Nvidia announces $2,999 Titan V, 'the most powerful PC GPU ever created'

It seems like Nvidia announces the fastest GPU in history multiple times a year, and that’s exactly what’s happened again today; the Titan V is “the most powerful PC GPU ever created,” in Nvidia's words. It represents a more significant leap than most products that have made that claim, however, as it's the first consumer-grade GPU based around Nvidia’s new Volta architecture.

That said, a liberal definition of the word “consumer” is in order here — the Titan V sells for $2,999 and is focused around AI and scientific simulation processing. Nvidia claims up to 110 teraflops of performance from its 21.1 billion transistors, with 12GB of HBM2 memory, 5120 CUDA cores, and 640 “tensor cores” that are said to offer up to 9 times the deep-learning performance of its predecessor.

Also it comes in gold and black, which looks pretty cool.

“Our vision for Volta was to push the outer limits of high performance computing and AI. We broke new ground with its new processor architecture, instructions, numerical formats, memory architecture and processor links,” Nvidia founder and CEO Jensen Huang says in a statement. “With Titan V, we are putting Volta into the hands of researchers and scientists all over the world. I can’t wait to see their breakthrough discoveries.”

There’s no word on when Volta might make its way into more attainable gaming-focused GPUs. Nvidia’s current Pascal architecture was introduced over a year and a half ago with the GTX 1080, which still remains one of the best gaming GPUs available — only really topped in performance by its turbocharged successor, the GTX 1080 Ti. Volta GPUs will be a lot more expensive for Nvidia to produce, so the company may want to ride out the Pascal wave as long as it can.

The Titan V is available today and is limited to two per customer.


110 TFLOPs? My 1080 does… 9. I feel so inadequate. But not $3k inadequate.

I’ll be interested to see how many pixels this thing can push. Silky smooth Retina-level VR graphics in a few years would be super neat.

A GTX 1080Ti is more than enough to push out silky smooth, retina level VR… You can even have a dual 1080Ti in SLI for about $1500 and run a 8K display with pretty good frame rates. 110TFLOPs of the Titan V is mainly aimed at Large Scale Deep Learning applications, which is why Nvidia also included 640 Tensor Cores. I don’t expect any enthusiast to buy this for gaming or productivity reasons though.

This is not entirely true. There are plenty of games that won’t give me a stable 100 FPS when running 3440×1440 on my 7700K @ 5.1GHz, 4133MHz RAM and GTX 1080 Ti.

And recommending SLI at this point in time is not all that wise. Multi-GPU (SLI and crossfire) support has plummeted the past 1-2 years. I used to have 2x GTX 980 and the amount of games I played that actually supported SLI was starting to really go down the drain and technologies like G-Sync didn’t work all that well when combined with SLI and whatnot. And with DirectX12 basically killing SLI and Crossfire all together and putting everything in the hands of the developers as they would need to successfully implement DirectX12 level of Multi-GPU support directly into the game and the game engine has proven to simply kill it altogether as no developer seems to care about adding it as there are so few people running Multi-GPU setups to begin with so it becomes a huge waste of development resources.

But yeah, this new TITAN V is all about compute like machine learning and AI. The branding is somewhat off here as GTX TITAN has always been a semi-professional and gaming card all in one. This time around this doesn’t seem to be the case. But it still lacks ECC memory so its obviously not a fully blown professional card like Quadro and Telsa but still.. Paying 3000 USD for something that seems to offer about twice the performance compared to a TITAN XP / GTX 1080 Ti is crazy for anyone not using these cards for compute.

Nvidia doesn’t support SLI anymore, but DX12 exposes all graphics adapters. It’s up to games to take advantage.

What? They support SLI, what are you talking about?

No support going forward. If you have it set up they’re not going to go back and break it of course, but they’re not actively pushing support in new games.

DX12 multiadaptor is where everything is going.

According to?

The Titan XP has SLI. This one does not seemingly, because it’s not really a gaming card I think, and they want to push you onto the Teslas with NVLINK if you wanted multi-GPU interlinks.

nVidia doesn’t support 3 or 4 way SLI anymore. 2 way still works fine, but the push behind it is mostly gone thanks to DX12

Yeah, that’s what I was thinking, 3 and 4 way are gone but 2 way is still around.

What do you mean Nvidia doesn’t support SLI anymore? They most certainly do. They just stopped supporting more than 2 cards in SLI, which arguably never worked well anyway.

This can’t be true? 110 Teraflops??? This is the most insane jump in technology that I’ve ever seen if this is really true. I’m at a loss for words.

I take it back. It’s only 15 Teraflops

Yeah they were referring to tensor operations.

Teraflop flop.

Tensor ops / CEO Math™

It looks like this GPU is aimed towards crypto-currency miners.

To me it doesn’t look anything like that. It is a massive loss if you try to crypto mine it with this thing with a huge installment tag.

I thought ASIC machines had obsoleted GPUs for crypto.

I think only for Bitcoin. Some crypto currencies such as Ethereum were designed to resist ASICs so you could use standard gaming GPUs instead.

100Tflops is from Tensor Ops. This is only a more modest upgrade from the Titan xP or 1080TI in CUDA core performance, which is where mining and gaming performance comes from.

So best mining card? Maybe, but excluding the price. For the price you could do better. The 100Tflops from tensor processing make this an AI training card.

For $3000 this is obviously a mining card. 1180 GTX probably gets released a few months later and is just as fast if not faster for games.

Obviously not, this is for CUDA and everything that goes with it (ie anything that benefits from mass parallelism such as linear algebra or more specifically deep learning).
There are better and cheaper alternatives for mining I’m sure.

Most of the performance is actually in the tensor cores rather than CUDA cores. Ho-hum CUDA upgrade, massive tensor upgrade, i.e for AI training.

Yep, read about it some more when I learned about the wtf-110-TFLOP. Made more sense.

Unless there is a mining algorithm that can make use of the Tensor cores then…no. Absolutely not.

Say it with me: Deep. Learning. That’s what this card is for.

View All Comments
Back to top ↑