About seven months back, Nvidia propelled the Tesla V100, a $10,000 Volta GV100 GPU for the supercomputing and HPC markets. This huge card was expected for specific markets on account of its huge bite the dust measure (815 mm sq) and enormous transistor check (21.1B). Consequently, it offered specific tensor centers, 16GB of HBM2, and hypothetical execution in specific workloads far above anything Nvidia had sent some time recently.
Today, at the Conference on Neural Information Processing Systems (NIPS), Jen-Hsun astonish propelled the same GV100 engineering in a customary GPU frame factor. Similarly as the GTX 1080 Ti is a trimmed-down rendition of the Nvidia Titan Xp, this new Titan V thins down in a few spots contrasted and the full-fat Tesla V100. Memory timekeepers are somewhat lower (1.7Gbps exchange rate, down from 1.75Gbps), and the GPU has three memory ways at 3,072 bits, instead of the 4,096-piece interface the Tesla V100 offers. It additionally offers only 12GB of HBM2, as opposed to the 16GB on the Tesla V100.
Nvidia is trumpeting the Titan V as offering 110 TFLOPS of drive, "9x that of its antecedent." We don't question that is actually valid, yet it's not a correlation with the single-exactness or twofold accuracy math we've ordinarily referenced while examining GPU FLOPS execution. It's a reference to Volta's execution change in profound learning undertakings over Pascal, and it' s determined by looking at Volta's tensor execution (with its specific tensor centers) against Pascal's 32-bit single-accuracy throughput. That doesn't mean the correlation is invalid, since Volta has particular tensor centers for preparing neural systems, and Pascal doesn't, however it's similar to looking at AES encryption execution on a CPU with specific equipment for that workload with another CPU that needs it. Is the examination reasonable? Totally. Be that as it may, it's reasonable just for the particular metric being measured, rather than being a generalizable experiment for the rate of change one CPU offers over the other.
Nvidia's expressed objective with the Titan V is to offer scientists who don't approach supercomputers or enormous iron HPC establishments a similar access to front line equipment execution that their comrades appreciate. While the GPU is estimated at an eye-popping $3,000 (in respect to the standard PC advertise), that is not especially contrasted and the run of the mill cost of a HPC server.
"Our vision for Volta was to push the external furthest reaches of elite registering and AI. We broke new ground with its new processor design, guidelines, numerical arrangements, memory engineering and processor joins," said Nvidia CEO Jen-Hsun Huang. "With TITAN V, we are placing Volta under the control of analysts and researchers everywhere throughout the world. I can hardly wait to see their leap forward revelations."
You can purchase a Titan V at the Nvidia store at the present time, yet we can't sincerely say we'd suggest one for anybody not working in these fields. Notwithstanding the "Titan" mark having initially appeared as a top of the line purchaser card with some particular logical process capacities, this GPU family has been moving back towards its logical registering research pulls for various years. While Nvidia will clearly bolster the GPU with a bound together driver display, I wouldn't hold my breath sitting tight for calibrated gaming support from a GPU family that so couple of clients will ever approach.