Not logged in » Login
View latest post
Back again TESLA GPUs
12.04.2015, 12:27
( 11 )
Hi again everybody,

we've got a new project based on NVIDIA Tesla GPGPU and back again we've got some questions about that. After some hardware adjustments, our costumer it's very happy with her C720 CELSIUS unit and the NVIDIA K40 GPGPU working on it. Finally we manage to slowdown fan speed decreasing sound pressure without creating a temperature problem but on future units I’ll recommend tower models for sure (check previous posts for that issue).

Anyhow, now, we have to create a new configuration with a couple of K40 or maybe the new K80 GPGPU. I've been reading about these new GPGPU and both are sold by Fujitsu although I can't find an answer to the incredible high prices on Fujitsu. I’ve look for an European NVIDIA reseller and the prices are more or less 50% less there than on System Architect, that doesn’t happen on QUADRO series. Talking about GPGPU of over 3,500€ that’s a lot of money and I can’t find a reason why I have to recommend to our costumer to buy Fujitsu’s units. Is there any reason I miss? On pictures, they seem exactly the same GPGPUs from PNY.

Moreover, for installing a couple of K40 the documentation only talks about using a RX350 S3 or TX300 S8 Primergy server models. No info about the new K80 model available at Fujitsu although it’s supposed to be available from 1 April. Also, although CELSIUS units can handle K20 and K20X there’s no mention about K40 or K80 models and even they are missing on System Architect configurators. In fact, K80 seems missing anywhere although it has been announced by Fujitsu, so no price idea. Need some info about that to make the correct configurations for our costumer.

We’ve test K40 on a CELSIUS unit without problems and seems that it’s just a problem on actualizing the documentation cause R93/40power or M73/40power models seems that can handle that GPGPUs without any problem. I guess, through the answers on this forum, that Fujitsu techs prefers Quadro than Tesla for CELSIUS even though some CELSIUS seems to be designed for HPC scenarios as we need them to. It’s that trueConfused Any reason for?

Looking for the prices on System Architect I can imagine a reason cause Quadro K6000 it’s a close configuration to K40 and there’s around 2,000€ difference between them. But that’s just because Fujitsu Tesla K40 price it’s incredible higher than what we can found directly buying the product to an NVIDIA distributors. In fact, Tesla K40 should be cheaper than Quadro K6000!!!! As Tesla is designed for pure HPC and Quadro seems to be designed better for graphic studio scenarios, shouldn’t be better Tesla option on HPC scenariosConfused

Finally, is there newer doc files about performance issues for high end computing, some of the docs we are using are quite old:

2008-03 Modular RAID for PRIMERGY (EN)
2011-09-15 wp-raid-controller-performance-ww-en
2013-07-04 wp-performance-report-primergy-pcie-ssd-iodrive2-ww-en
2013-08-01 wp-raid-controller-performance-2013-ww-en

Many thanks for the answers.

14.04.2015, 11:04
( 6 )
Hi Guillermo,

Your questions have been forwarded; we hope to have a reply for you soon.

With best regards

14.04.2015, 13:48
( 6 )
Hi Guillermo,

Here are the answers as provided by Fujitsu Support:

1. K40 is offered as server-only so far. So I can’t comment Server K40 costing.

2. K6000 was the CELSIUS offer instead (just ignore the connectors) of K40.

3. K6000 was always lower cost than K40, but recently Nvidia reduced significant cost of K40.

4. This was the reason workstation PM decided to introduce K40 as well, introduction in May 2015 (with special Fujitsu developed FAN module attached for R940).

5. K80 is already released in CELSIUS R940power (with special Fujitsu developed FAN module attached).

6. K80 to come as well in CELSIUS C740.

With best regards

16.04.2015, 11:46  /  Latest edited: 16.04.2015, 11:51
( 11 )
many thanks for the info, now everything it's a little bit more clear although I still have got some question left.

1. Maybe someone from Server DEpartment should review K20X and K40 prices and perhaps add also K80 to their configurators.

2. I guess from the answer that Fujitsu doesn't recommend K80 for any other CELSIUS than R940power or C740. From PSU point of view M740power could also handle at least one K80 cause it mounts a 1000W PSU and C740 just a 800W PSU (C740 has only room for one K80). In fact M740power maybe could also handle a couple of cards cause it has got the PCIe Gen3 x16 connectors, room enough and with just a couple of HD should work fine talking about power cosumption. That's true? System Architect doesn't allow that configuration.

3. Can't understand why System Architect even doesn't allow to configure a couple of K80 on a R940pwer. It allows a couple of M/K6000 but not a couple of Tesla. If we want to install a couple of K80 which solutions should we use? Primergy, Celsius,...?

4. Really K6000 has the same properties than Tesla K80Confused It's clear that both cards install the same processor and memory type but, for exemple, on Quadro family specifications there's no mention on peak double or single precision floating point performance. Could both card do exactly the same functions? I believe that Quadro should be optimized for graphical performance and Tesla just for pure mathematical use. Prices on NVIDIA distribuitors are not so far away between M6000 and Tesla K80. Perhaps a 400-500€ difference could be understand if performance it's also increased on a simulation scenario.

Again, many thanks for your help.

Page 1/1 1

Please login

Please log in with your Fujitsu Partner Account.


» Forgot password

Register now

If you do not have a Fujitsu Partner Account, please register for a new account.

» Register now