Nvidia’s New Titan V Pushes 110 Teraflops From A Single Chip
Updated, 12/8/2017, 2:45pm PT: Added credit for spec table
Nvidia announced a new Titan-class GPU that puts the previous model to shame. The company put the power of its GV100 Volta GPU into a desktop-class graphics card—for science!
Nvidia’s new Titan V offers an unprecedented level of computing performance in a single graphics card. The company said that the Titan V produces 9x more performance than the Titan Xp in deep learning compute tasks.
The Titan V features Nvidia’s GV100 GPU, which debuted earlier this year in the Tesla V100 data center card. The GV100 GPU is fabricated on TSMC 12nm FFN high-performance manufacturing process and boasts a massive 815mm2 die with 21.1 billion transistors. The chip features 5,120 Cuda cores for traditional GPU compute power,and 640 Tensor cores for deep learning. The Cuda cores and Tensors cores all operate at 1,200MHz with the potential to boost to 1,455MHz.
The Titan V cards also include 12GB of 1.7 Gb/s HMB2 memory that operates on a 3,072-bit memory bus and provides 653 GB/s of memory bandwidth. Nvidia said that the card also features a new combined L1 data cache and shared memory unit, which “improves performance while also simplifying programming.”
The Titan V’s Volta architecture also offers independent integer and floating-point data paths, which enables the GPU to handle workloads that require both computation and addressing calculations with better efficiency than previous GPU architectures could.
“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,” said Jensen Huang, Founder, and CEO of Nvidia. “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.”
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Comes With Software
Nvidia emphasized that the Titan V is meant for scientists and researchers. The company said that anyone who purchases a Titan V graphics card would be granted access to “Nvidia-optimised deep learning frameworks, third-party managed HPC applications, Nvidia HPC visualization tools, and the Nvidia TensorRT inferencing optimizer.”
This One’s Not For Games
Nvidia’s Titan series graphics cards were never meant for gamers. The cards are meant for scientists who can use the advanced computational capabilities of Nvidia’s graphics hardware. Though Nvidia doesn’t market the Titan directly to consumers, that doesn’t stop gamers with deep pockets from picking up the best GPU money could buy.
With the Titan V, gamers likely won’t be as enticed to drop one into their PC. The cards boast incredibly high Tensor compute performance, but it’s unclear how that would translate to gaming performance. What’s more, Nvidia isn’t asking $1,200 for the Titan V as it did with the Titan X, Titan X pascal, and Titan Xp. The Titan V is available now and you can order them directly from Nvidia's website, but this time around, Nvidia is asking for big bucks for the Titan-level card. If you want a Titan V, get ready to pony-up a whopping $2,999. That’s Titan Z territory, but the Titan V doesn’t include two GPUs as the Z did.
Spec Table Credit: AnandTech
| Product | Titan V | Tesla V100 (PCIe) | Tesla P100 (PCIe) | Titan Xp |
|---|---|---|---|---|
| CUDA Cores | 5,120 | 5,120 | 3,584 | 3,840 |
| Tensor Cores | 640 | 640 | N/A | N/A |
| Core Clock | 1,200MHz | ? | ? | 1,485MHz |
| Boost Clock(s) | 1,455MHz | 1,370MHz | 1,300MHz | 1,582MHz |
| Memory Clock | 1.7 Gb/s HBM2 | 1.75 Gb/s HBM2 | 1.4 Gb/s HBM2 | 11.4 Gb/s GDDR5X |
| Memory Bus Width | 3072-bit | 4096-bit | 4096-bit | 384-bit |
| Memory Bandwidth | 653 GB/s | 900 GB/s | 720 GB/s | 547 GB/s |
| VRAM | 12GB | 16GB | 16GB | 12GB |
| L2 Cache | 4.5MB | 6MB | 4MB | 3MB |
| Half Precision | 30 TFLOPS? | 28 TFLOPS | 18.7 TFLOPS | 0.19 TFLOPS (1/64 rate) |
| Single Precision | 15 TFLOPS | 14 TFLOPS | 9.3 TFLOPS | 12.1 TFLOPS |
| Double Precision | 7.5 TFLOPS? | 7 TFLOPS | 4.7 TFLOPS | 0.38 TFLOPS |
| Row 12 - Cell 0 | (1/2 rate) | (1/2 rate) | (1/2 rate) | (1/32 rate) |
| Tensor Performance (Deep Learning) | 110 TFLOPS | 112 TFLOPS | N/A | N/A |
| GPU | GV100 | GV100 | GP100 | GP102 |
| Die Size | 815mm2 | 815mm2 | 610mm2 | 471mm2 |
| Transistor Count | 21.1B | 21.1B | 15.3B | 12B |
| TDP | 250W | 250W | 250W | 250W |
| Form Factor | PCIe | PCIe | PCIe | PCIe |
| Cooling | Active | Passive | Passive | Active |
| Manufacturing Process | TSMC 12nm FFN | TSMC 12nm FFN | TSMC 16nm FinFET | TSMC 16nm FinFET |
| Architecture | Volta | Volta | Pascal | Pascal |
| Launch Date | 12/07/2017 | Q3'17 | Q4'16 | 04/07/2017 |
| Price | $2,999 | ~$10,000 | ~$6,000 | $1,299 |
Kevin Carbotte is a contributing writer for Tom's Hardware who primarily covers VR and AR hardware. He has been writing for us for more than four years.
-
bit_user That's actually cheaper than I thought it'd be. The V100 dies are enormous, and in extremely high demand. I guess enough of them had a bad memory channel that they decided selling them as Titans wouldn't cannibalize their Tesla market too badly.Reply
BTW, for comparison, the Quadro P100 sells for about $6k (although I believe it's fully-functional). -
bit_user Reply
Probably, but its drivers might not be very well optimized for gaming. The few gaming benchmarks I could find on the Quadro P100 were a bit underwhelming.20461421 said:Yeah, but can it run Crysis? ;)
-
shrapnel_indie Reply“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.”
Nvidia’s Titan series graphics cards were never meant for gamers.
With the Titan V, gamers likely won’t be as enticed to drop one into their PC. The cards boast incredibly high Tensor compute performance, but it’s unclear how that would translate to gaming performance. {...} If you want a Titan V, get ready to pony-up a whopping $2,999.
That is a painful price for unknown performance. I'd dare say it will trounce everything currently on the market for gamers, but will the Tensor units get in the way? Will this be the hint towards the possibility with the GTX-30 series? (GTX-20 series seems to be hinted at being just a Pascal refresh.) -
redgarl Well well well... not really the TFLOP I was looking for. Basically it is a 15 TFLOP card.Reply -
Rock_n_Rolla 5120 Cudas / 640 Tensors / 3,072 bit mem bus / 15 TFLOPs in SP / 1.7Gbps on HBM2 / 250 Watts TDPReply
-- Probably will perform the same as GTX1080 or somewhere in the numbers close to 1080Ti since its specialized driver are streamlined with Ai / DL / Data crunching.
I BET Nvidia will relase a compatible driver for it sometime soon specifically designed for usual desktop and gaming and if that does happen and since its a Volta, theres a high possibility it will beat 1080Ti...
... And playing Crysis with it on its very highest custom settings will be a walk in a park. IMO -
extremepenguin What I want to know about this card is how the drivers stack up for it in Citrix when used as a GPU resource for hosted programs. We currently use Tesla's for it and they are not cheap if we could take out some of the older Tesla's and replace them with these it would be a huge win for me licensing and performance wise.Reply -
extremepenguin As for those asking if it can play Crysis I am sure there is somebody already working with OpenAI to accomplish just that if for the only reason of making that question go away. The answer will be yes and it plays it better than you.Reply -
mynith This time they're right though, it's not designed for gaming. It'll do it well, I'm sure, but a lot of functionality goes unused. I'm not sure what you'd use tensors for in game development other than in the vertex shader.Reply -
bit_user Reply
I have it on pretty good authority that they're semi-decoupled from the normal execution pipeline.20462465 said:will the Tensor units get in the way?
The main impact that the tensor units and fp64 units have on gaming is just consuming die area (and making it more expensive) with hardware that games wouldn't use. Even for normal fp16 calculations, you still don't use the tensor units - they're hard-wired for computing tensor products, which shaders don't normally do.
It would be interesting to see which is faster - Quadro P100 or Titan V. The Quadro has more memory bandwidth, but the Titan V has a completely new ISA and more "cores".