Nvidia releases Volta-based Titan V for PC
Nvidia simply introduced it is latest Titan graphics card on the NIPS 2017 convention, the Titan V. NIPS stands for Neural Information Processing Systems, and the main focus is on AI and deep studying. Not surprisingly, the Titan V will do very effectively in these areas, and whereas it is ostensibly a card that can be utilized for gaming functions, I do not suppose that many players are going to be prepared to half with a cool $three,000 for the privilege.
That’s greater than twice the price of the present champion, the Titan Xp—what may presumably warrant such an expense? Simple: tons and many compute energy, because of the Volta GV100 structure’s Tensor cores, and backed by 12GB of HBM2 reminiscence. Here are the complete specs for the brand new behemoth, with the Titan Xp and GTX 1080 Ti supplied for distinction.
Even ignoring the deep studying Tensor cores, it is a beast of a graphics card. With 5120 lively CUDA cores, this by far the very best performing graphics chip from Nvidia—and there are nonetheless 4 disabled SMM clusters on this implementation of the GV100 structure. Clockspeed is barely decrease than on the Titan Xp, and fairly a bit decrease than the 1080 Ti, however the added cores greater than make up for that.
In uncooked numbers, the GPU portion of the Titan V can do 14.9 TFLOPS of single-precision FP32 computations, which is 31 p.c greater than the Titan Xp. Though do observe that the reported increase clocks are conservative estimates—the 1080 Ti for occasion runs at nearer to 1620MHz in most of my gaming benchmarks.
This can also be the primary pseudo-consumer graphics card from Nvidia to make the most of HBM2 reminiscence, however solely three of the potential 4 HBM2 stacks on GV100 are lively. With a reminiscence clockspeed of 1.7Gbps and a 3072-bit reminiscence interface, that is good for 653GB/s of bandwidth.
But the actual energy of the Titan V, not less than insofar as deep studying and AI calculations are involved, comes from the 640 Tensor cores. There are eight special-purpose Tensor cores per SMM cluster, and every Tensor core can carry out a 4×4 FMA (fused multiply add) operation per cycle. That’s 64 FLOPS for the multiply, and one other 64 FLOPS for the addition, so 128 FLOPS whole. Multiply that by 640, and it appears just like the Tensor cores run at a barely decrease clockspeed, as Nvidia reviews a complete of 110 TFLOPS—that might be round 1340MHz, although Nvidia could be together with different components right here as effectively.
Utilizing the Tensor cores would require customized code—it isn’t one thing that you’re going to instantly profit from in video games—so that is positively not meant to be the last word gaming graphics card for subsequent 12 months. In reality, I think we’ll see one thing like a GV102 core in some unspecified time in the future that utterly omits the Tensor cores, maybe even sticking with GDDR5X or GDDR6 reminiscence as a substitute of HBM2, however that it could possibly be a while earlier than such a product materializes.
Architecturally, Nvidia has additionally acknowledged previously that Volta is not simply Pascal with the addition of Tensor cores. The structure has been modified by way of thread scheduling and execution, reminiscence controllers, the instruction set, the core format, and extra. But regardless of the huge die measurement and gobs of computational energy, it is spectacular that the Titan V stays a 250W half. That bodes effectively for the true client Volta elements that we’ll finally see a while in 2018.
Or in the event you’ve acquired $three,000 burning a gap in your pockets, you should buy the Titan V on Nvidia’s web site proper now. Hell, why not purchase a number of and slap them in a system for some critical computational energy? Just do not plan on SLI, as there aren’t any SLI connectors. Not that you’re going to want these if you’re constructing an AI to take over the world.