Though Nvidia launched its 21 billion transistor Volta GPU structure again in Could, till now the chip has been used completely in compute playing cards—particularly, the Tesla V100 playing cards, which value about $10,000 for the PCIe model. However now Volta GPU is on the market in a graphics card: on the 2017 Neural Info Processing Programs convention, Nvidia CEO Jen-Hsun Huang introduced the Titan V, a $three,000 golden video card making the identical GPU—the Volta structure GV100—accessible to common end-users.
In contrast to the Tesla playing cards, the Titan V is a correct graphics card. It has three DisplayPort and an HDMI outputs, it makes use of the usual GeForce driver stack, and it’ll play video games. It will in all probability play them fairly effectively, although Nvidia hasn’t made any gaming efficiency claims up to now. That is as a result of though it is a graphics card, it is probably not being geared toward graphics purposes. Slightly, it is being aimed on the identical sorts of GPU-based computation (particularly machine studying) purposes that the Tesla playing cards are meant for, simply on a barely smaller scale. Titan V can be utilized in a workstation PC, relatively than a compute cluster in a datacenter.
With these related roles in thoughts, it is not too shocking that Titan V’s specs are extraordinarily much like Tesla V100’s. Each playing cards have 5,120 compute cores, and each have 640 machine learning-oriented “tensor cores” focusing on four×four matrix arithmetic. Precise clock speeds for the Tesla playing cards aren’t identified, however Titan V can enhance to 1.455GHz, barely greater than V100’s 1.370GHz. It is the reminiscence subsystem that has the largest distinction: Titan V has 12GB of HBM2 reminiscence, with a three,072-bit reminiscence bus. Tesla V100 has 16GB of HBM2 and a four,096-bit bus. The Titan’s reminiscence can be a hair slower, clocked at 1.70Gb/s in comparison with 1.75Gb/s.
Taken collectively, the brand new card has virtually an identical computational efficiency—13.eight trillion single precision floating level operations per second, 6.9 double precision TFLOPS, and 110 decreased precision machine studying TFLOPS in Titan V, in comparison with 14 single precision, seven double precision, and 112 machine studying TFLOPS on V100—however with reminiscence bandwidth of a mere 653GB/s, in comparison with 900GB/s for the compute card.
The excessive worth and compute focus imply that despite its bodily look, the Titan V is aimed fairly squarely on the “Professional” facet of the “prosumer” buyer that the Titan line has historically been geared toward. It is inevitable that some deep-pocketed avid gamers will choose up a Titan V and use it as nothing greater than a graphics card, however that is definitely not the core market.