The amplitude of this effect will vary from generation to generation, but it’s fair to say that the NVIDIA Pascal GPU architecture does this exceptionally well, and effectively creates a “discontinuity,” a term that is much-loved at NVIDIA.
Raw speed and horsepower is typically a good place to start to judge how good a new graphics processor is. These chips are the only kind of processor whose performance still grow at an incredible speed.
According to numbers released by NVIDIA, the GeForce GTX 1080 can outperform its GTX 980 predecessor by a factor of 1.6X to 1.8X in recent games. But it’s in VR that performance gets the biggest boost, with close to a 2.5X advantage.
A single GeForce GTX 1080 is faster than two GTX 980 cards linked via SLI, and it even manages to out-perform the current NVIDIA Titan X, a $1000+ graphics card reserved for a small elite. Watch Doom running at blazing speeds with the GTX 1080:
Relative to its power consumption, the performance is also impressive. The Pascal architecture is much more efficient for every Watt of power it uses. You may remember that the Maxwell architecture of the GeForce 9xx Series was already a good step up in compute density and power efficiency, but NVIDIA has managed to increase this by another 2X-3X, which is extremely impressive.
In fact, the GeForce GTX 1080 is so efficient that it requires a single power connector and will draw at most 180W of power.
Because of this efficiency, the GTX 1080 operating frequency can be pushed even further without overheating. The default 1733MHz frequency was pushed to 2114MHz during the NVIDIA launch event, and it was done using the out of the box air-cooled heatsink (at a 67C temperature). Previously, such speeds could only be obtained using complex liquid-gas apparatuses that would not be useful to most people.
This new GPU features 2560 NVIDIA compute cores (at 1607Mhz) versus 2048 for the GTX 980 (at 1126Mhz). With each core being more efficient, the specs suggest a total 9 TeraFLOPS of computing power for the GTX 1080, versus 5 TFLOPS for the GTX 980.
Previously, NVIDIA was using a 28nm semiconductor process. Also, the 16nm FinFET brings tremendous benefits in terms of power-efficiency and leakage control.
The GDDR5X graphics memory is another critical performance factor. It can move data at a speed of 10Gbps with a maximum of 8GB of memory available to the GPU. The GTX 980’s memory would max out at 7Gbps.
Critics may point out that memory such as HBM can go even faster, but they are however limited to 4GB at the moment, so it’s a matter of tradeoffs.
Backed by hardware support, the new 1000-Series GeForce cards also have a few (mostly) software tricks up their sleeves.
NVIDIA has announced Ansel, a very cool virtual camera that can be triggered inside a 3D app. When it happens, the app is frozen, and the user can move the camera around, and capture ultra-high resolution shots in 2D, 360-degrees, or VR to share later. I’ve talked about Ansel during the launch event.
The Simultaneous Multi-Display technology lets NVIDIA project a 3D scene properly on up to 16 different viewports/monitors. This is great for multi-display gaming since the projected image will not exhibit perspective distortions, making the immersion that much better.
For VR applications, it’s possible to use those multiple projections to match the physical curvature of VR headset lenses. Simultaneous Multi-Display is a geometric performance optimizer which lets the 3D engine project the scene 16 times without doing all the geometric transformations again. We’ve described Simultaneous Multi-Display in more details previously.
With a price of $600 for the GeForce GTX 1080, NVIDIA remains in the usual high-end range for consumer cards. However, with performance exceeding that of a Titan X, it is an extraordinary achievement and an equally extraordinary value to the end-user.
Even better, at $380 (June 10 2016), it’s possible to get the GeForce GTX 1070 which is also faster than a Titan X. Now, it is just nutty.
The NVIDIA “founders” edition cards will be a little more expensive at $700 for the 1080GTX (May 27). Those cards are designed, manufactured and sold by NVIDIA directly. They tend to look nicer than other OEMs’ cards, and the build quality is impeccable.