Hate to say I told you so. Or rather, I don't and so I'm going to gloat. Contrary to numerous comment protestations, Nvidia's 2016 graphics awesomeness has begun in the shape of its new GeForce GTX 1080 and 1070 cards. Based on the new Pascal architecture and teensy 16nm transistors, the new GPUs are exactly as expected. And yet also quite different. Meanwhile, AMD has dropped some hints regards the shape of Radeons to come. It all adds up to an exciting summer for PC graphics and a very good reason to put your GPU purchases on temporary hold, especially if VR is your bag...
So, yes, many doubted this would be happening so soon. But Nvidia has announced its first Pascal graphics chips for gamers, the GeForce GTX 1080 and 1070. Obviously, this is initially a paper launch, but I can assure you cards are incoming and I'll report back when I'm eyeballing actual hardware myself.
So, what's this new Nvidia clobber all about? The non-surprise aspects involve the shrunken 16nm bits made by production partner TSMC, including FinFET tech, and the Pascal architecture. Long story short, that all adds up to in excess of a double-generational leap in terms of the chip production technology. And it's ultimately that technology that underpins progress and improved performance for computer chips of all kinds.
The key specifications for the 1080 version start with 2,560 of Nvidia's CUDA cores (shaders to you and I), 160 texture units (this particular figure isn't fully confirmed) for, you know, processing textures, 64 render outputs for squirting pixels in the general direction of your display and 7.2 billion transistors.
For reference, the existing card that the 1080 replaces, namely the GTX 980, rocks in at 2,048 shaders, 128 textures, 64 render outputs and 5.2 billion transistors. Based on those numbers alone, the 1080 looks like a disappointment. A double-generational leap in production tech and not even 50 per cent more transistors or shaders? Really?
Nvidia's lean, mean pixel-pumping machine
Well, here comes the mitigation. For starters, Pascal shaders are not directly comparable to the Maxwell shaders in the old 980. Pascal shaders probably do more, er, graphicsy stuff. But more importantly, these new Pascal chips run way, way faster. The maximum boost clockspeed for the 1080 is 1,733MHz to the mere 1,216MHz of the old 980. In both cases, we're talking speeds for standard 'reference' cards. Board makers like Asus, MSI and the rest will inevitably cook up overclocked specials.
The rest of the spec list includes a 256-bit memory bus and GDDR5X memory with a data rate of 10Gbps – note these first Pascal boards do not boast fancy HBM memory, though GDDR5X is a minor novelty and features four data transmissions per cycle to the two of GDDR5. IE it's twice as fast at any given operating frequency.
Anyway, the upshot is a near doubling in raw computational performance over the 980, from 5TFLOPS to 9TFLOPS. Actual gaming performance will not be twice as fast, I suspect, but something in the region of 50 per cent faster in many cases... though with one critical exception. With Pascal, Nvidia has cooked up some special sauce for VR gaming.
Simultaneous Multi-Projection, for it is ye, is a new feature that allows for rendering of multiple viewpoints – up to 16 – in a single pass. At this stage, the technical details are a little foggy. But the key implication is a huge leap in VR performance.
Virtual reality rendering, of course, requires stereo images – one for each eye. The basics of Simultaneous Multi-Projection involve rendering the geometry for both eyes just once and then also discarding pixels that won't be rendered in a VR scenario before they are rasterised. As I said, the full details aren't clear, but Nvidia is more or less implying that you get two rendered images for the price of one which is exactly what you want for VR.
See, it's faster. But note the 'VR' small print...
Certainly, Nvidia is claiming that the 1080 will be twice as fast as even a GeForce Titan X for VR rendering. That's stunning if true and would make the 1080 an absolute, instant no brainer for VR fans. Simultaneous Multi-Projection also has beneficial implications for multi-monitor setups.
Elsewhere, further interesting features include the fancy new 'Ansel' screenshot tech that Alice mentioned previously. Again, this is effectively a technology that takes a snapshot of the whole 3D engine rather than a 2D frame. In other words, once captured, you can navigate around the frame in full 3D, apply filters and effects and even up the resolution and quality settings dramatically. So you can take a screen shot on a feeble £100 GPU and make it look like you're running a 5K panel and triple GTX 1080s. Apparently, the final output can extend to 61,440 by 34,560 pixels. Madness.
Inevitably, there's a whole bunch of other stuff including what's being billed as 'ray-traced' audio designed expressly for VR gaming. You can grab Nvidia's own overview here.
As for the slightly cheaper, cut-down 1070, full details haven't been released yet. The 1080 officially goes on sale May 27th, the 1070 follows on June 10th. Pricing is officially $599 and $379 for the 1080 and 1070. Call that £450 and £270. For the record, the 1080 is $50 more expensive than the 980 was at launch, which is disappointing given that it is very likely a smaller, cheaper GPU to manufacture.
It's arty. But is it farty?
Overall, these new chips are intriguing. I was expecting a bigger and more complex GPU, but a slower running GPU. The net result in terms of performance is probably about level with expectations but with that VR-performance kicker. In that context, the 1070 could well be the just-about-affordable sweetspot for VR gaming.
But what of AMD? Where Nvidia has Pascal, AMD will have Polaris, its own new family of chips. Nothing specific has been announced, but in recent interviews, AMD reps have been hinting that the target for the new GPUs is more mainstream than high end.
Wind your minds back to the Radeon HD 4000 series (cough) and you'll remember that AMD made a big song and dance about the futility of high end willy waving in the graphics market - not many can afford £500 graphics cards - and instead concentrating on maximising performance at a price more people can swallow. Nearer £200/$300 than £400/$600, in other words.
I expect we'll see a similar sales pitch for the new AMD Polaris cards. Whether this is by design or because AMD has slightly misjudged either its technology or what Nvidia has achieved probably doesn't matter. What does matter is that it should make for real choice for us gamers. Absolute VR power from Nvidia. Bang for buck from AMD. Sounds good to me.