Nvidia Volta Speculation Thread

A user named "Vega" who owns a TitanXp has obtained a TitanV and did some benchmarks:

3168851


https://forums.overclockers.co.uk/posts/31418225/
 
GamersNexus tested TitanV vs TitanXp:

Titan V is 41% faster in Doom, 27% faster in Sniper Elite 4, 36% faster in HellBlade, and 20% faster in TimeSpy. However it's only 5 to 10% faster in other titles like FireStrike, Ashes Of Singularity, Ghost Recon and For Honor. Since there is a clear disparity and inconsistency here, I guess the safest bet is that NVIDIA didn't really a release a proper driver for the card with all the necessary optimizations, opting instead to release a driver that offer the bare minimum of functionality in games. There is a serious frame pacing issue plaguing several titles already.

FWIW, the card boosts reliably to 1680MHz, overclocking can help make it to 1800MHz stable which adds another ~20% increase in performance. The card also appears to hit a thermal and power wall faster than Pascal. I am guessing lighter versions of Volta will alleviate these issues allowing the GPU to clock higher.

https://www.gamersnexus.net/guides/...marks-async-future-is-bright-for-volta/page-2
 
Last edited:
I did a bunch of crypto benchmarking on my card. I bought it for deep learning tasks but I have a lot of GPUs sitting around, so I do some mining on the side.

Outliers are ethash and lyra2z at a 2.2x speed increase. lyra2RE2 actually got slower. Most everything else is 1.5x faster.

As an aside, this card is really energy efficient. In every case, this card uses less power than the 1080ti while being 50% faster.

Code:
Algorithm  Titan V  1080ti   mult
LBRY       687      460      1.49
SKUNK      57.8     47.5     1.22
BitCore       36.3     23.1     1.57
poly       47.1     31.4     1.5
Groestl    74.9     58       1.29
xevan      crash      
keccak     1.63     1.18     1.38
Equihash   792      685      1.16
x17        25.85    18.3     1.41
x11evo     29.6     17.2     1.72
veltor     78.6     54.4     1.44
phi        44.4     29.4     1.51
skein      1260     842      1.50
sib        29.3     20.8     1.41
lyra2z     6.12     2.77     2.21
myr-gr     144.7    112.9    1.28
lyra2RE2   48.1     66.8     0.72
hsr        28.4     19.5     1.46
timetravel 54.3     39.9     1.36
c11        38.4     27.7     1.39
tribus     130.3    88.4     1.47
blakecoin  9.54     7.6      1.26
blake2s    8.6      6.26     1.37
ethash     67.3     30.5     2.21
                             1.53
 
That's really nuts. And bodes well for gaming volta. Or ampere, or whatever is next gaming generation.

AMD is probably quaking in their boots seeing this...

I would venture a guess that a lot of the power consumption reduction is due to HMB2 which is not going to be the case for consumer products.
 
I would venture a guess that a lot of the power consumption reduction is due to HMB2 which is not going to be the case for consumer products.
I guess the same applies to Vega?
So, cheaper Vega cards with GDDR5/6 would be in even larger perf/W disadvantage compared to Pascal/Volta/Ampere?
 
I would venture a guess that a lot of the power consumption reduction is due to HMB2 which is not going to be the case for consumer products.
But the upper Nvidia consumer cards would probably be using GDDR6 going forward, which has advantages over GDDR5/X.
.
 
Seems the TitanV has only 96 ROPs, same number as TitanXp. The number should be 128 ROPs just like the Tesla V100, but TitanV has 25% of them disabled. This could present a serious bottleneck though. Especially when considering the clock deficit the TitanV has compared to the TitanXP. This (in addition to drivers) might explain some of the inconsistencies we are seeing.

PCPer has the card and has posted a tear down video of it.
Some notes:
[*]This implementation has 25% of the memory and ROPs disabled, giving us 12GB of HBM2, a 3072-bit bus, and 96 ROPs.
[*]Clock speeds in our testing look to be much higher than the base AND boost ratings.
[*]So far, even though the price takes this out of the gaming segment completely, we are impressed with some of the gaming results we have found.
[*]The cooler might LOOK the same, but it definitely is heavier than the cooler and build for the Titan Xp.

https://www.pcper.com/news/Graphics...GPU-look-NVIDIA-TITAN-V-Unboxing-and-Teardown
 
So the earlier reports of boosting to 2Ghz were wrong, the card gains nicely in some games in GN's review with more power. nvidia have similar problem as AMD now where the chips aren't scaled?

Waiting for duderandom84 to get one for a clock for clock comparison.
 
So the earlier reports of boosting to 2Ghz were wrong, the card gains nicely in some games in GN's review with more power. nvidia have similar problem as AMD now where the chips aren't scaled?

Waiting for duderandom84 to get one for a clock for clock comparison.
Well it is not really an efficient dedicated FP32 GPU die but a huge mixed-precision die, and like I mentioned pretty obvious would hit a thermal and power wall when compared to a Titan Xp.
If this was just Volta arch expanding Titan Xp by 42% then the efficiency would be there and probably hit a bit higher with the clocks while also performing better in games with less throtting, I am pretty sure Nvidia has mentioned in the past in discussions at events that doubling the SM per GPC/TPC (P100 and V100) does not really favour gaming workloads.

Logically if they expand cores by 40% then before any arch improvements you have a very high power demand, the node they are on is a slight advantage over the last TSMC 16FF+, and so does the improvements to the Volta arch.
But it is a huge die optimised beyond FP32 cuda cores-gaming due to being more scientific orientated.
 
Last edited:
So the earlier reports of boosting to 2Ghz were wrong
It can be easily overclocked to 1.9GHz though. That's pretty solid given the massive size of to the chip.
nvidia have similar problem as AMD now where the chips aren't scaled?
Vega scaling problems are universal, TitanV have spotty scaling in some cases at best. It usually scales pretty close to its advantages in FP32, despite being hampered by reduced clocks and less ROPs. A lighter version of Volta with higher clocks and more ROPs (sacrificing FP64 and Tensor) can easily take care of the problem. Volta at 1.8GHz gains significant chunk of performance. Also Frame pacing issues suggest a not fully optimized driver as of yet.

German site hardwareluxx.de got themselves a Titan V and ran Luxmark 3 as well as GPU pi:
https://www.hardwareluxx.de/index.p...volta-architektur-im-gaming-test.html?start=6
Though they completely destroyed their gaming benchmarks by using a gen 3 Core i7, thus having their entire high end GPU stack massively CPU limited! Even their 1080Ti is barely 10% faster than their regular 1080!
 
Last edited:
It can be easily overclocked to 1.9GHz though. That's pretty solid given the massive size of to the chip.

Yeah, that's what I implied by the stating that power limit increase allows it to perform much better where it can.

Vega scaling problems are universal, TitanV have spotty scaling in some cases at best. It usually scales pretty close to its advantages in FP32, despite being hampered by reduced clocks and less ROPs. A lighter version of Volta with higher clocks and more ROPs (sacrificing FP64 and Tensor) can easily take care of the problem. Volta at 1.8GHz gains significant chunk of performance. Also Frame pacing issues suggest a not fully optimized driver as of yet.

I wasn't thinking of Vega scaling(compared to what?), more like Fury vs 390X. I'm presuming this has same number of ROPs as previous Titan and that'll affect scaling unless something else in Volta's architecture prevents it. 1080 doesn't do 2x of 1060 in some part due to 1060 having 48ROPs instead of 32. And the problem being that nvidia face in terms of drivers, optimizing this thing for games. Maybe they will be going back to similar SM for their gaming cards as Pascal when they're out later and this thing is forgotten.
 
PugetSystems has a quick and dirty test up for now:
https://www.pugetsystems.com/blog/2017/12/12/A-quick-look-at-Titan-V-rendering-performance-1083/

Interesting results for V-Ray and FurryBall RT' 60% to 65% faster with the Titan V.
I would assume they kept it within spec as would be used with workstations, which makes the results even more promising as it has slightly lower Boost clocks than Titan xP, just that one can push the Titan xP closer to the ragged edge but not sure that is suitable for most users anyway longterm without changing the cooling solution for lower noise.
 
Last edited:
Back
Top