Updated, August 5: The GTX Titan X, powered by the GP102 Pascal GPU, is now available to buy in the UK, Europe, and US. The Titan X is currently only available through Nvidia's own online store, priced at £1,100 in the UK and €1,300 in the Eurozone. At the time of writing, the Titan X is in stock in the UK. Americans can pick a new Titan X up for a relatively cheap $1,200, though it appears to be out of stock currently.
Original story
Forget the GTX 1080: there's a new slab of graphics card hotness on the way from Nvidia, and its name is, er, the GTX Titan X. Yes, Nvidia has taken its most expensive graphics card and given it a Pascal-architecture makeover. $1200—UK price TBC, but probably £1,100—buys you 11 teraflops of FP32 performance, which is a significant 24 percent jump over the 8.9 teraflops of the GTX 1080, and just over 60 percent higher than the 6.6 teraflops of the original Titan X.
The new Titan X launches on August 2 in the US and Europe. At first it'll only be available from the Nvidia website, but it will percolate down to other retailers soon after.
The Titan X is powered by a new chip, GP102, which packs in 3584 CUDA cores. While Nvidia hasn't revealed the amount of Streaming Multiprocessors (SMs), texture units, and the like, if the company uses a similar architecture to the GP104 chip (as used in the GTX 1080 and GTX 1070), expect a 40 percent boost in SMs over the GTX 1080 to 28. The chip runs at a 1417MHz base clock and 1531MHz boost clock.
Backing up GP102 is 12GB of GDDR5X memory running at an effective 10GHz and attached to a wide 382-bit bus, resulting in a 480GB/s of memory bandwidth, or a 50 percent increase over the GTX 1080. Power is provided by one 8-pin and one 6-pin PCIe connector, with a max TDP of 250W, the same as the previous Titan X. Connectivity consists of DisplayPort 1.4, HDMI 2.0b, and dual-link DVI. Externally, the Titan X uses the same multifaceted cooler of the GTX 1080, albeit now finished in black.
Interestingly, while the new GTX Titan X features the same number of CUDA cores as Nvidia'sTesla P100 GPU—as used in deep learning and science applications—it is using a different chip. With previous Titans, Nvidia has simply taken its biggest chip and given it a prosumer makeover, but that doesn't seem to be the case here. Those hoping for a return of the stellar FP64 performance of the original Titan may be disappointed here, although Nvidia is pushing the card's 44 TOPs of INT8, a measurement for neural network inference performance.
It's also a wee bit disappointing not to see HBM2 being used and Nvidia opting for GDDR5X. Yes, HBM2 is still in its infancy (the P100 is the only GPU that uses it) and costs are high. But given this is Nvidia's flagship GPU, and it costs $1200, its inclusion would have been a pleasant surprise. It would also have given AMD a poke, especially since the company is planning to use HBM2 in its upcoming Vega architecture.
The reveal of the Titan X follows the launch of the GTX 1080, the GTX 1070, and the GTX 1060, all of which were released over the past two months. That's a fierce release schedule by anyone's standards. Quite why Nvidia is so keen to get the Titan X out of the door, particularly when it has zero competition at the top end of the market, is something of a mystery. Still, with 11 teraflops, the Titan X might finally be the card that gives us sweet 60FPS 4K gaming.
Source: ars Technica
No comments:
Post a Comment