Single page Print

Sizing 'em up
Do the math involving the clock speeds and per-clock potency of the GM204 cards, and you'll end up with a comparative table that looks something like this:

Peak pixel
fill rate
(Gpixels/s)
Peak
bilinear
filtering
int8/fp16
(Gtexels/s)
Peak
shader
arithmetic
rate
(tflops)
Peak
rasterization
rate
(Gtris/s)
Memory
bandwidth
(GB/s)
Radeon R9 285 29 103/51 3.3 3.7 176
Radeon R9 280X 32 128/64 4.1 2.0 288
Radeon R9 290 61 152/76 4.8 3.8 320
Radeon R9 290X 64 176/88 5.6 4.0 320
GeForce GTX 770 35 139/139 3.3 4.3 224
GeForce GTX 780 43 173/173 4.2 4.5 288
GeForce GTX 780 Ti 45 223/223 5.3 4.6 336
GeForce GTX 970 75 123/123 3.9 4.7 224
Asus Strix GTX 970 80 130/130 4.2 5.0 224
GeForce GTX 980 78 156/156 5.0 4.9 224

The rates above aren't destiny, but they do tend to be a pretty good indicator of how a given GPU will perform. Since the GM204 can run at higher clock speeds than the GK110, the GeForce GTX 980 is able to give even the mighty GTX 780 Ti a run for its money in terms of shader arithmetic—with a peak rate of five teraflops—and rasterization. The 980 trails a bit in the texture filtering department, but look at that pixel fill rate. Nothing we've seen before comes all that close.

Contrast that prowess to the GTX 980's relatively modest memory bandwidth, which is no higher than the prior-gen GTX 770's, and you might ask some questions about how this new balance of resources is supposed to work. The answer, it turns out, is similar to what we saw with AMD's Tonga GPU a couple of weeks back.

Nvidia Senior VP of Hardware Engineering Jonah Alben revealed in a press briefing that Maxwell makes more effective use of its memory bandwidth by compressing rendered frames with a form of delta-based compression. (That is, checking to see whether a pixel's color has changed from a neighboring pixel and perhaps only storing information about the amount of change.) In fact, Alben told us Nvidia GPUs have used delta-based compression since the Fermi generation. Maxwell's compression is the third iteration. The combination of better compression and more effective caching allows Maxwell to reduce memory bandwidth use substantially compared to Kepler—from 17% to 29% in workloads based on popular games, according to Alben.

So what happens when we try 3DMark Vantage's color fill test, which is limited by pixel fill rate and memory bandwidth, on the GTX 980?

Yeah, that works pretty darned well. The GTX 980 paints over twice as many pixels in this test as the GK104-based GTX 770, even though the two cards have the same 224 GB/s of memory bandwidth.

On paper, the GTX 980's other big weakness looks to be texturing capacity, and in practice, the 980 samples textures at a lower rate than its competition. The GTX 970 even falls slightly behind the GTX 770 in this synthetic test, just as it does on paper. We'll have to see how much of a limitation this weakness turns out to be in real games.

The GM204 cards have some of the highest rasterization rates in the table above, and they make good on that promise in these tests of tessellation and particle manipulation. The GTX 980 sets new highs in both cases.

In theory, the GeForce GTX 780 Ti has more flops on tap and higher memory bandwidth than the GTX 980, so it should perform best in these synthetic tests of shader performance. In reality, though, Maxwell delivers on more of its potential. Even with a big memory bandwidth handicap, the GTX 980 outperforms the GTX 780 Ti in both benchmarks. Only AMD's big Hawaii in the Radeon R9 290X is more potent—and not by a huge margin.