Nvidia Infuses DGX-1 With Volta, Eight V100s In A Single Chassis

[Guest]

Nvidia's new Tesla V100-powered DGX-1 packs eight Volta GPUs crammed into a 3U chassis to deliver a whopping 960 TFLOPs of power from 40,960 CUDA cores.

Nvidia Infuses DGX-1 With Volta, Eight V100s In A Single Chassis : Read more

 

[Janissaire]

high end gaming for rich peoples lolz,you can tell it by looking at the chassis online&the premium look of the graphic card.

 

[bit_user]

NVLink 2.0 increases throughput to 10X that of a standard PCIe connection (300GB/s).

Comparing aggregate vs. point-to-point bandwidth.
: (

A single NVLink2 is really just ~56% faster than PCIe 3.0 x16.

 

[drajitsh]

Paul please check. You alluded that it CAN be used a gpu, "Yes, the DGX-1 can play Crysis " . Tesla cards do NOT have display out hardware. You do NOT get DVI/DP/HDMI ports in tesla cards.

 

[coolitic]

I believe a few tesla cards had display output in the past at least, for whatever reason. But yeah, most don't.

 

[Fiqar_]

I'm so sick of everyone following apple in terms of design. Be orginal for christs sake! This looks like something straight out of an Apple launch event! I mean the premium look alone gives it away!

 

[bit_user]

I'm so sick of everyone following apple in terms of design. Be orginal for christs sake! This looks like something straight out of an Apple launch event! I mean the premium look alone gives it away!

Wow, of all the things to pick on...

Anyway, for those of us who don't watch Apple launch events, what are you claiming resembles them? I have no idea what you mean by "premium look".

 

[Karadjgne]

Theres no issue with not having DP out on that monstrosity, it'll never be used in a pc, not with a 3200w necessity power consumption. The only way anyone will use it with games is via network.

 

[bit_user]

Theres no issue with not having DP out on that monstrosity, it'll never be used in a pc, not with a 3200w necessity power consumption. The only way anyone will use it with games is via network.

There's allegedly a Quadro GP100 graphics card. So, it's plausible there could someday be a Quadro GV100.

If this is to be believed, it looks like the drivers still need work:

http://www.videocardbenchmark.net/gpu.php?gpu=Quadro+GP100&id=3721

 

[Karadjgne]

Dunno, I've always wondered about that, can't be just the drivers. Look at the Quadro's out now, like the $800 m4000, it's got 1664 Cc, 192 mem interface and 8Gb vram and DP 1.2 vrs the 1080ti that has double the Cc, 352 mem interface, 11Gb vram and DP 1.4 yet is $100 cheaper. And yet the gtx can't do what that m4000 can, so I'm guessing there's considerably more to what else is on those cards than just the physical specs, cuz in gaming the m4000 is gonna get Stomped by the gtx, yet in Cad, the gtx is gonna get Stomped and it can't all be just the drivers. The Quadro's / FirePros were never really good at gaming.

 

[Elrabin]

Dunno, I've always wondered about that, can't be just the drivers. Look at the Quadro's out now, like the $800 m4000, it's got 1664 Cc, 192 mem interface and 8Gb vram and DP 1.2 vrs the 1080ti that has double the Cc, 352 mem interface, 11Gb vram and DP 1.4 yet is $100 cheaper. And yet the gtx can't do what that m4000 can, so I'm guessing there's considerably more to what else is on those cards than just the physical specs, cuz in gaming the m4000 is gonna get Stomped by the gtx, yet in Cad, the gtx is gonna get Stomped and it can't all be just the drivers. The Quadro's / FirePros were never really good at gaming.

Why are you comparing an older model to a newer model?

M4000 is based on Maxwell. A more fair comparison would be to the P5000 which is based on Pascal, same as the 1080

GTX 1080 has 2560 CUDA cores at 1733 mhz and 8gb of GDDR5x

The P5000 has 2560 CUDA cores at an unspecified clock speed and 16gb of GDDR5x

They should perform almost identically in gaming and workstation tasks once clock speed is taken into account.

I'm using a GTX 1080 in Nvidia IRAY 3d renders and it stomp the snot out of a M4000 for that task.

http://www.migenius.com/products/nvidia-iray/iray-benchmarks-2016-3

 

[Karadjgne]

I was comparing the m4000 and gtx1080ti because they were the closest I could see in price range of @$750.

The gtx1080 is $500
The Quadro p5000 is $2000.
And you are pretty much stuck for PNY? Are ppl seriously paying that much cash for the extra 8Gb of ram and a different set of drivers?

That was my comparison, and now there's an 8gpu combination tesla card that's $150k and needs a 3.2kw power source?

What gives? Why do the Quadro's cost so much more, for so much less and yet in that benchmark, the GP100 bombed the scores at $7000 as compared to a $500 gtx1070.

What could possibly be the reasoning for the DGX-1, a server based piece of equipment, shared by multiple ppl at the same time, if 8 individual pc's with 1080ti's will get better performance at a fraction of the cost per individual station. You could outfit over 210 ppl with 1080ti in individual workstations for the price of that one Tesla card or at least 100ppl with sli. Or is this all about saving time using gpu based rendering vrs cpu rendering

 

[Elrabin]

I was comparing the m4000 and gtx1080ti because they were the closest I could see in price range of @$750.

The gtx1080 is $500
The Quadro p5000 is $2000.
And you are pretty much stuck for PNY? Are ppl seriously paying that much cash for the extra 8Gb of ram and a different set of drivers?

That was my comparison, and now there's an 8gpu combination tesla card that's $150k and needs a 3.2kw power source?

What gives? Why do the Quadro's cost so much more, for so much less and yet in that benchmark, the GP100 bombed the scores at $7000 as compared to a $500 gtx1070.

What could possibly be the reasoning for the DGX-1, a server based piece of equipment, shared by multiple ppl at the same time, if 8 individual pc's with 1080ti's will get better performance at a fraction of the cost per individual station. You could outfit over 210 ppl with 1080ti in individual workstations for the price of that one Tesla card or at least 100ppl with sli. Or is this all about saving time using gpu based rendering vrs cpu rendering

Where to begin......

Quadro and Tesla and Grid cards are designed to go into servers which have different airflow requirements, specifically that servers have very high CFM fans that blow front to back and quadro/tesla/grid cards are either fanless with inlet/outlet front/back or have fans that also blow front/back

They also never change the design of a particular card from the beginning of production to the end of production, guaranteeing the design for the server lifecycle which will be 3-5 years. This is important for OEM qualification and testing.

OEMs like HP, Dell, IBM, etc test and certify that professional cards work in their systems and conform to specific dimensions and thermal requirements

Also, the drivers are guaranteed to work with ISV(independent software vendor) certification

For example, if you buy X server, with Y quadro/grid/tesla card running Z driver, it will guarantee to work with vendors products such as 3d studio max, Autocad, Solidworks and more.

Professional workloads also are MUCH larger than consumer ones. An autocad model might consume 16gb of VRAM or more

A random GTX card will not have any of the above

A company will pay more for the above because it reduces their risk and down time.

Big companies that rely on their HPC or computing clusters or VDI farm for productivity and revenue will not fuck around with spending 4-8x less per card because if they have issues with a GTX card, be it hardware or software, they have no path to resolve.

Consumer graphic card manufacturers will send out RMA's in weeks, not hours and they have no software troubleshooting for ISVs

Quadro/Tesla/Grid you will get a replacement card from HP/Dell/IBM/etc in as little as four hours depending on your warranty and you have full software support. The server vendor will open a ticket with both the ISV and Nvidia to resolve

I work in Enterprise IT, if you have to ask "Why are you spending so much?" You're not the target audience

 

[Karadjgne]

Finally, an answer in English not Techy. Thank you.
So basically, as I understand it, something along the lines of the DGX-1 is warranted because of the sheer size and scope of the workload in like a CGI movie vrs the speed inherent in working on just a few particular frames individually, the sum of the whole being more than just the sum of the individual parts. And time is money.

 

[Charles Alden]

The DGX workstation is just like any other computer, just with very, very powerful hardware. It can run multiple 4K, even 8k displays and yes, it can play Crysis for Linux.

 

[bit_user]

The DGX workstation is just like any other computer, just with very, very powerful hardware. It can run multiple 4K, even 8k displays and yes, it can play Crysis for Linux.

The spec sheet says it has 3x DisplayPort. Doesn't say what revision, but up to 4k.

https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/dgx-station/dgx-station-data-science-supercomputer-datasheet-v4.pdf

So, the V100's are driving the DisplayPorts? Interesting. I figured they'd be compute modules like those in the DGX-1, but they look like they could be PCIe cards, perhaps with NVLink going over the top.

If you know anything more, do share. Not like I've got $69k to spare on one...