Thursday, December 7th 2017
NVIDIA Announces TITAN V "Volta" Graphics Card
NVIDIA in a shock move, announced its new flagship graphics card, the TITAN V. This card implements the "Volta" GV100 graphics processor, the same one which drives the company's Tesla V100 HPC accelerator. The GV100 is a multi-chip module, with the GPU die and three HBM2 memory stacks sharing a package. The card features 12 GB of HBM2 memory across a 3072-bit wide memory interface. The GPU die has been built on the 12 nm FinFET+ process by TSMC. NVIDIA TITAN V maxes out the GV100 silicon, if not its memory interface, featuring a whopping 5,120 CUDA cores, 640 Tensor cores (specialized units that accelerate neural-net building/training). The CUDA cores are spread across 80 streaming multiprocessors (64 CUDA cores per SM), spread across 6 graphics processing clusters (GPCs). The TMU count is 320.
The GPU core is clocked at 1200 MHz, with a GPU Boost frequency of 1455 MHz, and an HBM2 memory clock of 850 MHz, translating into 652.8 GB/s memory bandwidth (1.70 Gbps stacks). The card draws power from a combination of 6-pin and 8-pin PCIe power connectors. Display outputs include three DP and one HDMI connectors. With a wallet-scorching price of USD $2,999, and available exclusively through NVIDIA store, the TITAN V is evidence that with Intel deciding to sell client-segment processors for $2,000, it was a matter of time before GPU makers seek out that price-band. At $3k, the GV100's margins are probably more than made up for.
The GPU core is clocked at 1200 MHz, with a GPU Boost frequency of 1455 MHz, and an HBM2 memory clock of 850 MHz, translating into 652.8 GB/s memory bandwidth (1.70 Gbps stacks). The card draws power from a combination of 6-pin and 8-pin PCIe power connectors. Display outputs include three DP and one HDMI connectors. With a wallet-scorching price of USD $2,999, and available exclusively through NVIDIA store, the TITAN V is evidence that with Intel deciding to sell client-segment processors for $2,000, it was a matter of time before GPU makers seek out that price-band. At $3k, the GV100's margins are probably more than made up for.
135 Comments on NVIDIA Announces TITAN V "Volta" Graphics Card
Thats the strategy my friend. Offer something at 10000 euro, than something similar to it at 3000. 3000 will then seems like a bargain. Buts ist not, ist 3000 ya, know.
how much does it cost nvidia to put together the Titan V ? No more than 500 EUR i believe. They should seel it at 1000 euro tops, 700 would be a much better Price.
But they want to cash the fk up ! So they put it first at 8000, make it scarce and anavailable, preparing the sheep psychologicaly into thinking that when the 3000 Card will come, it is cheap as fk, so you "Need" to buy it, because so cheap at 3100 euros ist a damn miracle, that what it is. Damn, you won the lottery, you can buy the Card at 3100 Euros, ist so cheap, who wouldnt want to buy such a Card.
Unlike the software argument of previous cards, the Volta architecture is AI focussed. I'll be very surprised if the attainable gaming range has the Tensor cores.
Then again, if Nvidia can make them work for gaming, hell, bite me backwards. The new GeForce AI GTX 2080ti - it knows where to render before the game does, shoot opponents before they've even pressed the 'w' key.
Sign me up.
I generally don't care about colours, so I'm not going to start giving a rat's ass now.
If you ask me this card at $3K has more sense than a $1200 TXp which is 5% faster than 1080Ti at $500 more, with exactly the same GP102 core and only 1GB more vram. This is to V100 what 1080Ti is to TXp.
DOUBLE-PRECISION
7 TeraFLOPS
SINGLE-PRECISION
14 TeraFLOPS
DEEP LEARNING
112 TeraFLOPS
Since the Titan V is 110TFlops using the tensor cores I would guess 13.75 for Single precision. (x8)
For reference,
1080ti is 11.3 sp, Titan Xp is 12.1 sp
Vega64 is 12.66-13.7 Single and 25.3-27.5Tflops half precision/packed math (not custom tensor cores, can be used as general compute)
P100 can operate at half precision for the same doubling in performance but nothing else in the pascal family can, I am assuming the V100 can as well for non tensor tasks as well as this new Titan Volta.
www.servethehome.com/nvidia-titan-v-extracting-value-deep-learning-enhancements/
www.nvidia.com/en-us/titan/titan-v/
But well if it clocks as high as pascal can, it's 2*2.1GHz*64*80=21,5TFlops of fp32 max. Though good luck for keeping that cool :D