NVIDIA Announces World's Most Powerful Professional GPU - The Quadro GV100

[Raevenlord]

NVIDIA CEO Jensen Huang today announced to the world what is being touted as the world's most advanced GPU for professionals - the Quadro GV100. Based on NVIDIA's Volta architecture and sporting 5,120 CUDA cores, this is a behemoth of a professional graphics card, eclipsing all previous NVIDIA offerings by a wide margin.

The GV100 was built to usher in all kinds of workload performances - from machine learning and HPC to actually being the only professional graphics card actually capable of processing workloads around the just announced RTX raytracing suite. That alone is a boon that's sure to get these cards moving from the presses and unto professional's hands all the faster. The 5,120 CUDA cores join the 32 GB of HBM2 memory and up to 7.4 TFLOPs of power for double-precision rendering, 14.8 TFLOPs for single-precision workloads, 29.6 TFLOPs half-precision, and 118.5 TFLOPs for deep learning through its Tensor Cores. The previous high-end Quadro, the GP100, offered 10.3 TFLOPs for single-precision rendering.





This graphics card is a behemoth in all senses. The 5,120 CUDA cores, 640 Tensor cores, 320 texture units and 128 ROPs of hardware on this GPU run at some 1450 MHz boost clocks. The HBM 2 memory is a 1.7 Gbps affair on a 4096-bit bus, which naturally features ECC (error correction for data-critical workloads). The TDP stands at a (relatively) innocuous 250 W - for this big a chip, it's actually rather tame, and could make it (mostly a drop-in replacement for the Quadro P100, which sports TDP a tad lower at 230 W.



The Quadro GV100 has dropped the GP100's sole DVI port, leaving it with 4x DisplayPort 1.4 connectors. The card also features standard Quadro Sync and Stereo connectors for synchronized refresh and quad-buffered stereo respectively. Naturally, this Quadro GV100 features support for dual NVIDIA NVLink operation, which will allow prospective users to enjoy the full benefits of the interface, particularly low latency data transfers, remote memory access, and memory pooling.

View at TechPowerUp Main Site

 

[T4C Fantasy]

Boost Clock Speed: 1447 vs 1455 xD

Quadro GV100
Boost Clock: 14.817TFLOPs + 118.538TFLOPs = 133.335TFLOPs

TITAN V
Boost Clock: 14.899TFLOPs + 119.193TFLOPs = 134.092TFLOPs

 

[dj-electric]

With 32gb of memory, this thing is an absolute unparalleled monstrosity

People who train machines know how much

 

[Xzibit]

Its a good thing it has ECC

The Register: 2 + 2 = 4, er, 4.1, no, 4.3... Nvidia's Titan V GPUs spit out 'wrong answers' in scientific simulations

 

[TheinsanegamerN]

And no news on Ampere or new gaming cards. Another year of the same GPUs!

 

[T4C Fantasy]

And no news on Ampere or new gaming cards. Another year of the same GPUs!

i doubt Ampere is real, just Volta and Turing

 

[TheoneandonlyMrK]

Its a good thing it has ECC

The Register: 2 + 2 = 4, er, 4.1, no, 4.3... Nvidia's Titan V GPUs spit out 'wrong answers' in scientific simulations

Kinda nice how they have a replacement for such scientists now though eh.
Nice card though.

 

[the54thvoid]

Its a good thing it has ECC

The Register: 2 + 2 = 4, er, 4.1, no, 4.3... Nvidia's Titan V GPUs spit out 'wrong answers' in scientific simulations

Yeah, the real scientists use the Quadro cards. If you want to run serious and precise mathematical models that require ECC, don't skimp on a $3000 Geforce card that doesn't have it.

Not really sure why, tbh, the Titan V even exists apart from making some quick bucks from eager early adopters.

 

[Xzibit]

Not really sure why, tbh, the Titan V even exists apart from making some quick bucks from eager early adopters.

Yeah, especially since they changed the EULA

TPU: NVIDIA Forbids GeForce Driver Deployment in Data Centers

Be funny if Titan line evolves to be the mining card

 

[Vya Domus]

Yeah, the real scientists use the Quadro cards. If you want to run serious and precise mathematical models that require ECC, don't skimp on a $3000 Geforce card that doesn't have it.

It does rise some questions though , it's the same silicon , there is a chance that whatever flaw was in the Titan V is still present.

 

[R-T-B]

It does rise some questions though , it's the same silicon , there is a chance that whatever flaw was in the Titan V is still present.

It's probably memory related, which ECC would correct.

There is always a chance, I suppose.

 

[the54thvoid]

It does rise some questions though , it's the same silicon , there is a chance that whatever flaw was in the Titan V is still present.

I did a bit of light reading - ECC is an added redundancy which duplicates a code for a given calculation. It's not so much an error on Volta as the simple fact, ECC is not avaialble on that line. Titan V does not have ECC because it's cheaper than a Volta Quadro. It's like saying my phone camera is faulty because it doesn't have OIS, when in reality, OIS is an added algorithm to create a better image. Nothing is broken, simply not included.

From Nvidia:

“All of our GPUs add correctly. Our Tesla line, which has ECC [error-correcting code memory], is designed for these types of large scale, high performance simulations. Anyone who does experience issues should contact support@nvidia.com.”

 

[Xzibit]

From Nvidia:

“All of our GPUs add correctly. Our Tesla line, which has ECC [error-correcting code memory], is designed for these types of large scale, high performance simulations. Anyone who does experience issues should contact support@nvidia.com.”

Dont think its an ECC issue but that where Nvidia is directing the solution. Which lines up with their EULA move.

After repeated tests on four of the top-of-the-line GPUs, he found two gave numerical errors about 10 per cent of the time. These tests should produce the same output values each time again and again.

We have repeatedly asked Nvidia for an explanation, and spokespeople have declined to comment.

 

[Sasqui]

Yeah, especially since they changed the EULA

TPU: NVIDIA Forbids GeForce Driver Deployment in Data Centers

Be funny if Titan line evolves to be the mining card

And here I thought they were going to forbid users from playing Far Cry 5 with it

 

[Vya Domus]

I did a bit of light reading - ECC is an added redundancy which duplicates a code for a given calculation. It's not so much an error on Volta as the simple fact, ECC is not avaialble on that line. Titan V does not have ECC because it's cheaper than a Volta Quadro.

I wasn't talking about ECC , that's only relevant to the memory component. There is the floating point arithmetic inside the silicon that can crate generate errors from the get-go for example.

 

[Ferrum Master]

I did a bit of light reading - ECC is an added redundancy which duplicates a code for a given calculation. It's not so much an error on Volta as the simple fact, ECC is not avaialble on that line. Titan V does not have ECC because it's cheaper than a Volta Quadro. It's like saying my phone camera is faulty because it doesn't have OIS, when in reality, OIS is an added algorithm to create a better image. Nothing is broken, simply not included.

From Nvidia:

“All of our GPUs add correctly. Our Tesla line, which has ECC [error-correcting code memory], is designed for these types of large scale, high performance simulations. Anyone who does experience issues should contact support@nvidia.com.”

EIS not OIS. OIS requires hardware just as ECC does thou... a bad analogy.

 

[ensabrenoir]

...I'll be looking for a review of this on Linus Tech Tips caus he'll probably get four to go along with his two 28 core zeons on that insane overclockable dual socket Asus board (with custom e.k waterblocks) so he can try to play minecraft on it or something equally ridiculous.

 

[T4C Fantasy]

Tpu needs a tech show
Start hiring xD @W1zzard

 

[DeathtoGnomes]

mmm Professional.... ya

 

[the54thvoid]

So, we wait for an article that says quadro is also producing errors before we confirm it's not ECC related. All mathematical calculations are prone to errors on hardware so ECC is required for scientific rigor.
Everyone is jumping to a hardware fault conclusion when as usual, none of us know. From what I read, it's also a very specific calculation and one could argue, ECC, available on proper (not gimmicky Titan brand) hardware would render this error void.

From the Register itself:

A spokesperson for Nvidia has been in touch to say people should drop the chip designer a note if they have any problems. The biz acknowledged it is aware of at least one scientific application – a molecular dynamics package called Amber – that reportedly is affected by the Titan V weirdness

Storm, teacup and Nvidia will fix.

 

[bug]

It does rise some questions though , it's the same silicon , there is a chance that whatever flaw was in the Titan V is still present.

There's not the slightest amount of evidence so far that the issue is in the silicon. It could be in the firmware, driver of the software used.

What I'm interested in, is how this handles RTX. But I guess it will be a while before someone manages to get a hold of both the hardware and the software needed to benchmark that.