NVIDIA Geforce GTX 1080

“Twice the performance of a GTX Titan X, and three times as efficienct.” Jen-Hsun Huang’s announcement of the new Geforce GTX 1080 perhaps was a bit over-enthusiastic, and atypically so for the high-powered CEO, but there you have it. NVIDIA claims that they’ve arrived at an unprecedented leap in graphics performance with this generation of GPUs thanks to the Pascal architecture, improvements in memory technology, and a new production process. This is a big jump in GPU performance for sure, but keep in mind that NVIDIA’s claims are valid when looking at what the GTX 1080 can do well that the Titan X can’t. Let’s dive into the announcement of the cards for now, and catch up on the rest of NVIDIA’s announcements later.

NVIDIA Geforce Pascal hardware comparison

GTX 1080 GTX 1070 GTX 980 Ti GTX 980
 GPU family name  Pascal Pascal Maxwell Maxwell
 CUDA Core count  2560 2816 2048
 Single-precision throughput 9.0 TFLOPS 6.5 TFLOPS 5.6 TFLOPS 4.3 TFLOPS
 Base clock 1607 MHz 1000 MHz 1127 MHz
 Boost clock 1733 MHz 1076 MHz 1216 MHz
 Memory bus width 256-bit 256-bit 384-bit 256-bit
 Memory bandwidth 320 GB/s 337 GB/s 224 GB/s
 Outputs Displayport 1.4, HDMI 2.0b, DVI Displayport 1.4, HDMI 2.0b, DVI Displayport 1.2a, HDMI 2.0, DVI Displayport 1.2a, HDMI 2.0, DVI
 HDCP 2.2 support Yes Yes Yes Yes
 Maximum temperature 95° C 95° C 92° C 98° C
 Thermal power limit 180W 250W 165W
 Power connectors 1x 8-pin 1x 6-pin, 1x 8-pin 2x 6-pin
 Launch price $599 $379 $649 $549

Immediately, you can see that things have been irrevocably changed with this announcement, but there’s more still. The Geforce GTX 1080 is set to launch on 27 May 2016 worldwide, while the GTX 1070 trails it on 10 June 2016, just under a month later. Why things are this staggered is unknown at this point. NVIDIA’s never made a fuss like this in recent years, but that’s partially because they also haven’t had a new production process to introduce at the same time.

Let’s talk about the specs for a moment. Pascal is made using TSMC’s 16-nanometer FinFET production process, which allows NVIDIA’s GPUs to have wildly different characteristics to older architectures on larger nodes. It’s about three times as power efficient as Maxwell, and a lot of this is down to the lower amount of voltage being pushed through the chip, as well as the lower temperatures and leakage that it has to deal with. FinFET is much better when it comes to leakage than older processes, and this is what helps keep the GTX 1080 so cool. While running a demo by id Software on stage, Huang pointed out that the GTX 1080 demo was running above 2.0GHz clock speeds, with a GPU clock of about 5.5GHz, but still managed to stay cool at 67° C. That’s a pretty big achievement.

The GTX 1080 promises to be a monster GPU. It’s more than twice as fast as a GTX 980, and on paper almost matches two GTX 980 Ti cards in SLI. NVIDIA kept on making references to the GTX Titan X, but likely no-one is using that card for gaming purposes because its raw performance is much less impressive than the GTX 980 Ti, especially at current prices. Recent benchmarks put the Titan X and the GTX 980 Ti within 5% of each other, so a lot of that computing power goes wasted.

At 9.0 TFLOPS, it’s also faster than the AMD Radeon R9 Fury X, but AMD’s card settles in at 8.6 TFLOPS. NVIDIA’s specification is also based on the card operating at boost frequencies, but this happens pretty much all of the time on current Maxwell cards. Keep this in mind when looking at the graphs that follow below a little later. nvidia-GTX-1080-Founders-Edition-cooler

In terms of design, NVIDIA built one of the most advanced reference coolers in the world. The one for Pascal will be available as a “Founders Edition” version of the two cards, and carry a price premium over the standard model – $699 for the GTX 1080 and $449 for the GTX 1070. NVIDIA has reworked the shroud and heatsink design and added in a vapor chamber as well, which they claim is more efficient in terms of performance and drawing away heat than anything they’ve designed previously. It looks the business too, and is angled and aggressively shaped. A full metal backplate covers the GTX 1080’s rear, but NVIDIA did not say if the GTX 1070 also gets one.

NVIDIA Geforce Pascal launch (1)

Clock speeds are up, a great benefit of moving to FinFET. Huang claimed that seeing overclocks above 2.0GHz on the GPU was common, so this is a welcome reveal that means that enthusiasts will have a lot of leeway to play with. NVIDIA claims it’s able to do this thanks to reworked power circuitry and a new power phase design that allows the GPU to be fed cleaner and more stable power. Looking at the image above, the GTX 980 has a huge variation in its sine wave, while the GTX 1080 easily fits into that with less variation in power delivery.

There’s nothing available on the GTX 1070 just yet in terms of clock speeds, but they should be similarly high compared to Maxwell graphics cards. The power requirements are really insane, though – just one 8-pin PEG power connector is required for the GTX 1080, and a 500W power supply comes recommended.

NVIDIA Geforce Pascal launch (2)

NVIDIA’s performance metrics for the GTX 1080 hail it as the new leader in graphics performance, but as I expected, they’re massaging the performance improvements in anticipation of a new node change that might last longer than expected. As a result, the company claims that the relative gaming performance of the GTX 1080 is about twice as fast as the Titan X in the scenario where a Titan X is just as energy efficient.

The above graph is a bit misleading in that regard, and I don’t expect raw performance improvements of 200% over the previous generation. We’ll have to wait and see what the performance profile really is, but it’s still impressive nonetheless. The relative gaming performance is about the same as two GTX 980 Ti cards in SLI, which is unbelievably efficient.

NVIDIA Geforce Pascal launch (3)

Pascal is also much better suited to VR gaming than previous architectures, and some of the new features which I’ll detail later today are much better done on Pascal than Maxwell. The GTX 1080 is up to six times more effective for a given wattage than the GTX Titan X, and delivers more than twice the performance for the same amount of power as a GTX 980. Those are ridiculous numbers for a card of this size and complexity, and NVIDIA’s engineers should pat themselves on the back in this regard. It would be interesting to see how the Radeon Fury X stacks up to the GTX 1080 when given the same workloads. AMD’s GCN architecture is still very capable for VR workloads.

That’s all the interesting bits for the cards so far. NVIDIA has yet to go into the specifications of the GTX 1070, and as of this writing the card’s web page on the Geforce website hasn’t gone live yet. Review samples of the two cards should be making their way around the world as we speak, and the NDA can’t be far off now.

Stay tuned to NAG Online for a look at the other new things NVIDIA got to announce in their Geforce event later today.