In July 2016, NVIDIA launched TITAN X, based on Pascal, promising 11 teraflops (TFLOP) of computing performance. In April of this year, NVIDIA increased performance a bit with a fully unlocked Pascal GP102 core in the form of TITAN Xp, which has 12 TFLOP computing power.
Now, NVIDIA is ready to unleash a completely new GPU for researchers and scientists, the TITAN V. The card was properly announced at the Conference and Workshop on Neural Information Processing Systems (NIPS) in Long Beach , California. The new TITAN V is based on the NVIDIA Volta architecture, and with that change there is a gigantic leap in computing performance with certain workloads. According to the company, the TITAN V redesigned Tensor Cores, which have parallel floating and independent data paths in parallel, help to drive an incredible 9x increase in Tensor Performance compared to the TITAN Xp, reaching the vertiginous 110 TFLOP for Deep Learning workloads. The performance of a single precision will probably reach a much lower level of 14 TFLOP, more or less depending on the frequency of the GPU.
All that power is included in the GV100 GPU of 21,100 million transistors that is manufactured with a high performance 12 nm FFN process from TSMC, customized for NVIDIA. The TITAN V is loaded with 5120 CUDA cores, 640 cores, and has base / pulse clocks of 1200MHz and 1455MHz respectively. The HBM2 memory runs at 1.7 Gbps with a 3072-bit interface (for an effective bandwidth of 653 GB / sec). While the GV100 GPU powered by the NVIDIA Tesla V100 accelerator is paired with 16 GB of HBM2, the TITAN V comes with 12 GB of HBM2. Other details that NVIDIA is willing to share at this time include the fact that TITAN V has a new combined data cache L1 and a shared memory unit to improve overall performance. Reportedly, energy efficiency has also dramatically improved with Volta; NVIDIA is claiming great performance gains, within the same power envelope as Pascal.
"Our vision of Volta was to push the external limits of high performance computing and artificial intelligence, we have opened a new path with its new processor architecture, instructions, numerical formats, memory architecture and processor links, "said NVIDIA CEO Jensen Huang . "With TITAN V, we are putting Volta in the hands of researchers and scientists from around the world, I can not wait to see its breakthrough discoveries."
As expected, the TITAN V will be added to NVIDIA GPU Cloud, which provides a quick access to a fully optimized deep learning software stack to advance artificial intelligence (AI) development.
"As a partner of NVIDIA Inception, we rely on NVIDIA to offer innovative solutions to the AI market, today, its solutions evolve to better support an important phase of the development of artificial intelligence: experimentation", added Jeremy Antoniuk, COO of Deep Cognition. "With TITAN V and NGC, the Deep Learning Studio of Deep Cognition has developed a powerful desktop solution for the development, iteration and training of artificial intelligence.
Since TITAN V is designed for use in environments research, do not expect to see an "enthusiastic class" price on the card, however, the TITAN V is priced at $ 2,999 and is available right now at the NVIDIA store.