We'll begin with a series of synthetic tests aimed at exposing the true, delivered throughput of the GPUs. In each instance, we've included a table with the relevant theoretical rates for each solution, for reference.
| Peak bilinear
| Peak bilinear
|GeForce GTX 680||34||135||135||192|
|GeForce GTX Titan||42||196||196||288|
|GeForce GTX 690||65||261||261||385|
|Radeon HD 7970 GHz||34||134||67||288|
|Radeon HD 7970 GHz CrossFire||67||269||134||576|
In the past, performance in this color fill rate test has been almost entirely determined by memory bandwidth limitations. Today, I'm not so sure. None of the solutions achieve anything like their peak theoretical rates, but I think that has to do with how many color layers this test writes. The Titan's certainly not 50% faster than the GTX 680, despite having literally 50% higher memory bandwidth.
The multi-GPU solutions, meanwhile, scale up nicely in these synthetic tests.
Among the single-GPU solutions, the Titan is very much in a class by itself here.
Tessellation and geometry throughput
|GeForce GTX 680||4.2||192|
|GeForce GTX Titan||4.4||288|
|GeForce GTX 690||8.2||385|
|Radeon HD 7970 GHz||2.1||288|
|Radeon HD 7970 GHz CrossFire||4.2||576|
The last couple generations of Nvidia GPU architectures have been able to sustain much higher levels of polygon rasterization and throughput than the competing Radeons. The big Kepler continues that tradition, easily leading the GeForce GTX 680, even though its peak theoretical rasterization rate isn't much higher. Most likely, the GK110's larger L2 cache and higher memory bandwidth should be credited for that outcome.
|GeForce GTX 680||3.3||192|
|GeForce GTX Titan||4.7||288|
|GeForce GTX 690||6.5||385|
|Radeon HD 7970 GHz||4.3||288|
|Radeon HD 7970 GHz CrossFire||8.6||576|
These tests of shader performance are all over the map, because the different workloads have different requirements. In three of them—ShaderToyMark, POM, and Perlin noise—the Titan and the Radeon HD 7970 GHz are closely matched, with the 7970 taking the lead outright in ShaderToyMark. That's a nice reminder that the 7970 shares the same peak memory bandwidth and is within shouting distance in terms of peak shader throughput.
I had hoped to test OpenCL performance using at least one familiar test, LuxMark, but it crashed when I tried it on the Titan due to an apparent driver error. Anyhow, I think the GPU computing performance of these chips deserves a separate article. We'll try to make that happen soon. For now, we're going to focus on the Titan's primary mission: gaming.