Review: Nvidia GeForce GTX 780 Ti

Introduction

Nvidia's high-end graphics cards have held a seemingly unassailable performance lead for most of this year. The GeForce GTX Titan came in swinging in February and smashed all before it. Nvidia followed this up with the slightly slower GeForce GTX 780 in May, and the green team had the premium market all to itself until a few weeks ago. Due to an honest lack of competition, this performance monopoly has meant the return of the £500-plus cards.

AMD Radeon R9 290X and R9 290 have finally arrived on the scene and matched, or even exceeded, Nvidia's performance, particularly when gaming at high resolutions, with the added benefit of being cheaper than the comparison GeForces. Now, in a two-pronged attack at wrestling back the hearts, minds and wallets of the enthusiast, Nvidia has taken an axe to GeForce GTX 780 and GTX 770 pricing, making them somewhat comparable with AMD's offerings, and promised a new benchmarking-topping card in the form of the $699 GeForce GTX 780 Ti.

Today, we reveal just how Nvidia intends to regain or reinforce - depending upon how you construe current benchmark results - its position as the provider of the best gaming graphics cards on the market.

Nvidia has three paths that it can take to advance the performance of the current GeForce GTX 780: it can increase the core and memory frequencies - this works really rather well with partner cards - boost the number of shader-units, or do both. After having seen the performance of the Radeon R9 290X, which is roughly analogous to the GTX Titan, Nvidia has decided to take that third path. Let's show you how with the usual Table of Doom.

Architecture analysis - the full-fat GeForce GTX 780 Ti

You'll likely know the GTX 780 uses an incomplete GK110 die. The full-fat implementation calls for 15 SMX units; Titan uses 14 and GTX 780 12, with each containing 192 shaders and 16 texture units. GTX 780 Ti, meanwhile, increases this to the maximum 15 supported by the Kepler architecture. Yup, this is Nvidia going for an all-out performance assault on the nascent Radeons.

Nvidia also boosts the GPU Boost speed to a supported 928MHz, though, going by previous experiences, the GPU may well run faster when gaming. Crunching the official numbers shows that GTX 780 Ti has, via architecture and frequency, 19 per cent more shading power than said Titan. Keeping all this extra shading power requires an upgrade to the backend of the GPU, so Nvidia increases the memory speed from 6,000MHz to 7,000MHz - an uptick of almost 17 per cent.

Though it's clear that GeForce GTX 780 Ti will be a fair bit faster than Titan in games, by dint of its processing and memory advantages, do note that it ships with a 3GB framebuffer, implemented to keep costs in check and Nvidia's profit margins high. GTX Titan remains the premier card in our minds, equipped with a 6GB buffer and massively higher double-precision throughput, but it is also $300 dearer.

Expressed in pure gaming terms, the GeForce GTX 780 Ti is the fastest card ever built by Nvidia. It's highly unlikely that Nvidia will release a faster reference GPU based on the present Kepler architecture, for it has used all available architecture resources in constructing the GTX 780 Ti.

The question of GeForce hegemony is moot; the one that matters is whether this GeForce, representing the pinnacle of Kepler GPU engineering, has enough chutzpah to overcome the AMD Radeon R9 290X - which, let's not forget, is $150 cheaper.

As an aside, though the card looks identical to the GeForce GTX 780 and Titan, Nvidia has reworked the power-delivery system such that the GPU can dynamically pull power from the 6-pin, 8-pin, or PCIe lanes when overclocking, using the trio of sources to balance the juice coming in - a feature not present on previous Nvidia cards. Perhaps more importantly, this means that the GTX 780 Ti uses a slightly different PCB to the other high-end cards, thus stopping add-in card partners from launching their own versions quickly. Nvidia expects partners to have custom cards in roughly a month's time.

Vagaries of board design notwithstanding, let's get on over to them there benchmarks.