“It’s insane,” Nvidia CEO Jen-Hsun Huang proudly admitted during a GeForce GTX 1080’s reveal, holding a graphics label aloft. “The 1080 is insane. It’s roughly insane amounts of performance… a 1080 is a new king.”
He wasn’t joking. The long, barren years of stalled GPU record are over, and this savage is badass.
A hulk burst for GPU-kind
As different as it is, a vast opening burst of a GTX 1080 (starting during $599 MSRP, $699 Nvidia Founders Edition reviewed) doesn’t accurately come as a surprise.
Faltering graphics processor routine record left graphics cards from both Nvidia and AMD stranded on a 28-nanometer transistor node for 4 prolonged years—an roughly infinite length of time in a lightning-fast universe of complicated technology. Plans to pierce to 20nm GPUs fell by a wayside due to technical woes. That means a 16nm Pascal GPUs violence inside a GTX 1080’s heart (and AMD’s forthcoming 14nm Polaris GPUs) paint a burst of two full routine generations.
That’s nuts, and it alone could emanate a large fanciful burst in performance. But Nvidia didn’t stop there.
Pascal GPUs adopted a advanced FinFET “3D” transistor technology that done a initial mainsteam coming in Intel’s Ivy Bridge mechanism processors, and a GTX 1080 is a initial graphics label powered by GDDR5X memory, a supercharged new chronicle of a GDDR5 memory that’s come customary in graphics cards for a few years now.
On tip of all that, Nvidia invested significantly in a new Pascal pattern itself, quite in tweaking efficiencies to boost time speeds while concurrently shortening energy requirements, as good as many some-more under-the-hood goodies that we’ll get to later—including extended asynchronous discriminate facilities that should assistance Nvidia’s cards perform improved in DirectX 12 titles and fight a vital Radeon advantage.
Oh, and did we discuss all a new facilities and performance-enhancing program alighting alongside a GTX 1080?
Note: Because this is a vital GPU advancement, we’ll spend some-more time than common deliberating under-the-hood sum and tech specs. If that’s not your thing, burst to page two for contention on a GTX 1080’s large new technical wonders and page three for a new consumer-facing features. Performance speak starts on page four.
Let’s flog things off with an Nvidia-supplied spec piece comparison of a GTX 1080 vs. a predecessor, a GTX 980. (Side note: The tiny fact that a company’s comparing a GTX 1080 directly opposite a GTX 980 is noteworthy. Usually, GPU makers examination new graphics cards opposite GPUs dual generations behind in examination materials. The GTX 960 was compared opposite a GTX 660—not a GTX 760—in Nvidia’s central materials, for example.)
Here, some of a advantages to switching to 16nm burst out immediately. While a “GP104” Pascal GPU’s 314mm2 die distance is extremely smaller than 398mm2 die in a comparison GTX 980, it still manages to fist in 2 billion some-more transistors overall, as good as 25 percent some-more CUDA cores—2560 in a GTX 1080, contra 2048 in a GTX 980.
And collect adult your jaw! The GTX 1080 indeed rocks definitely ridonkulous 1,607MHz bottom time and 1,733MHz (!!!!) boost time speeds—and that’s usually a batch speeds. We managed to holder it to over 2GHz on atmosphere but violation a persperate or tinkering with a card’s voltage. Add it all adult and a new graphics label blows a prototype out of a H2O in both gaming opening and discriminate tasks, leaping from 4,981 GFLOPS in a GTX 980 all a approach to 8,873 GFLOPS in a GTX 1080.
Diving even deeper, any Pascal Streaming Multiprocessor (SM) facilities 128 CUDA cores, 256KB of register record capability, a 96KB common memory unit, 48KB of L1 cache, and 8 hardness units. Each SM is interconnected with a GP104 PolyMorph engine that handles zenith fetch, tessellation, viewport transformation, zenith charge setup, viewpoint correction, and a intriguing new Simultaneous Multi-Projection record (which we’ll get to later), according to Nvidia.
A organisation of 5 SM/PolyMorph engines with a dedicated raster engine forms a Graphics Processing Cluster, and there are 4 GPCs in a GTX 1080. The GPU also facilities 8 32-bit memory controllers for a 256-bit memory bus, with a sum of 2,048KB L2 cache and 64 ROP units among them.
That segues easily into another technological allege in Nvidia’s card: a memory. Despite rocking a 256-bit train a same distance as a predecessor, a GTX 1080 managed to pull a altogether memory bandwidth all a approach to 320GBps, from 224GBps in a GTX 980. That’s interjection to a 8GB of cutting-edge Micron GGDR5X memory inside, that runs during a peppery 10Gbps—a full 3Gbps faster than a GTX 980’s already rapid memory. How quick is that, really? Nvidia’s GTX 1080 whitepaper sums it up:
“To put that speed of signaling in context, cruise that light travels usually about an in. in a 100 picosecond time interval. And a GDDR5X IO circuit has reduction than half that time accessible to representation a bit as it arrives, or a information will be mislaid as a train transitions to a new set of values.”
Implementing such rapid memory compulsory Nvidia to redesign both a GPU circuit pattern as good as a house channel between a GPU and memory dies to harsh specifications—a routine that will also advantage graphics cards versed with customary GDDR5 memory, Nvidia says.
Pascal achieves even larger information transfers capabilities interjection to extended memory application technology. Specifically, it builds on a delta tone application already found in today’s Maxwell-based graphics cards, that reduces memory bandwidth final of organisation like colors together. Here’s how Nvidia’s whitepaper describes a technology:
“With delta tone compression, a GPU calculates a differences between pixels in a retard and stores a retard as a set of anxiety pixels and a delta values from a reference. If a deltas are tiny afterwards usually a few pieces per pixel are needed. If a packaged together outcome of anxiety values and delta values is reduction than half a uncompressed storage size, afterwards delta tone application succeeds and a information is stored during half distance (2:1 compression).”
The new Pascal GPUs perform 2:1 delta tone application some-more effectively, and combined 4:1 and 8:1 delta tone application for scenarios where a per-pixel tone movement is minimal, such as a darkened night sky. Those are targets of opportunity, though, given a application needs to be lossless. Gamers and developers would dispute if GeForce cards started screwing with picture quality.
Using tone application to revoke memory needs isn’t new during all—AMD’s Radeon GPUs also do it—but Nvidia says that between this new, some-more effective form of application and GDDR5X’s benefits, a GTX 1080 offers 1.7x a sum effective memory bandwidth of a GTX 980. That’s not unfair during all, and it takes some of a prick out of a card’s miss of insubordinate high-bandwidth memory, that debuted in AMD’s Radeon Fury cards, despite in capacities singular to 4GB.
The Pascal GPU’s technological enhancements and burst to 16nm FinFET also make it impossibly energy efficient. Despite resolutely outpunching a Titan X, a GTX 1080 sips only 180 watts of energy over a singular 8-pin energy connector. By comparison, a GTX 980 Ti sucks 250W by 6-pin and 8-pin connectors, while a 275W Fury X uses a span of 8-pin connectors. The GTX 1080 does a lot some-more opening with a lot reduction power.
Next page: New features! Async compute, coexisting multi-projection, and more