GeForce GTX 780 Ti: Nvidia strikes back, retaking the performance crown from AMD

The past few weeks haven’t been kind to Nvidia. AMD’s 290 and 290X products are incredibly aggressive, both in terms of clock speed and their overall performance. As a result, Nvidia, which began the fall comfortably in charge of the upper GPU product stack, has been rocked back hard on its heels. The company responded to the R9 290X with a sharp price cut to the GTX 780, but even that slash left the card in trouble, neatly pinned between the R9 290 and the R9 290X. Clearly, Nvidia needed a bigger gun.

Fortunately, they had one on tap. Meet the GTX 780 Ti.

GK110: Fully armed and operational

We’ve been over Nvidia’s GK110 GPU multiple times in the past year, ever since it debuted as a new, workstation-class product meant for the scientific computing and professional GPU space. The chip is a monster — both in terms of gaming performance and die size. At 551 square millimeters and 7.1 billion transistors, this is the largest, most complex single GPU Nvidia has ever built. And up until now, all the shipping implementations of GK110 — including the GTX Titan and GTX 780 — have been handicapped. While the core contains 2,880 cores total, Nvidia has always disabled at least one of the compute clusters (SMX in Nvidia lingo). Thus the GTX Titan had 2,688 GPU cores, while the GTX 780 had just 2,304.

The GeForce GTX 780 Ti, on the other hand, has all 15 SMXs enabled and all cores operational. Its texturing resources, already significant, have increased further, up to 240 TMUs (from 224 on the GTX Titan). ROP count, however, remains fixed at 48. As such, its pixel fillrate is up only slightly from the Titan’s 40.2GPixels/s, to 44.5GPixels/s. Memory bandwidth is also up somewhat, to 336GB/s, from 288GB/s on the Titan and GTX 780. Unlike AMD, which uses a relatively pedestrian 1250MHz memory clock but compensates with a 512-bit RAM bus, Nvidia has opted for a narrower 384-bit bus but a much higher effective clock speed, at 1750MHz.

The GTX 780 Ti, however, isn’t a complete upgrade from the Titan. It lacks two features that characterized that GPU — its 6GB RAM buffer and its double-precision performance. Of the two, the lack of support for double-precision floating point is unsurprising — Nvidia’s decision to include full-speed DP in Titan turned that GPU into a steal for buyers who might want it, but didn’t want to shuck out full price for a K20X workstation cards. How many GPUs were actually sold for this purpose, however, remains unclear. Not many, in any case.

The 3GB RAM limit is potentially more of an issue — something we’ll try to look at in some detail here, despite being officially limited to a 1920×1080 monitor. We’ll discuss that a little later on.

Next page: 4K gaming, and whether 3GB of RAM is enough

• 1 of 4
• Next