The skinny on NVIDIA’s latest and greatest GPU based on the GK110 GPU, previously seen in Tesla supercomputers.
The skinny on NVIDIA’s latest and greatest GPU based on the GK110 GPU, previously seen in Tesla supercomputers.
Titan:- a person or thing of enormous size, strength, power, influence, etc.: a titan of industry.
Back in the mists of time, or at least the halcyon days of 2011, the first whispers of NVIDIA‘s Kelper GPUs were heard. Rumour told of an enormous beast with thousands of shaders, wide memory bus and lashing of onboard memory. As speculation turned to reality the vision dimmed – the first GeForce of this series was a smaller, more efficient affair.
But then the rumours picked up once again. Tesla, NVIDIA’s server-platform compute line was to feature a huge Kepler GPU at its core, the fabled GK110. Then: despair – surely these professional parts would never make it into the mainstream. Today, we can tell you that they will.
GeForce GTX Titan
The GK110 makes its first desktop outing in the the guise of the GeForce GTX Titan, a dual-slot single-GPU card designed for desktop systems, evidenced by the modest power requirements of 8+6 Pin PCI-E power. Compared to the GK104 – NVIDIA’s previous flagship GPU seen on the GTX 680 – Titan features 57% more CUDA cores and a wider 384bit memory bus, back up once again by GDDR5 memory clocked at 6008MHz. The total transistor count of 7.1 billion puts it at pretty much twice the size of the GK104 and almost three times the size of Intel’s Core i7 3960X CPU.
One might expect that the size of the GPU would necessitate very modest clock speeds, but although the clocks for the Titan are below that of the typical GHz+ of the GTX680 they’re not low by any means. Titan reference clocks are pegged at 837MHz, which are then raised to 876MHz via GPU Boost 2.
GPU Boost 2
The first Kepler GPUs brought with it GPU Boost: an on-board dynamic overclocking system which pushed the speed of the CUDA cores above the default based on available power. With Titan, NVIDIA have implemented the second iteration which they’re calling GPU Boost 2, targeting a higher overclock threshold and improved utilisation thanks to better diagnostic tools and targeting of specific temperatures.
The improved process allows better balancing of performance and acoustics, with greater value now being apportioned to efficient aftermarket or other 3rd party cooling solutions such as those seen from MSI and GIGABYTE this past year. Extreme overclockers can also ‘Over-Volt’ the GPU on cards released by supported AICs, although NVIDIA are clear that running the GPU at high voltages for extended periods of time can reduce longevity.
Cooling and Acoustics
The GK110 is a far larger GPU than the the GK104, with an implication that it will have a higher TDP and therefore greater cooling requirements. In accordance with this the reference cooling design of the GeForce GTX Titan has been augmented considerably versus that of the GTX 680 by extending the heat dissipating fins further towards the exhaust. We wouldn’t expect the fan quality to diminish, so the overall cooling package should be both cool and very quiet as reference designs go. Proper utilisation of fan profiles in conjunction with GPU-Boost should serve to improve the experience still further.
Position in the Marketplace
One of the first questions that will come up is whether Titan is replacing the current flagship card – the dual-GPU GTX 690. Word from NVIDIA is that Titan will be sold alongside all the current GPU range, targeting those using especially high-resolution monitor setups which would take advantage of the cards wide bus and 6GB video memory. The GTX 690 will be realigned (in marketing) to be tailored for very high image quality on single-monitor setups up-to 1900*1200, including 3D.
As you would expect from NVIDIA discrete GPUs Titan is compatible with SLI – in this case 3-way rather than the 4-way capability of the GTX 680. NVIDIA pitch that extreme solution as the only option for Crysis 3 maxed out in surround-gaming, although you should expect to pay a pretty penny for the privileged.
Compute Capability
NVIDIA’s GTX 680 was a departure from the Fermi-based 480 and 580 in that double-precision performance was hobbled by comparison. Titan rectifies this by enabling far greater double-precision performance across CUDA cores, at the expense of overall clock speed. This should benefit those who utilise GPU-accelerated compute applications greatly, those you will need to ensure that the setting is enabled/disabled appropriately in the driver suite beforehand.
General Information
At this time we cannot be certain about the GTX Titan release quantities, but as it’s based on a server part we wouldn’t expect them to be plentiful. Price is similarly in the high-end range: at an MSRP of £837 it’s considerably more than a single GTX 690, never mind a GTX680. This is unsurprising given that those with very high resolution setups are the target market – after all, multiple displays of any size and quality don’t come cheap.
Modest power requirements of 600W from a quality PSU and an only 10.5 inch length on the reference design also make it appropriate for small form factor PCs, unlocking great gaming potential in a small and quiet package.
Titan is due for release, including benchmarks, soon. You may indeed see it up for pre-order when this article goes live. If you’re contemplating the card we would suggest you carefully consider how appropriate the card is to your setup, planned or current.
Fore more detailed information direct from NVIDIA visit http://www.nvidia.com/titan, and be sure to check out the product launch trailer:
