Puget Systems print logo

https://www.pugetsystems.com

Read this article at https://www.pugetsystems.com/guides/1080
William George (Puget Labs Technician)

NVIDIA Titan V Surprise

Written on December 8, 2017 by William George
Share:

NVIDIA's CEO, Jen-Hsun Huang, dropped a bit of a bombshell at the NIPS conference yesterday: the launch - and immediate availability - of the next graphics card in NVIDIA's Titan series. It is called the Titan V, with V referring to the new Volta architecture it is based on. Titan naming has been all over in the past, so don't try to make any sense of it alphabetically. After all, V comes before X and Z... even those this card is newer than the Titans which used those letters.

What is most interesting about this new Titan V, though, is where it is aimed. In the past, Titan cards have toed the line between mainstream and professional graphics realms. Some Titans have been advertised as gaming cards, some have not. Some Titans bore the GeForce branding of NVIDIA's mainstream graphics cards, more recent models have not. The very first Titans had great compute performance, which was lost on some of the later models, but has returned according to the specs NVIDIA released last night.

So what does that mean for the new Titan V's performance? And what does the new Volta architecture bring to the table? We will be getting a pair of Titan Vs in the near future for testing, at which point we will be able to answer these questions more definitively, but for now here is what we can surmise from the card's specs and what is known about Volta and the GV100 chip (which is also used in the Tesla V100):

  • More CUDA cores but lower clock speed than the existing Titan Xp. That should translate to better general graphics performance, but not a massive change. This card should perform well in games and mainstream GPU applications, but it won't be twice the speed of current cards in those situations (despite being more than twice the price).
  • The same 12GB of memory as recent Titan models, but a faster type of memory. Instead of GDDR5X memory, found on the Titan Xp and GTX 1080 cards, the Titan V has HBM2. This allows for much higher memory bus width and clock speeds, and in turn higher overall memory bandwidth. It isn't as fast as the enterprise-grade Tesla V100, which has ~33% more VRAM and bandwidth, but it is the fastest memory on any NVIDIA card outside the Tesla line.
  • Double precision calculation speed has not been artificially limited on the Titan V, as it is on most of NVIDIA's consumer video cards. This hearkens back to the original Titan's feature set, and it means that workloads using FP64 will be many times faster than on other Titan or GeForce cards from recent years. NVIDIA quotes performance of 6.9 TFLOPS in double precision mode, which is more than 18 times the Titan Xp's 0.38 TFLOPS rating!
  • Tensor cores are a new GPU feature that NVIDIA has added in the Volta architecture, and which are present on the GV100 GPU and thus the Titan V. They are specifically tailored to perform single and half-precision floating point calculations (FP32 & FP16) for use in machine learning frameworks. They won't help in more mundane GPU applications, but if you are working in Tensorflow, Caffe2, etc then they will provide a big boost in performance over previous generations of NVIDIA cards.

So what does all of that mean for computer users? It depends on what you are doing. For gamers, not much. The upper-end GeForce 1070 / 1080 series cards, or even a Titan Xp, will give you much better performance per dollar... even if the Titan V might technically be a little faster overall (we'll have to wait till gaming sites get their hands on it to know for sure). For graphics professionals, running GPU-accelerated programs like Premiere Pro, Resolve, OctaneRender, and the like, also not a lot. We will be testing some of those programs with the Titan V, but from the specs alone I don't expect it to be able to justify its high price tag for most content creators.

But for developers working on machine learning, deep learning, and AI? This is where the Titan V will shine!

It provides similar specs / performance to the Tesla V100 which costs more than three times as much, while also having standard HDMI and DisplayPort outputs that you need in a workstation (as opposed to a headless server, where most Teslas reside). It also has a high quality heatsink and fan for cooling, another thing that is missing from the Tesla cards that require specialized cooling from the system that houses them. Being able to get this sort of horsepower for the full range of half, single, and double precision calculations in a tower workstation or even a compact desktop will be great for folks working in this space. I am excited to see what our own Dr. Don Kinghorn finds when he gets his hands on a Titan V! Check out his HPC Blog on our website to see content from him about it in the coming weeks.

Tags: NVIDIA, Titan, V, GPU, graphics, video, card, Volta, GV100, machine, deep, learning, AI, HPC, compute, double, single, half, precision, FP64, FP32, FP16, tensor
COFASA

Considering this card it aimed at AI and considering all the features it has the price seems to be about right.

The problem is the consumism that always arises with this cards people but them for gaming even though nvidia always releases a more affordable card with similar gaming performance under the 80ti naming scheme

They make pricing go throught the roof, that explains why although the titan xp war originally launched for $1200 it can be found for $1500 in some places.

Posted on 2017-12-09 18:53:05

Yeah, I agree that for the specs and focus this card has it is priced reasonably. A large part of why I wrote this blog post was to steer gamers and content creators *away* from the Titan V, since it doesn't look like it will give good price:performance for those applications. We'll likely get Volta-based GeForce cards sometime in 2018, for the wider "mainstream" user base.

Posted on 2017-12-11 16:54:26
Ryan Julian

Do you have a firm number from NVIDIA on the FP16 performance? The assumption would be 2xFP32 = 30TFLOPs, but I can't find a data sheet with any claims.

Posted on 2017-12-09 19:55:28

I have not seen FP16 numbers that I was sure of yet. AnandTech's article initially listed those, but then removed them - I'm not sure why, but I can't recall what they were and I suspect they may have found out they were not correct or something anyway.

If you are able to make use of the Tensor cores, which are optimized for FP16 calculations, I think you'd see even more than 30TFLOPS... but if you need to stick with just using the normal CUDA cores, then your estimate is probably a good ballpark. We'll have to wait and see, I guess. I'm not sure if anything Dr. Kinghorn tests will use FP16.

Posted on 2017-12-11 16:51:38
qwertpo

This is good news to me - the NVIDIA website doesn't say anything about FP64 cores, but that's what I need for my work. I've been living with a Titan Black for a loooong time.

Posted on 2017-12-15 16:55:56

Yeah, that is one place the Titan V really shines! I hope I am not stealing Don's thunder, but he did some testing and saw real-world results of over 4 TFLOPS with FP64 workloads on the Titan V. NVIDIA's specs indicate it should go even higher, probably depending on what exactly you are doing, but either way that is a huge improvement over previous Titans, GeForce, and even most Quadro cards.

Posted on 2017-12-15 17:05:25