Nvidia CEO Jen-Hsun Huang loves a spectacle. So while you’re reading this sentence, the GPU-manufacturer’s founder is regaling thousands of gamers at a 24-hour gaming event the company is sponsoring in Los Angeles with the announcement of Nvidia’s latest graphics processor, code-named “Big Maxwell.”
Despite Nvidia’s draconian efforts to prevent a recurrence of the leaks that presaged the announcement of its Shield Tablet, details of the new GPUs have been dripping into the Interwebs for several weeks. Here’s the official information: The new GPU will initially be available in two SKUs, the $329 GeForce GTX 970 and the $549 GeForce GTX 980. The GTX 970 has 1664 processor cores running at a base frequency of 1050MHz (boost clock of 1178MHz), while the beefier GTX 980 has 2048 cores running at 1126MHz (boost clock of 1216MHz). The GTX 970 and the GTX 980 both have a 256-bit interface to 4GB of GDDR5 memory running at an effective speed of 7Gbps.
The mounting brackets on cards following Nvidia’s reference design will have one dual-link DVI connector, three DisplayPort 1.2 interfaces, and one HDMI 2.0 port. HDMI 2.0 support is particularly important because it supports 4K resolution at a refresh rate of 60Hz (the older HDMI 1.4b standard is limited to a refresh rate of just 30Hz, which is fine for movies, but terrible for games).
Four new features
In addition to making games run faster—the key promise behind every new GPU—Nvidia announced four new features that promise to make games look more realistic: Dynamic Super Resolution, Multi-Frame-Sampled Anti-aliasing (MFAA), Voxel Global Illumination, and VR Direct. Here’s a brief overview of those technologies.
Dynamic Super Resolution, according to Nvidia, enables the new GPUs to deliver 4K-quality graphics on a 1080p display. The processor effectively renders 4K resolution in the GPU’s frame buffer, and then applies a Gaussian filter to downsample the image to 1080p when it's output to the monitor. This promises to increase visual fidelity without taking a hit in frame rate. The feature is enabled by default in Nvidia’s GeForce Experience utility, and the end user may increase the resolution as high as 5K if desired. Scott Herkelman, general manager of Nvidia’s GeForce business unit, demoed the effect for me at a briefing on Tuesday, and it is impressive.
Aliasing describes the visual distortion that occurs when a GPU renders an image. In the most common examples, non-square images can have jagged edges, and moire can spoil the detail in intricate patterns. Anti-aliasing, obviously, is a means of coping with this phenomenon, and multi-sample anti-aliasing (MSAA) is the most effective technique.
MSAA works by having the GPU sample multiple locations of each pixel in a scene, and then fully rendering and combining those pixels in the image that’s ultimately rendered on the display. In 4X MSAA, four samples of each pixel are taken. For all its effectiveness, MSAA is very computationally expensive and can significantly reduce frame rate when it’s enabled in a game.
Nvidia’s Multi-Frame Anti-aliasing (MFAA) functions in a similar fashion, but it samples pixels over two frames while rotating the pixels. The technique, according to Nvidia, delivers results that are very similar to MSAA without the performance hit that MSAA exacts. Herkelman says MFAA is up to 30 percent faster than MSAA.
Voxel Global Illumination (VXGI) is the third big improvement to be found in the GeForce GTX 970 and -980. Briefly, it’s a new lighting technique that promises to deliver much better image quality without requiring a significant performance hit. As Herkelman explained it, a voxel (a portmanteau for volumetric pixel) stores information about the light in a rendered environment.
Games designed to run on previous-generation GPUs depend on what Herkelman described as “hand-placed lights.” A few light sources in a game environment are placed in fixed locations, so the manner in which the light behaves and its impact on the environment is predetermined in the game’s programming code. That’s part of the reason why only certain objects in a game can be destroyed, the programmer must hard code the manner in which the light and shadows will be rendered on those objects.
VXGI enables programmers to place dynamic light sources in a game’s environment. Objects in the game can be assigned properties that determine how they’ll reflect or absorb the light cast upon them. With this technology, the game programmer can simply place light sources, and those lights will react dynamically to any changes in the environment. In the rendered image of the Apollo moon landing, above, Herkelman changed viewpoints by moving the camera all around the scene, and the lighting changed accordingly. The way objects reflected or absorbed light changed dynamically, and cast shadows moved in relation to the point of view.
Oculus Rift support
Nvidia’s VR Direct technology is intended to benefit upcoming virtual reality headsets such as the Oculus Rift. This bundle of features is designed to reduce latency, allow headsets to take full advantage of Nvidia’s dual-GPU SLI technology, and improve stereo-vision support. Herkelman demoed a new space-combat game for me Tuesday, and it looked leagues better than anything I saw over the course of Nvidia’s 3D Vision initiative.
Like the original Maxwell architecture Nvidia announced in February, these new GPUs adopt power-savings tricks originally developed by Nvidia’s Tegra CPU team. Where a GeForce GTX 680 card with 2GB of memory has a thermal design power (TDP) rating of 195 watts, and a GeForce GTX 780 with 3GB of memory has a TDP of 250 watts, the TDP for the GeForce GTX 980 with 4GB is just 165 watts.
Despite the lower power consumption, Nvidia says the GTX 980 will deliver 5 teraflops of single-precision compute performance, compared to 3- and 4TFLOPs for the GTX 680 and -780, respectively (a teraflop is one trillion floating-point operations per second). The new GTX 970, meanwhile, has a TDP of 145 watts and delivers 4TFLOPs of compute performance. Thanks to the lower TDPs, reference-design cards will have two 6-pin power connectors to supplement the power the cards draw from the PCIe bus.
In the wake of this product launch, Nvidia is discontinuing its GeForce GTX 780Ti, GTX 780, and GTX 770. The suggested retail price for basic boards based on its GeForce GTX 760 has been reduced to $219.
Update: This article was updated to correctly report the video outputs on Nvidia reference-design cards. The source of the incorrect information came from a presentation provided by Nvidia, but I should have noticed that it didn't match the photograph of the card.