NVIDIA TITAN V

How to build your own Skynet

LOOKING AT THE CALENDAR on the wall, we realize that Nvidia’s Pascal debuted about two years ago now. It used to be that we could expect some massive changes in that time, but today’s Titan Xp is built off that same original GP100 that first appeared way back in the Tesla P100. That was an insane part, built for machine learning, and the GP100 never showed up in a consumer product (unless you count the $7,499 Quattro GP100). Nvidia has decided to try a new approach with its GV100 core—which sits at the heart of the Tesla V100 and goes after a similar market as the P100—and has released the prosumer and machine learning focused Titan V.

Nvidia’s Titan series of graphics cards has always been more about status than practicality, and waiting a few months could usually net you nearly the same level of performance for about half the price. This time, rumors are swirling about Volta, suggesting we might not even see GeForce variations, and instead will get Ampere GPUs later this year. That’s not really a problem, assuming Ampere is to the GV100 what the GP104 was to the GP100. The problem is that GV100, for all its performance, has a lot of extra design elements that simply aren’t ideal for graphics workloads.

Take the 64 CUDA cores per SM (streaming multiprocessor), with 32 FP64 (floating point) cores, and the ability to do twice as many packed math (FP16) operations as FP32 ops. FP64 isn’t something games use or even need, though there may be some algorithms where the packed math support proves beneficial. Even more than this, the presence of the new Tensor cores—eight per SM—takes a lot of space, and these cores are special purpose constructs designed to accelerate FP16 machine learning operations. Toss in HBM2, and you have a very large, very expensive core that’s not a great fit for consumer workloads.

That’s why Nvidia is marketing the Titan V not at extreme gaming, but at machine learning researchers. Without such a product, the only way to get access to Nvidia’s Tensor cores would be through DGX-1 servers, or a few cloud computing options. But the DGX-1 is back-ordered, and even if it wasn’t, it runs at $149,000 per server, including eight Tesla V100 cards. Three grand for a Titan V that includes all the core features is a bargain by comparison, and opens the door for university researchers to come to grips with the technology before moving to larger supercomputing installations.

In our review of the Falcon Northwest Tiki last issue, we noted that the Titan V is a potent solution for both machine learning and graphics work, easily claiming top honors in our gaming test suite. But how does the Titan V perform if we throw caution to the wind and overclock it? We managed to push the core clocks up by 150MHz and the HBM2 by 125MHz. Combined with maxing out the power limit—and the fan speed— performance improves by a substantial 25–30 percent, but the GPU starts hitting thermal and power limits, meaning there’s potentially even more headroom available.

Even redlining a GTX 1080 Ti or Titan Xp, you’re not coming anywhere near the overclocked Titan V. However, it is worth noting that the Titan V doesn’t include SLI connectors, so while running multiple cards for machine learning purposes is possible, you’re not going to be able to improve gaming performance beyond what you get with a single card.

Regardless of whether Volta GPUs show up in GeForce cards in the coming months, some variation on the gaming aspects of the Titan V will almost certainly appear. Give us 12GB GDDR6 and the same 5,120 FP32 CUDA cores (or more) but in a smaller, more affordable package, and you can call it Volta, Ampere, Turing, or Marie. Whatever the name, it will be awesome, and Dream Machine 2018 is already on the drawing board. As for the Titan V, unless you are after a machine learning tool, save your money and wait for the next generation non-Titan cards to arrive.

Nvidia Titan V

ALESSANDRO’S BATTERY Fastest GPU around; overclockable; machine learning.

FLAT BATTERY Exorbitant pricing; power/ thermal limits; wait for next-gen GeForce.

$2,999, www.nvidia.com