Surprise, surprise! No rumors, no leaks. Not even a peep. And NVIDIA pulls out a new TITAN graphics card out of nowhere. The TITAN V was revealed at the 2017 Neural Information Processing Systems conference, with CEO Jen-Hsun Huang showing it off, to everyone’s surprise. This monster sports the GV100 GPU, same as the Tesla V100 accelerator.
The 815 mm2 GV100 die present on the TITAN V, has an insane CUDA core count of 5120, with 640 of the much advertised tensor cores specialized for neural network interfacing and other AI related workloads. The older TITAN cards while not exactly aimed for the PC video gaming community, passed as gaming cards with performance on par with their x80Ti counterparts. But the TITAN V is a different animal altogether. NVIDIA seems to be focusing purely on compute performance here, and the inclusion of the tensor cores and the venue of announcement again, was the Neural Information Processing Systems conference, which pretty much cements it.
If you compare it to the Tesla V100, the cards are astonishingly similar. The Titan V gets the full GV100 GPU along with all the fancy tensor cores. The thing that is different though is the memory and the lack of NVLink interface. These TITAN Vs are mainly going to be discarded Teslas, so it’s not that surprising if you take that into account. These are TESLA GPUs slapped onto TITAN PCBs, with pricey looking golden shrouds. These new TITANs don’t support SLI or any multi-GPU configurations, which is another new for the TITAN cards, again indicating that it’s not aimed towards gamers.Coming to the other specs, the TITAN V has 12GB HBM2 memory, clocked at 1.7Gbps, coupled with a 3072 bit bus, netting a bandwidth of 653GB/s. The Tesla V100 on the other hand has 16GBs of HMB2 memory, with a 4096 bit wide bus and a resulting bandwidth of 960GB/s. The base core clock is 1300 Mhz, with a boost clock of 1455 Mhz, slightly higher than the Volta based Tesla accelerator.
The card features a premium looking golden shroud which makes it look all the more expensive, which it actually is. But the shroud’s design is actually the same as all other TITAN cards. The ports are also the standard offerings, 3 HDMI and 1 DP. The TDP is rated at 250W. A curious thing to note about the PCB is that it has NVlink connectors but the shroud blocks them. It would be interesting to note if NVlink actually works on this card once you circumvent the shroud.To the common person, the 3000 US$ price tag is intimidating to say the least. But the TITAN V is basically a cheaper TESLA which is otherwise available for a steep price of around 10K US$. So while, it may seem like an unnecessary addition to NVIDIA’s already loaded GeForce lineup, in certain circles, mainly compute-heavy AI driven ones, this card should probably do well.
Now let’s recap.
This card is NOT for gamers. While it is powerful in terms of graphics performance than all the present consumer cards, it is still more focused towards compute- oriented tasks and neural network interfacing.
110 TFLOPs of deep learning peformance, AKA Tensor performance. A compact super-computer now available to the general public. Well, general public with fat paychecks anyway. This should benefit sponser-funded scientists and researchers.
Free AI Software on NVIDIA GPU Cloud. This is the silver lining for developers working in the field of AI and neural networks. Consumers can gain immediate access to the latest GPU-optimized AI, deep learning and HPC framework by signing up at no charge for an NVIDIA GPU Cloud account. This container registry includes NVIDIA-optimized deep learning frameworks, third-party managed HPC applications, NVIDIA HPC visualization tools and the NVIDIA TensorRT™ inferencing optimizer.
Another cool video making the GPU look like a Ferrari just like ever other NVIDIA flagship “trailer”.