for your answers. I did not want be a dick with my first post :) , i just remember days 3 years ago, when Radiance himself would specifically post over here, that buying gtx460 with 2GBs of RAM is the best option for most...and recently i had a subjective feeling, you dont get this kind of advices over here anymore and have to rely just on yourself or other users experience and their good will to share that experience to find these things out.
But what if that information is wrong or outdated? Remember OctaneRender is a commercial endeavor and with that there comes some responsibility to provide correct information... And currently we don't have the resources to provide you comprehensive and reliable information about hardware.
Timmaigh wrote:EDIT> Just found this on local HW site:

...

seems Nvidia recommends Quadro for Octane :)
See, that's just wrong. The price per performance is too high for Quadros. And Quadros are based on the same GPU as GeForce cards anyway. Which is why I would not recommend to buy anything else than GeForce for Octane. You usually get more performance by using multiple GPUs instead of buying one extremely expensive one. That's the beauty of Octane, you can just add more GPUs as long as your mainboard/powers supply/case supports it. So use this feature ;)

Cheers,
Marcus
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra
Post Reply

Return to “Off Topic Forum”