Out of nowhere, NVIDIA has printed the NVIDIA Titan V this day at the 2017 Neural Recordsdata Processing Programs convention, with CEO Jen-Hsun Huang flashing out the card on stage. A mere 7 months after Volta used to be announced with the Tesla V100 accelerator and the GV100 GPU inner it, NVIDIA continues its breakneck trot by releasing the GV100-powered Titan V, accessible within the marketplace this day. Geared in direction of a decidedly extra compute-oriented market than ever before, the 815 mm2 behemoth die that is GV100 is now accessible to the broader public.
|NVIDIA Compute Accelerator Specification Comparability|
|Titan V||Tesla V100
|Memory Clock||1.7Gbps HBM2||1.75Gbps HBM2||1.4Gbps HBM2||eleven.4Gbps GDDR5X|
|Memory Bus Width||3072-bit||4096-bit||4096-bit||384-bit|
|Single Precision||13.8 TFLOPS||14 TFLOPS||9.three TFLOPS||12.1 TFLOPS|
|Double Precision||6.9 TFLOPS
(1/2 of rate)
(1/2 of rate)
(1/2 of rate)
|a hundred and ten TFLOPS||112 TFLOPS||N/A||N/A|
|Manufacturing Route of||TSMC 12nm FFN||TSMC 12nm FFN||TSMC 16nm FinFET||TSMC 16nm FinFET|
|Originate Date||12/07/2017||Q3’17||This autumn’sixteen||04/07/2017|
For the spec sheet now we bear long past forward and lined it up against NVIDA’s different Pascal cards, and for loyal reason. While the Titan sequence of cards might well well bear started existence as a prosumer card in 2013, since then NVIDIA’s GPU designs bear become increasingly extra divergent between compute and graphics. And even supposing the outdated Titan Xp used to be in accordance to the extra graphics-centered GP102 GPU, the card itself used to be basically (however no longer thoroughly) pitched as an entry-level compute card, for purchasers who wished a (somewhat) low-rate manner to attain FP32 compute and neural network inferencing in workstations and small clusters.
The Titan V, by extension, sees the Titan lineup at final switch loyalties and commence the use of NVIDIA’s excessive-finish compute-centered GPUs, on this case the Volta architecture basically basically basically based V100. The tip consequence is that as an various of being NVIDIA’s prime prosumer card, the Titan V is decidedly extra centered on compute, in particular attributable to the combination of the trace tag and the uncommon characteristic dwelling that comes from the use of the GV100 GPU. Which isn’t to verbalize that it is seemingly you’ll per chance well well perhaps also’t attain graphics on the card – that is unexcited very unparalleled a video card, outputs and all – however NVIDIA is at the initiating promoting it as a workstation-level AI compute card, and by extension focusing on the GV100 GPU’s uncommon tensor cores and the extensive neural networking efficiency advantages they provide over earlier NVIDIA cards.
On this sense the Titan V is a return to affect of kinds to the knowledgeable aspect of prosumer for the Titan household. One in all the genuine claims to status for the genuine Titan used to be its excessive efficiency in specialised FP64 compute workloads, something that used to be lost on the later Titan X and Titan Xp. By switching to NVIDIA’s specialised excessive-finish compute GPUs, the Titan V regains its formerly lost compute capabilities, the total whereas also gaining the total compute capabilities NVIDIA has launched since then. It’s no mistake that Jen-Hsun launched the card at a neural networking convention, as that is a smartly-behaved chunk of the knowledgeable computing target audience that NVIDIA is concentrating on with the card.
Apparently, evaluating it to the PCIe Tesla V100, I’m shocked by unbiased how finish the cards are in ideas and efficiency. NVIDIA has confirmed that the Titan V gets the GV100 GPU’s fleshy, unrestricted FP64 compute and tensor core efficiency. To the most realistic of our info (and from what NVIDIA will comment on) it doesn’t seem that they’ve artificially disabled any of the GPU’s core ideas. What does separate the Titan from the Tesla then from a efficiency standpoint is moderately easy: reminiscence skill, reminiscence bandwidth, and the shortcoming of NVLink functionality. There are also a preference of smaller variations between the cards that assist to differentiate them between server and workstation – similar to passive versus energetic cooling, NVLink, and the increase insurance policies – however in any other case for purchasers who are running a small preference of cards, the Titan V’s characteristic dwelling is remarkably finish to the unparalleled extra dear Tesla V100’s, which is a extremely titillating constructing because it goes to uncover unbiased how assured NVIDIA is that this won’t undermine Tesla gross sales.
Transferring on and diving into the numbers, Titan V ideas 80 streaming multiprocessors (SMs) and 5120 CUDA cores, the same quantity as its Tesla V100 siblings. The variations reach with the reminiscence and ROPs. In what’s clearly a salvage piece for NVIDIA, one of the crucial card’s Four reminiscence partitions has been carve, leaving Titan V with 12GB of HBM2 connected via a 3072-bit reminiscence bus. As every reminiscence controller is said to a ROP partition and 768 KB of L2 cache, this in turn brings L2 down to Four.5 MB, as successfully as reducing down the ROP depend.
When it comes to clockspeeds, the HBM2 has been downclocked somewhat to 1.7GHz, whereas the 1455MHz enhance clock genuinely suits the 300W SXM2 variant of the Tesla V100, despite the fact that that accelerator is passively cooled. Particularly, the preference of tensor cores bear no longer been touched, despite the fact that the wonderful a hundred and ten DL TFLOPS rating is lower than the 1370MHz PCIe Tesla V100, as it might well per chance well well appear that NVIDIA is the use of a clockspeed lower than their enhance clock in these calculations.
For the card itself, it contains a vapor chamber cooler with copper heatsink and sixteen vitality phases, involved within the 250W TDP that has become fashioned with the single GPU Titan units. Output-wise, the Titan V brings three DisplayPorts and 1 HDMI connector. And as for card-to-card dialog, PCB itself looks to bear NVLink connections on the pinnacle of the PCB, however these gape to bear been intentionally blocked by the camouflage to prevent their use and are presumably disabled.
As talked about earlier, NVIDIA is unsurprisingly pushing this as a compute accelerator card, in particular brooding about that Titan V ideas tensor cores, and conserving the TITAN branding versus GeForce TITAN. However there are these of us who know higher than to buy folk won’t tumble $3000 to use the most recent Titan card for gaming, and whereas gaming is no longer the main (or even secondary) center of attention of the card, you also is no longer going to appear NVIDIA denying it. In that sense the Titan V goes to be handled as a jack-of-all-trades card by the firm.
To that finish, no gaming efficiency info has been disclosed, however NVIDIA has confirmed that the card makes use of the fashioned GeForce driver stack. Now whether or no longer these drivers bear genuinely be optimized for the GV100 is one more topic thoroughly; Volta is a brand novel architecture, markedly so every so steadily. Speaklng thoroughly off the cuff here, for graphics workloads the card has extra resources than the Titan Xp in nearly every meaningful metric, however it be also a smaller distinction on paper than you would assume.
As for NVIDIA’s intended market of compute and AI users, the Titan V will be supported by NVIDIA GPU Cloud, which incorporates a preference of deep finding out frameworks and HPC-connected instruments.
If the golden camouflage didn’t already indicate so, the Titan V is also carving out a brand novel eye-watering trace level, losing in at $2999 and on sale now at the NVIDIA store. NVIDIA has, to this level, been promoting Tesla V100 merchandise as swiftly as they can fabricate them, so I’m no longer going to be shocked if the Titan V sees an identical destiny. The $3000 trace tag is moderately excessive, even by Titan requirements, however with the rare Tesla V100 PCIe card going for spherical $10,000, the Titan V is markedly more cost effective. In level of truth in some respects I’m shocked NVIDIA is promoting a GV100 card for thus runt; these are GV100 salvage ingredients that establish no longer manufacture the carve for Tesla – so the many will seemingly be throwing them away – however it unbiased goes to uncover how assured NVIDIA is that it is no longer going to undermine the Tesla household.
At any rate, for NVIDIA knowledgeable users who bear been seeking to dip their toes into Volta however didn’t desire a fleshy-fledged Tesla card, the Titan V is clearly going to be a usual card. Over the final two years NVIDIA’s AI efforts bear been firing on all cylinders, and by bringing a GV100 card down to unbiased $3000, inquire of to gape them crack originate the market that unparalleled further. I dare direct the thought that of the “prosumer” Titan has died with this card, however for the unexpectedly growing knowledgeable compute market, this looks to be precisely the affect of card that a form of developers bear been ready for.