Enlarge (credit: Nvidia)

Although Nvidia launched its 21 billion transistor Volta GPU architecture back in May, until now the chip has been used exclusively in compute cards—specifically, the Tesla V100 cards, which cost about $10,000 for the PCIe version. But now Volta GPU is available in a graphics card: at the 2017 Neural Information Processing Systems conference, Nvidia CEO Jen-Hsun Huang announced the Titan V, a $3,000 golden video card making the same GPU—the Volta architecture GV100—available to regular end users.

Unlike the Tesla cards, the Titan V is a proper graphics card. It has three DisplayPort and one HDMI outputs, it uses the standard GeForce driver stack, and it will play games. It'll probably play them quite well, although Nvidia hasn't made any gaming performance claims thus far. That's because, although it's a graphics card, it's not really being aimed at graphics applications. Rather, it's being aimed at the same kinds of GPU-based computation (especially machine learning) applications that the Tesla cards are meant for, just on a slightly smaller scale. Titan V would be used in a workstation PC, rather than a compute cluster in a datacenter.

With these similar roles in mind, it's not too surprising that Titan V's specs are extremely similar to Tesla V100's. Both cards have 5,120 compute cores, and both have 640 machine learning-oriented "tensor cores" that specialize in 4×4 matrix arithmetic. Exact clock speeds for the Tesla cards aren't known, but Titan V can boost to 1.455GHz, slightly more than V100's 1.370GHz. It's the memory subsystem that has the biggest difference: Titan V has 12GB of HBM2 memory, with a 3,072-bit memory bus. Tesla V100 has 16GB of HBM2 and a 4,096-bit bus. The Titan's memory is also a hair slower, clocked at 1.70Gb/s compared to 1.75Gb/s.

Read 2 remaining paragraphs | Comments