One of the reasons I liked the Supermicro case is that it can operate as a tower or can be rack-mounted, so if the noise was an issue, I could migrate to a machine room. There is another issue with desk-side use in the US: with 120V outlets, the PSU only supply 1100W. I found one can buy an external 1500W step-up transformer for about $70, but this would likely add to the heat/noise issue.
By coincidence, just got word of Renderstream’s updated web page. They sell a variety of customizable GPU workstations, including ones based on the Supermicro MB / case combo you’re asking about.
But they also have a new option for a new EIGHT-WAY GPU box. Now that, my friends, would be a fun toy.
I’ve always just built my own machines, but this is obviously something special.
Looks like the SuperMicro it is.
But before I take the plunge for eternity and forever hold my peace, does anyone know of a socket C32 or G34 Opteron motherboard which can take 4 Teslas?
Final attempt to buy AMD.
Looks like this might be based on the Colfax CXT8000 4U Rackmount Server (Colfax GPU servers). For my application, I think this would be quite nice. I have a large, read-only dataset I’d like to distribute between GPUs, but I think this would be too slow over Infiniband. Having 8 GPUs (with 24-48 GB DDR5, in the fall) on the same node would be very useful.
Not yet, but the platform was only released a month ago. Most of the first generation boards from the usual suspects (Tyan, Supermicro) are really server boards, most only have a single SR5690 north bridge and at most 2 x 16 lane PCI-e slots.
The VDACTr8 and VDACTr8x2 and the other eight GPU systems share a common lineage. Our systems differ in that they are customized with the user and application in mind and then built in an ISO9001 facility, with CUDA loaded, initialized, tested for compatibility and then burned in at 94F to 102F for twenty-four hours where if there are any failures the systems are fixed and re-burned in and ships when it finally makes the grade. For system builders we also offer components and kits.
For database we offer up to 144GB of memory that we can configure between main memory and RAM-disk. Finally we offer full world-wide on-site service warranty plans except for South Africa, which we hope to remedy soon.
We have four-years of GPGPU experience where we continually strive to make a better product. One of the things we are looking for is a better burn-in test one that fully loads the GPUs and logs the performance. We would love to hear the forums input.
Thanks
John
On my machine, only n-body sdk sample increase gpu temperature much and makes fan be noisy. I run all samples.
I apologize for my misinformation. I am not affiliated with Colfax or any other vendor, and did not realize who is the OEM for these 8-PCI servers. I have no reason to endorse one vendor’s product over another, so I apologize if I appeared to be doing so. Thanks for the info!
Thanks for the apology but no offense was taken. It was an FYI only. we are just excited to be involved with the community.
Thank you
john
Can anyone confirm a supermicro server mobo with 7 PCIE slots. I saw a review in the past but dont know if its ever made it to market.
The Supermicro X8DTH-6F has 7 (x8) PCI-E 2.0 ( in x16 slot).
Thanks much.
Anyone know of any AM3 790FX or 890FX motherboards with 4 Tesla support.
I see that the MSI 790FX-GD70 has double width support and 4 x16 slots, but a) no on-board display so one slot needs to go to that and B) probably no space for 4 anyway.
Maybe even an AM2+ motherboard with 4 Tesla support. The only one I can see is the Foxconn Destroyer (which I am investigating).
By “Telsa support” I presume you mean 4x16 PCI-e slots + onboard graphics? Then right now I think the answer is no. Three T10 Telsa + a GT200 or Fermi GPU is about as good as you will do (it is the same with all the Intel X58 consumer boards too). 890FX hasn’t officially been released yet, but I doubt it will change anything. I know ASUS and MSI both have 4x16 slot 790FX boards that should work.
I wouldn’t recommend any of the older AM2+ Nvidia 780 chipset boards - they are really old, BIOS updates stopped a long time ago, and the current DDR3 platforms are considerably better in just about every respect.
Of course the new Tesla C2050 is starting to trickle out to customers - those can act as a GPU, so that might be another option to consider.