Nvidia has the three fastest graphics cards on the market in its Titan X, GeForce GTX 1080, and GeForce GTX 1070, but it isn't waiting for the competition to catch up. A new GeForce GTX 1080 Ti is imminent, based on the same GP102 processor as Titan X, and it's aimed at at the big gap between its $700 $600 (the price drop was a nice little nugget from this evening's announcement) GeForce GTX 1080 Founders Edition card and the $1,200 Titan X.The new 1080 Ti is $700, and it's coming next week.
Expect performance to land closer to the latter than the former, though. All 28 of the chip's SMs are turned on, yielding 3584 CUDA cores and 224 texture units at a rated GPU Boost clock rate of 1.6GHz (that's slightly higher than the Titan X).
The back-end does receive a small haircut: One of its 32-bit memory controllers is disabled, yielding an aggregate 352-bit pathway. This does some funny things to the configurations GeForce GTX 1080 Ti supports, and as a result, the card includes 11GB of memory. Despite the narrower memory path, GeForce GTX 1080 Ti actually offers more theoretical bandwidth than Titan X due to the introduction of 11 Gb/s GDDR5X memory (it moves up to 484 GB/s to the $1200 card's 480 GB/s).
In introducing GeForce GTX 1080 Ti, Nvidia made a point of discussing the challenges tied to pushing memory to 11 Gb/s. By overcoming them, together with Micron, it created opportunities to offer other SKUs with higher-performance memory. Thus, Nvidia said it plans to sell GP104 GPUs to board partners for GTX 1080 cards with 11 Gb/s GDDR5X and GP106 GPUs for GTX 1060s with 9 Gb/s GDDR5.
Cutting one controller naturally takes out a corresponding ROP cluster, leaving 88 ROPs. It trims a 256KB slice of L2 cache, too. What remains is 2,816KB, down from Titan X's 3MB.
| GPU | GeForce GTX 1080 Ti | Titan X (GP102) | GeForce GTX 1080 (GP104) | Titan X (GM100) |
|---|---|---|---|---|
| SMs | 28 | 28 | 20 | 24 |
| CUDA Cores | 3584 | 3584 | 2560 | 3072 |
| Base Clock | ? | 1,417MHz | 1,607MHz | 1,000MHz |
| GPU Boost Clock | 1600 MHz | 1,531MHz | 1,733MHz | 1,075MHz |
| GFLOPs (Base Clock) | ? | 10,157 | 8,228 | 6,144 |
| Texture Units | 224 | 224 | 160 | 192 |
| Texel Fill Rate | 358.4 GT/s | 342.9 GT/s | 277.3 GT/s | 192 GT/s |
| Memory Data Rate | 11 Gb/s | 10 Gb/s | 10 Gb/s | 7 Gb/s |
| Memory Bandwidth | 484 GB/s | 480 GB/s | 320 GB/s | 336.5 GB/s |
| ROPs | 88 ROPs | 96 | 64 | 96 |
| L2 Cache | 2816 KB | 3MB | 2MB | 3MB |
| TDP | 250W | 250W | 180W | 250W |
| Transistors | 12 billion | 12 billion | 7.2 billion | 8 billion |
| Die Size | 471 mm² | 471 mm² | 314 mm² | 601 mm² |
| Process Node | 16nm | 16nm | 16nm | 28nm |
Although we typically praise Nvidia's cooling solutions for exhausting waste heat rather than recirculating it in your case, our thermal testing typically shows these designs underperforming the axial fan-based coolers from board partners, which often facilitate lower temperatures and generate less noise.
Nvidia used this feedback to improve the GeForce GTX 1080 Ti's power supply, making it more efficient, thereby dissipating less heat. What results is a lower operating temperature at a given fan speed, or similar temperatures at lower fan speeds. The benefit comes from doubling up on the GeForce GTX 1080's dualFET configuration for each of the card's seven power phases.
Company representatives also say the 1080 Ti's cooler is better than previous versions. First, the output panel sheds its DVI port, leaving the card with three full-sized DisplayPort connectors and one HDMI interface. For folks who still require DVI connectivity, a dongle will come bundled with the card. The expansion bracket is reportedly redesigned as well, allowing significantly more airflow.
Of course, it remains to be seen how the GeForce GTX 1080 Ti fares in our performance, thermal, acoustic, and power testing. Cards should be arriving soon, so you can count on our usual thorough evaluation in the days to come.
