NVIDIA launched the GTX Titan X to much acclaim in March, and it was immediately hailed as the greatest GPU to date. Something of a halo product - and with a price to match - its monster three thousand shader GM200 core and 12GB frame buffer gave the impression that it was designed for two uses: hardcore single precision compute and high resolution scene rendering.
The consumer space has been relatively quiet by comparison. NVIDIA's GeForce GTX 980 continues to cast a heavy shadow over the performance end of the market, but even the flagship GeForce card has a long way to go before it can hit the performance heights of the Titan X. Given both the yawning price and performance gap between the GTX 980 and Titan X it was inevitable that the next GeForce GPU would serve to dominate this end of the market. Today NVIDIA are launching that card - the GeForce GTX 980 Ti.
Based on the best Maxwell has to offer.
The GeForce GTX 980 Ti is a card with a lot in common with the premier GTX Titan X, but tailored to more closely match the demands of performance gamers who perhaps don't have £900 to spend on a new GPU. Rather than a supped-up version of the GTX 980 with higher clock speeds and the like, it's a slightly slim-line Titan X designed to give consumers an excellent 4K experience whilst pushing the performance/watt enveloper beyond what was possible with the Kepler-based GeForce flagships.
Featuring the same GM200 core as the Titan X, the GTX 980 Ti has two Streaming Multiprocessors (SMMs) disabled of the twenty-four on the die, thus leveraging a total of 2816. This compares favourably with the 2048 available on the GTX 980, yet at 3072 the Titan X still rules the roost. The GTX 980 Ti also has commensurately more texture units and ROPs than the GTX 980, 176 and 96 respectively. At 1000MHz base/1076 MHz boost, core frequencies don't differ significantly from the Titan X; all things being equal you might expect raw performance to be almost 90% of its bigger sibling at a considerably lower price. Not bad.
Of course, the next part of the equation is the memory subsystem, an area where NVIDIA has come in for some flack in recent months. The GTX 970 was listed as a 4GB VRAM, but their modification to the GM104 core meant that 512MB of that frame buffer ran considerably slower than the rest. Nonetheless it was still an impressive card, and served as a salient lesson in being up front about hardware specs.
The GTX 980 Ti incorporates a 6GB frame buffer over the same 384-bit-wide memory bus as the Titan X. Clocked at 3500MHz (GDDR5) it has an overall data rate of 7Gbps and a gargantuan total memory bandwidth of 336 GB/s, and that's before accounting for the Third Gen Delta Colour Compression algorithm implemented in NVIDIA's Maxwell architecture.
A 6GB GDDR5 frame buffer is 50% larger than the GTX 980 and most variants of AMD's R9 290X, and more per core than AMD's flagship R9 295X2. The reason for this is simple - it's unlikely that those spending more than £500 on a single GPU will still game at 1080p, and even at 1440p some titles (most notably GTA V) are utilising more than 4GB of video memory. To push playable frame rates and high quality settings at 4K resolutions you need to marry a high performance GPU with a large pool of high-performance memory, something that the GTX 980 Ti does in spades.
Taking the GTX 680 as a baseline NVIDIA project that the GTX 980 Ti scales to 4K roughly three times as well, almost twice as well as the GTX 780 Ti. For those gaming at higher resolutions on older cards, that alone seems like compelling reason to give the GTX 980 a look. In the real world that translates to playable frame rates in the three most modern games - the Witcher 3, GTA 5 and Project Cars, which isn't quite possible with the older GTX 680 NVIDIA expect you might replace with this card.
On The Horizon - supporting DirectX 12 Feature Level 12_1
The next big thing, at least in terms of software, will be the release of Windows 10 this summer. With Windows 10 will come DirectX 12, an updated graphics API which includes a vast array of performance innovations; but it wouldn't be a DirectX revision without confusing granulatory.
Baseline DirectX12 support is Direct3D 12 API support. Encompassing a low overhead instruction set, greater control over graphics resources and asynchronous compute support, this is the core of what makes DirectX12 such a promising development. On NVIDIA's side all GPUs from the 400-series onwards (i.e. Fermi, Kepler and Maxwell-based hardware) will support the Direct3D 12 API on Windows 10 through driver updates.
DirectX 12 Feature Level 12_0 is the next step up in DirectX 12 support. GPUs which conform to this standard are capable of Tiled Resources (loading only relevant portions of a texture into memory, rather than the entire texture), Bindless Textures and Typed UAV Access; all of these aim to reduce CPU overhead and cut some of the sharp edges off techniques introduced in previous versions of DirectX. Kepler should and Maxwell will support at least Feature Level 12_0.
Feature Level 12_1 is where things get a little smart. At this level Microsoft are introducing two new technologies to DirectX12's feature set, whilst NVIDIA are going a stage further in their DirectX 12 support for Maxwell.
Rasterization is the means by which a vector shape (i.e. triangle or other polygon) is converted into an array of square pixels. Traditional rasterisation techniques place sample points in the middle of pixels, and then define a triangles' coverage pixels based on it covering sample points. By its very nature this tends to be quite imprecise and favouring speed, and often results in coverage problems at triangle edges.
A Conservative Raster algorithm, key to DirectX 12 Feature Set 12_1, defines a pixel as covered if any part of the pixel is covered by the triangle. Employing hardware resources allows developers to create more realistic effects, perhaps most notably with ray-traced shadows as in the example below.
As you can can see, Ray-traced shadows using the Conservative Raster show none of the aliasing typical to previous ray-traced shadows calculated without this technique. In essence this means it's possible to have pixel-perfect shadows which are less computationally expensive than other implementations.
Raster Ordered Views
A less well heralded Key Feature of Feature Set 12_1 is Raster Ordered Views (ROVs), and again will be part of Maxwell's core feature DirectX12 set. ROVs give developers control over the order in which shader operations are performed, which in tern allows for more flexibility in the design and implementation of blend and transparency algorithms.
Volume Tiled Resources
VTR is a new methodology designed for complex visual effects such as fluid dynamics or smoke effects which are volumetric in their nature. These effects require important additional information parameters such as temperature or density, but typically are stored and stream separately. VTR incorporates this information directly into the tiled texture as a third dimension (hence the 'Volume' Tiled Resource), but the information can continue to be streamed selectively such that only relevant tiles are ever loaded. This reduces both memory and CPU overhead.
Implementing VTR allows users to create more complex and previous computationally expensive visual effects into games such as Sparse Fluid Simulation, which is a technique often used in the creation of more realistic smoke effects. Not strictly part of DX12 Feature Set 12_1, it's nonetheless supported by NVIDIA's Maxwell architecture.
Back To The Hardware
The GTX 980 Ti sees the return of the excellent reference cooler seen on the Titan X and earlier flagship cards. Combining a vapour-chamber and 'blower' rear thermal exhaust for hot air, it has in the past proved to be more than adequate in keeping NVIDIA's high-end GPU under control. It will need to be on top form as the GTX 980 Ti features a 250W TDP, far higher than the 980 and on a par with it's bigger brother, the Titan X.
PSU power requirements are commensurately high for a modern GPU. NVIDIA recommend a 600W PSU (ensure that it's a decent quality too of course) with one 6-pin and one 8-pin PCI-Express power connector, the same as the GTX 780 Ti but more than (for instance) a GTX 680. If the Titan X is any indication overclocking the GPU is expected to be particularly demanding on the PSU. Be sensible with your component choices and budget accordingly.
The reference design features three DisplayPort 1.2a connectors, one HDMI and one DVI connector; naturally the GTX 980 Ti is compatible with NVIDIA's G-SYNC frame sync technology with a suitable G-SYNC monitor. Up to 4-way SLI is supported for multi-GPU configurations.
Speaking of budget, NVIDIA's MSRP for the GTX 980 Ti will now occupy the top end of NVIDIA GeForce product stack, displacing the GTX 980 at $649, and will be bundled with a code for Batman: Arkham Knight. To compensate the GTX 980 is moving to an MSRP of $499, whilst the GTX 970 and 960 will remain at $329 and $199 respectively. In the UK the GTX 980 Ti is priced at £539.99 (inc. V.A.T.); €605,99 (excl V.A.T) is the price on the continent, where V.A.T. varies. Cards should be available as you read this, whilst partner variants will be on sale in the coming weeks.
Rich's review of the GTX 980 Ti will be our definitive take on NVIDIA's latest GeForce flagship. For now, we'll leave you with a final look at the card in all its glory.