Titan V announced, $3000 volta gpu

[casiofx]

5120 cores, 640 tensor cores, 12GB HBM2, 110teraflops deep learning, 15teraflops FP32, and designed to burn your wallet to the ash.

https://www.nvidia.com/en-us/titan/titan-v/

 

[krumme]

Great i will take the midrange version at 1500 euro then. Sorry about the price lol but its an indication this rocks.

 

[Sable]

tbh, fair play to them. There's no competition for the 1080ti but they know people always want something faster so there's bound to be early adopters for this who'll pay crazy money. So they can rake in the money for this as it's ready for release and just hold off releasing anything else till they feel like it as the 1080Ti is still bloody pricey and still selling and also still the best "none-halo" card.

 

[krumme]

Its amd adrennalin drivers comming next week. Thats the reason for this announcement.
I think that means more reasonably consumer volta already january or february. But i guess its about gddr6.
Great news anyway all over.
I hope for a 1080 replacement and compettitor for upcomming small vega.

 

[casiofx]

I suspect the next geforce would not use gv100 but the one nvidia announced, ampere, probably without tensor cores.

3000 is a bargain for reseachers to use the tensor cores, instead of shelling out 15k usd plus for tesla v100 like that.

While for us this thing is not really a consumer product anymore.

 

[krumme]

Yeaa its professional
https://www.anandtech.com/show/1213...dia-titan-v-video-card-gv100-for-3000-dollars

Single precision is not that improved at 13.8

Build on 12nm and still 815mm2...!

 

[JoeRambo]

Single precision is not that improved at 13.8

Those looking for SP can already buy much cheaper cards. Eye watering numbers in DP and tensor ops tho. Too bad memory capacity and bw was cut down, otherwise this thing would have some serious applications.

 

[Campy]

I don't think this will be very good for gaming, but we'll see. Maybe I'll be pleasantly surprised.

Good to see the new architecture make it's way closer to consumer level.

 

[Yotsugi]

Its amd adrennalin drivers comming next week. Thats the reason for this announcement.

That's silly and you know it.

Those looking for SP can already buy much cheaper cards. Eye watering numbers in DP and tensor ops tho. Too bad memory capacity and bw was cut down, otherwise this thing would have some serious applications.

It still has some decent applications.
It's basically a non-headless cutdown V100.
Perfect for personal ML/DL workstation.

 

[casiofx]

The titan xp originally had FP64 to be 1/32 of FP32. But the driver update boosted titan xp’s performance in professional applications. Is it still 1/32?

 

[Yotsugi]

The titan xp originally had FP64 to be 1/32 of FP32. But the driver update boosted titan xp’s performance in professional applications. Is it still 1/32?

Yes.
Driver update won't make FP64 units to magically appear on-die.

 

[NTMBK]

Its amd adrennalin drivers comming next week. Thats the reason for this announcement.

This has literally nothing to do with gaming. It's a deep learning monster, but barely an improvement over a 1080ti for FP32. They want to sell this to deep learning scientists.

 

[Yotsugi]

They want to sell this to deep learning scientists.

DIY DL/ML workstations are still pretty niche.

 

[NTMBK]

DIY DL/ML workstations are still pretty niche.

If you've already got a perfectly good Dell/HP workstation, you can buy this and drop it in, let you train DNNs 10x faster.

 

[imported_bman]

Yeaa its professional
https://www.anandtech.com/show/1213...dia-titan-v-video-card-gv100-for-3000-dollars

Single precision is not that improved at 13.8

Build on 12nm and still 815mm2...!

Those TPUs and extra L2 cache probably eat up a lot of die space.

Waterblocked and with an aggressive overclock this card might be able to break 20TFlops. I suspect ~305W will be the power limit of the card and this is likely an already power hungry GPU so overclocking might be limited. Hopefully the move to HBM creates enough headroom that it allows a 2.11GHz OC, the OC required for 20TFlops. Though at least it should be able to provide twice the performance of a 1080 with a moderate-high overclock, that pretty much guarantees 4k@60 on ultra settings on all games (mins will still dip to the 40s on some) assuming a Volta TFlop => Pascal TFlop for games.

 

[Yotsugi]

Those TPUs and extra L2 cache probably eat up a lot of die space.

No and no.
TPUs are small-ish FP16 ALUs. The wiring probably costs them more than the actual die area.
But FP64 eats a LOT of die area.

Waterblocked and with an aggressive overclock this card might be able to break 20TFlops.

Not on this chip, it will enter immolation stage very soon.
GV102?
Possibly.

 

[NTMBK]

Those TPUs and extra L2 cache probably eat up a lot of die space.

As do all those FP64 and FP16 cores. This isn't a successor to the Titan Xp, it's a successor to the Tesla P100/Quadro P100- if you want to compare die area, compare those two.

 

[imported_bman]

No and no.
TPUs are small-ish FP16 ALUs. The wiring probably costs them more than the actual die area.
But FP64 eats a LOT of die area.

Not on this chip, it will enter immolation stage very soon.
GV102?
Possibly.

Yeah FP64 is probably the biggest culprit when it comes to die space.

I don't think temps will be the killer for overclocking this card, the power limit will. Titan XP on water can hit above 2GHz while staying around 45C.

 

[Yotsugi]

Titan XP on water can hit above 2GHz while staying around 45C.

GP102 is smaller, significantly so.
Did anyone try overclocking Quadro P100?

 

[imported_bman]

GP102 is smaller, significantly so.
Did anyone try overclocking Quadro P100?

I have not seen anyone post P100 benches for gaming let alone hack a waterblock onto one, there are probably some P6000 that have been waterblocked. Most of that large die will not be utilized while gaming and those temps for the Titan XP @ 2GHz can be achieved even on a hybrid kit using a 120mm AIO radiator, I can't see a 240mm or 320mm radiator having any trouble with temps for the V100.

 

[Flapdrol1337]

Should beat a titan x by 30% or so in gaming right? Kind of surprised they're launching a consumer version of this, even if it is at a ridiculous price.

 

[dark zero]

It is supposed to crush Titan X by 50%.... If not is Titan Z all over again, kills HBM2 and 12 nm process at the same time...

The clocks are pretty lower to even Vega 64...

 

[Qwertilot]

It crushes the Titan X utterly out of sight in terms of what it is really trying for (Deep learning compute in particular.).

 

[NTMBK]

It is supposed to crush Titan X by 50%.... If not is Titan Z all over again, kills HBM2 and 12 nm process at the same time...

The clocks are pretty lower to even Vega 64...

The Titan X was an FP32 focused card, this isn't. It's all about the deep learning tensor cores, and the FP64 cores.

 

[krumme]

That's silly and you know it.

It still has some decent applications.
It's basically a non-headless cutdown V100.
Perfect for personal ML/DL workstation.

Comming as lightning without thunder?

Ok lets wait and see reviews of this and the new adrennaline drivers. This card could have been launched one month before or one month later. The timing is made to keep mindshare. Secure total performance control. Secure brand. Until next gen sp16 arives.
Nobody will care if a rx64 is as fast as a 1080ti in february or if this is 815mm2 or 3000 usd.