Out of the blue, NVIDIA has unveiled the NVIDIA Titan V today at the 2017 Neural Information Processing Systems conference, with its CEO Jen-Hsun Huang showing the card on stage. Just 7 months after Volta was announced with the Tesla V100 accelerator and the GV100 GPU within it, NVIDIA continues its dizzying pace by launching the Titan V with GV100 engine, available today for sale. Aimed at a market decidedly more computer-oriented than ever, the gigantic die of 815 mm 2 which is GV100 is now available to the general public.
|Specification comparison of the computation accelerator NVIDIA|
|Titan V|| Tesla V100
| Tesla P100
|Cores tensioners||640||640||N / A  N / A|
|Memory Clock||1.7Gbps HBM2||1.75Gbps HBM2||1.4Gbps HBM2||] 11.4Gbps GDDR5X|
|Memory bus width||3072-bit||4096-bit||4096-bit  384-bit|
|Memory bandwidth||653GB /||900GB / sec||720GB / sec||547GB / sec|
|Single precision||13.8 TFLOPS||14 TFLOPS||9.3 TFLOPS||12.1 TFLOPS|
|Double precision|| 6.9 TFLOPS
| 7 TFLOPS
| 4.7 TFLOPS
| 0.38 TFLOPS
| Performance of Tensor
|110 TFLOPS||112 TFLOPS||N / A||N / A|
|TDP||250W||250W||250W||250W  Form factor||PCIe||PCIe||PCIe||PCIe|
|Manufacturing process||TSMC 12nm FFN  TSMC 12nm FFN||TSMC 16nm FinFET||TSMC 16nm FinFET|
|Release date||07/12/2017||Q3 & # 39; 17||Q4 & # 39; 16||04/07/2017|
|Price||$ 2999||~ $ 10000||~ $ 6000||$ 1299|
For the spec sheet that I followed and aligned with the other NVIDIA Pascal cards, and for good reason. While the Titan card series may have started its life as a prosumer card in 2013, since then the NVIDIA GPU designs have become increasingly divergent between computing and graphics. And although the previous Titan Xp was based on GPU GPU2 graphics-centric, the card itself was primarily (but not only) released as an entry-level computing card, for customers who needed a form (relatively ) cheap to do FP32 computation and inference of neural networks in workstations and small clusters.
The Titan V, by extension, sees the Titan line-up finally switch loyalties and start using GPUs focused on NVIDIA's high-end computing, in this case the Volta architecture based on V100. The end result is that, instead of being the main NVIDIA Prosumer card, the Titan V is decidedly more focused on the calculation, particularly due to the combination of the price tag and the set of unique features that come from the use of the V100 GPU. What does not mean that you can not make graphics on the card, this is still a video card, outputs and everything, but NVIDIA is promoting first and foremost as a workstation-level artificial intelligence computer card, and by extension focusing on the unique tensor cores of the V100 GPU and the huge performance advantages of neural networks they offer over previous NVIDIA cards.
In this sense, the Titan V is a kind of return to the prosumer's professional genre for the Titan family. One of the original claims to fame for the original Titan was its high performance in specialized FP64 computing workloads, something that was lost in the later Titan X and Titan Xp. By switching to NVIDIA's specialized high-end computing GPUs, the Titan V recovers its previously lost computer capabilities, while also obtaining all the computing capabilities that NVIDIA has introduced since then. It is not a mistake that Jen-Hsun presented the card at a neural network conference, since this is a large part of the professional computer audience that NVIDIA is targeting with the card.
Interestingly, comparing it with the Tesla V100 PCIe, I'm I'm surprised how close the cards are to features and performance. NVIDIA has confirmed that the Titan V obtains the complete and unrestricted FP64 computation of the V100 GPU and the performance of the tensioner core. As far as we know (and what NVIDIA will comment on), it does not appear that they have artificially disabled any of the main features of the GPU. What separates the Titan from the Tesla from a performance point of view is quite simple: memory capacity, memory bandwidth and the lack of NVLink functionality. There are also a number of small differences between the cards that help differentiate between the server and the workstation, such as passive versus active cooling, NVLink and support policies, but otherwise for customers running a small number of clients. cards, the Titan The feature set of V is notably closer to the much more expensive Tesla V100, which is a very interesting development as it shows how safe NVIDIA is that this will not undermine Tesla's sales.
Advancing and immersing itself in the numbers, Titan V has 80 transmission multiprocessors (SM) and 5120 CUDA cores, the same amount as their Tesla V100 brothers. The differences come with memory and ROP. In what is clearly a salvage part for NVIDIA, one of the 4 memory partitions of the card was cut off, leaving Titan V with 12 GB of HBM2 connected through a 3072-bit memory bus. Since each memory controller is associated with a ROP partition and 768 KB of L2 cache, this in turn reduces L2 to 4.5 MB, as well as reducing the ROP count.
In terms of clock speeds, the HBM2 has been deciphered slightly at 1.7 GHz, while the 1455MHz boost clock actually matches the 300W SXM2 variant of the Tesla V100, although that accelerator cools passively. Notably, the number of cores has not been touched, although the official rating of 110 DL TFLOPS is lower than the PCIe Tesla V100 of 1370 MHz, since it seems that NVIDIA uses a clock speed lower than its impulse clock in these calculations .
For the card itself, it features a steam chamber cooler with copper heatsink and 16 power phases, all for the TDP of 250 W that has become standard with the models of Titan GPU individual. As for the output, the Titan V brings 3 DisplayPorts and 1 HDMI connector. And as for card-to-card communication, the PCB itself seems to have NVLink connections on top of the PCB, but these appear to have been intentionally blocked by the cover to prevent its use and are presumably deactivated.
As mentioned above, NVIDIA, as expected, promotes this as a computer accelerator card, especially considering that Titan V has tension cores, and keeps the TITAN brand in opposition to GeForce TITAN. But there are those who know better than to assume that people will not spend $ 3000 to use the latest Titan card for games, and although games are not the primary (or even secondary) focus of the card, you will not see NVIDIA deny it either. In that sense, the company will treat the Titan V as an illegal trading card.
To this end, no information has been released on the performance of the games, but NVIDIA has confirmed that the card uses the standard Stack of GeForce drivers. Now, if those drivers have actually been optimized for the GV100 it's something completely different; Volta is a new architecture, remarkably so sometimes. Here, speaking only of the bracelet, for graphics workloads, the card has more resources than Titan Xp in almost all significant metrics, but it is also a minor difference in paper than you think.
Regarding the computer and artificial intelligence market of NVIDIA users, the Titan V will be supported by the NVIDIA GPU Cloud, which includes a series of deep learning frameworks and tools related to HPC.
If the golden shroud did not suggest it, the Titan V is also creating a new dazzling price point, priced at $ 2999 and now on sale at the NVIDIA store. NVIDIA, to date, has been selling Tesla V100 products as fast as they can produce them, so it will not surprise me if the Titan V sees a similar destination. The $ 3000 price tag is quite high, even by Titan standards, but with the rare Tesla V100 PCIe card for around $ 10,000, the Titan V is significantly cheaper. In fact, in a way, I'm surprised that NVIDIA is selling a GV100 card for so little; these are GV100 salvage pieces that do not make the cut for Tesla, so the alternative would be to throw them away, but it only shows how safe NVIDIA is that it will not undermine the Tesla family.
Anyway, for professional NVIDIA users who have been looking to get their feet into Volta but did not want a full Tesla card, the Titan V will clearly be a popular card. In the last two years, NVIDIA's AI efforts have been running at full speed, and by taking a GV100 card to just $ 3000, we hope to see them open up the market much more. I would venture to say that the idea of the titan "prosumer" has died with this card, but for the rapidly growing professional computer market, this seems to be exactly the kind of card that many developers have been waiting for.