2016 Awesomeness: Nvidia's New Pascal Graphics
Oh my God, it's full of transistors...
If it was a car it would be a gold-wrapped, kleptocrat-owned Bugatti Veyron ostentatiously double parked outside a Knightsbridge hotel. It's still bloated, it's still overly complex and you still can't afford it. But it's a graphics chip and a harbinger of things you might actually be able to buy. I give you Nvidia's new Pascal GP100, a 15.3 billion transistor beast and the beginnings of that 2016 awesomeness I promised for the new year. In other words, if you're thinking of buying a new graphics card, you might want to hold fire. Meanwhile, Intel has also taken the wraps off a massive new chip you can't afford and the final piece the Laird Gaming Dungeon™: Driver Edition has arrived. Yup, I'm liking 2016.
Nvidia's Pascal, then. It's a whole new family of graphics chips to replace the incumbent Maxwell crew. The first to break cover is known as GP100, which also happens to be the flagship chip of the family. The snag is that it's been announced as a Tesla-branded high-performance parallel compute product rather than a gaming video card.
You could argue that's an academic distinction given that the price of the future gaming card based on GP100 will be out of reach of almost all of us. Instead, what matters is that the new Tesla P100 board confirms that Nvidia's new Pascal chips are incoming in all their 16nm FinFet glory.
I've been through this before, but the short version is that PC graphics chip have stagnated in recent years and 2016 is going to see what amounts to a double-jump in chip manufacturing tech from 28nm transistors to 16nm. And that should translate into a much bigger leap in performance than usual with this new generation.
You can rest assured this is one graphics card that can play Crysis. Well it could if it was a graphics card. Which it isn't. Oh well
For the record, the new GP100 chip is an absolute monster at around 600mm2 and over 15 billion transistors. For context, an eight-core Intel Core i7 is just 356mm2 and Nvidia's previous biggest graphics chip, as found in the Titan X, rocks in at just eight billion transistors.
Anyway, other highlights include a crazy 4,096-bit memory bus, just like AMD's Radeon R9 Fury but with second-gen HBM or high-bandwidth memory and 16GB of the stuff instead of the Fury's arguably problematic 4GB buffer.
Elsewhere, you might argue the GP100's key metric is its 3,840 eye candy-creating shader thingies, or CUDA cores as Nvidia likes to call them (a few of which are disabled in the Tesla board). That compares with 3,072 from Nvidia's current gaming monster, the Titan X.
Nvidia's Pascal GP100, yesterday
However, for Pascal, the internal structure has changed, with each of Nvidia's so-called Streaming Multiprocessors (SM) specced up with 64 CUDA cores where the outgoing Maxwell family has 128 CUDA cores per SM.
What's more, Nvidia has loaded this new chip with lots of double-precision compute performance that's irrelevant to gaming but will soak up some of that epic transistor count. The point being that there's extensive architectural change here and a direct comparison of the CUDA core count probably makes no sense, especially for GP100.
But here's the point. Nvidia says the chip is in production right now as is the fancy new HBM2 memory. One day, maybe late this this year but perhaps further out, it will make its way into the Geforce Titan Über Alles, or whatever the hell they call it, and it will be pointlessly expensive and hardly any of us will buy it.
Pathetic 2015 humans and their puny Titan Xs beware, Pascal is coming...
But they've made this mega GPU work. And that bodes well for the smaller Pascal chips many of us might buy. I would guesstimate late summer for the first Pascal for us lot, a replacement for the current GeForce GTX 980, with more mainstream members of the Pascal family following swiftly.
Thus, the main moral here is that if you are in the market for a new graphics card, hold on a few months. This kind of technological leap doesn't happen every year. 2016 is going to be special.
So, at worst whatever you're planning on buying will be a bit cheaper. At best you may be able to buy dramatically more performance for your money. Watch this space.
As for Intel, it has just released its new Broadwell-E/P CPUs. A bit like Nvidia and the new Pascal chip, Broadwell-E is being released first as a heavy iron product for servers. And just like Pascal, the numbers are bonkers.
This is what a 24-core block diagram looks like...
The top Broadwell-E/P will have 22 cores and 44 threads – count 'em! (actually the top chip has 24 cores in silicon, it's crazy stuff) - but it's likely by the time it makes it into desktop PCs, those numbers will be 10 and 20 with a baseclock around 3.0GHz and a Turbo clock of about 3.5GHz.
Of course, as is the case with all of these high end chips, it lags a generation behind the mainstream Skylake Core i3, i5 and i7s in terms of architecture. But the point here is that if AMD actually delivers with its new Zen chips, Intel has a cupboard full of chips up to 22 cores with which to respond.
I'm sure it won't need to go anywhere near the full 22 core chip on the desktop. But the 12 core, 14 core or even 18 core variants?
And, of course, what with DX12 and Vulkan promising much better threading support for games and VR potentially shaking things up, too, having loads of CPU cores might actually be relevant for gaming. It'll be fun finding out.
And finally, thanks to Playseat.com I finally have access to a proper driving pod. As they say on the BBC, other makes of driving seat are available, but this is the one currently resident in my house.
I am excited. Are you excited?
Just having the seat sitting there has me genuinely more excited about the whole escapade. I do love a good bucket seat. Anyway, the empirical work of data gathering has begun, so tune back in a few weeks or so from now for what I shall confidently pitch as, but won't remotely be, the definitive dissertation on modern driving games.