Overclocked NVIDIA TITAN V Benchmarks Emerge

Megalith

Megalith

24-bit/48kHz
Staff member
Joined
Aug 20, 2006
Messages
13,000
A redditor has assembled a collection of links to some early, unofficial benchmarks of the TITAN V, and they suggest that it’s a pretty impressive card. Owners should be able to overclock its memory out of the box, as the results appear to have a 110 to 130 MHz higher clock than the listed frequency.
 
I'm looking forward to seeing real world benchies. Synthetic benchmarks really don't do it for me, though I know they have their purpose.
 
  • Like
Reactions: DF-1
like this
http://hwbot.org/submission/3730827_h2o_vs_ln2_gpupi___1b_2x_titan_xp_2sec_189ms

Screen_Shot_2017-12-09_at_22.29.09.png


clearly the card as i expected is aimed at compute performance more than gaming, but man that gpupi speed is insane (its #6 above, hwbot doesn't have Titan V yet in db so the guy subbed them as Xp's)
 
Gideon said:
Well some people were quite wrong about 30% to 50% better performance then a 1080ti. Granted it's not a gaming card but that is kind of underwhelming as well.
Click to expand...
Its CPU limited.
 
Gideon said:
Well some people were quite wrong about 30% to 50% better performance then a 1080ti. Granted it's not a gaming card but that is kind of underwhelming as well.
Click to expand...

Don't see how they made those predictions. We've been seeing some pretty high correlation between gflop and real performance, between generations for the last few generations, if we're only looking nvidia vs nvidia or AMD vs AMD.

The fact is, the gflop rating of Volta's cuda cores have the same performance as Pascal's when adjusted for core count and speed.

This means Volta Cuda Cores are effectively = Pascal Cores.

1080Ti/Titan X/XP are basically sitting at about 12tflops, while Volta is rated at 15tflops, or about a 25% raw boost in performance.

Sure there are some differences, mainly the Tensor cores and mem bandwidth. But Tensor cores aren't for game, and memory bandwidth is typically a throughput limiter.
 
Once they pare it down for gaming and get rid of the compute focus, im sure itll do well. Already clocking well for such a monster die. im guessing 1080Ti performance for 1080 level wattage. Im an AMD guy but I sure do love that Nvidia is pushing the performance envelope. We have seen though, moving into the future its gonna be multichip packages, AMD has got a leg up with its infinity fabric and HBM experience. Love competition, Nvidia is killin it lately, I hope AMD's Navi launch bring core clocks up and HBM2 can scale its clocks to feed these beasts.
 
I've seen people calling it a consumer card in various places here and there. For $3,000, as a consumer card, it better come with one of Huang's kids.
 
psyclist said:
Once they pare it down for gaming and get rid of the compute focus, im sure itll do well. Already clocking well for such a monster die. im guessing 1080Ti performance for 1080 level wattage. Im an AMD guy but I sure do love that Nvidia is pushing the performance envelope. We have seen though, moving into the future its gonna be multichip packages, AMD has got a leg up with its infinity fabric and HBM experience. Love competition, Nvidia is killin it lately, I hope AMD's Navi launch bring core clocks up and HBM2 can scale its clocks to feed these beasts.
Click to expand...

I don't think there's anything here that AMD can do and Nvidia cannot. And honestly, I'd rate Nvidia's HBM experience above AMD's: Nvidia has used HBM for two of their Tesla HPC-focused generations now, successfully. External interconnects are not hard.

iamjanco said:
I've seen people calling it a consumer card in various places here and there. For $3,000, as a consumer card, it better come with one of Huang's kids.
Click to expand...

It's not a Quadro or Tesla, ergo (super) high-end 'consumer'. It's also significantly cheaper than the compute-only product that it is based on.
 
psyclist said:
moving into the future its gonna be multichip packages
Click to expand...

I fall for that idea for over a decade, without much success. gaming GPUs and multichip never worked well. i quite like NVIDIA's solution: just make the biggest die a silicon node allows.

SLi is dead, long live 815mm² chips
 
geok1ng said:
I fall for that idea for over a decade, without much success. gaming GPUs and multichip never worked well. i quite like NVIDIA's solution: just make the biggest die a silicon node allows.

SLi is dead, long live 815mm² chips
Click to expand...

For the last time: SLI IS NOT MULTI CHIP MULTI CHIP IS NOT SLI

SLI is two cards with two schedulers with two memories, each with the mostly repeated instructions and information.

Multi chip is ONE scheduler accessing available resources (chips) with a cache on each, and ONE shared memory bus.

NVIDIA claims it can't be done. And I understand their latency concern issues. But I have done a little math using Cyclomatic complexity of the pipeline and it might be possible even with complex shaders. Each chip represents a node on the graph. Interchip communication (Edges on the call graph) can be minimal as the pipeline tends to be strongly connected processing wise, eliminating the number of latency data sync calls. The trick is data culling and early rejection of triangles.
 
Last edited by a moderator:
DigitalGriffin said:
it might be possible even with complex shaders
Click to expand...

R&D budgets of 2 billions USD are not unheard of. As NVIDIA has managed to capture more than 80% of the GPU market using die sizes that are on average much larger than AMD's, i would say that their budget was not misspent.I would like to see AMD succeed on GPU as they have on CPU market, but i just don't see it coming. Intel was delivering 5% better performance every 2 years. NVIDIA is delivering 20% better performance every 18 months.
Massive dies FTW.
 
Squall_Rinoa89 said:
the performance to price for the 1080 Ti vs Titan V is like dismal. Suuure its not a "Gaming" GPU, but really... for almost twice the Cuda Cores, its performance is TERRIBLE.

...pass.
Click to expand...
Compute performance is 75-100% higher than Titan Xp.
upload_2017-12-11_8-41-53.png


For CUDA, subtract 15% from the Tesla result to get the Titan result.
upload_2017-12-11_8-42-47.png


This is a return to a prosumer card. You're getting 80% of a Tesla V100 for 30% the cost.
 
Armenius said:
Compute performance is 75-100% higher than Titan Xp.
View attachment 46015

For CUDA, subtract 15% from the Tesla result to get the Titan result.
View attachment 46016

This is a return to a prosumer card. You're getting 80% of a Tesla V100 for 30% the cost.
Click to expand...

For the vast majority of users though it's wasted transistors as we will never really gain from those tensor units while gaming.

I could use it for my ai projects however
 
You must log in or register to reply here.
Back
Top