Nvidia's new Maxwell graphics kit, then. It's out but what's it all about? Epic performance density and power efficiency is the elevator pitch, with a spot of improved cryptocurrency hashing thrown in for good measure. But are the first new Maxwell boards – the GTXs 750 and 750 Ti – the bomb or a bum deal?
And so to Maxwell. That's the codename for Nvidia's latest and greatest GPU architecture. In time it will entirely replace Kepler, which is the basis for nearly all the GeForce 600 and 700 series GPUs to date, including the Titan models.
Confusingly, the first Maxwell GPUs are actually badged Geforce GTX 750 and 750 Ti, which doesn't exactly scream 'new architecture'. But we'll come back to the branding shenanigans in a moment. Let's deal with Maxwell as an overall architecture and then have a look at the first members of the new family.
The big news is that Nvidia is pitching Maxwell as 'mobile first'. In other words, Maxwell is designed as a mobile architecture first and then adapted for desktop. Shades of Intel Core? Absolutely.
As a corollary, Nvidia is claiming the big win with Maxwell is efficiency or performance per watt. How much of this is expediency is a tricky question. I reckon there's little doubt Nvidia was hoping to have access to a new, smaller production node from production partner and uber chip fabber TSMC by now.
As it is, they're still stuck on 28nm and that means launching big Maxwells isn't an option. Instead, the first GPUs are relatively low end and so the message can't be about record-breaking performance.
That said, Maxwell certainly has extensive new mobile optimisations. It does indeed look a lot like Kepler re-jigged with a view to power efficiency.
So in terms of functionality, not all that much has changed. There's no whizz-bang new rendering tech. The basic DX11 feature set is carried over from Kepler, which helps to make sense of the 700-series branding for the first Maxwell-based boards. That's actually good news if you own a Kepler card because it means you're not suddenly going to be locked out of some wondrous new rendering feature.
Instead, it's a case of taking existing technology and rearranging it for more efficient operation. Result is a claimed doubling of performance per watt. If true, it's genuinely impressive – remember these first Maxwells don't get the benefit of a new production node. Of course, as the first truly mobile optimised architecture, you'd expect big gains. Ye olde low hanging fruit syndrome. Future gains probably won't be as dramatic.
Anyway, the headline stuff involves a new block design. Kepler was based on a streaming multiprocessor block known as an SMX containing 192 shader cores, four warp schedulers and eight dispatch units, the latter two effectively feeding the 192 shaders as a homogeneous pool.
With Maxwell, the SMX becomes an SMM and now contains just 128 shaders. However, those 128 shaders are split into four groups of 32, each with its own warp scheduler and two dispatch units.
What's a warp scheduler, you ask? Let's not go there. Just know that Nvidia says the new 128-core SMM delivers 90 per cent the performance of the old 192-core SMX. Yeah, really. Needless to say, the new 128-core SMM is also much smaller than the old 192-core SMX. So you can have much more performance from a given size of GPU. Or the same performance in a much smaller, lower-power chip.
To be frank, Nvidia was already doing pretty spectacular things with Kepler when it came to the size of its GPUs and the performance they kicked out, compared to AMD at least. Maxwell looks truly epic by this metric.
There are plenty of other changes, including a massive uptick in the amount of on-die cache memory. Nvidia has also upped Maxwell's general compute ante – it will be much more competitive with AMD than Kepler when it comes to mining cryptocurrencies, if that sort of thing is your bag. But that shift from SMX to SMM is the bit you really need to comprehend.
As for the first Maxwells out of the gate, the aforementioned Geforce GTX 750 and 750 Ti, they're based on the new GM107 GPU. Price-wise, we're looking at roughly £90 and £115 respectively, which will give you an idea of how they're positioned - they're entry-level gaming GPUs.
Respectively, they sport 512 and 640 shaders and share the same maximum boost clock of 1,085MHz. At this point, it's worth noting that the equivalent chip from the old Kepler family, GK107, had just 384 shaders.
If there is a catch with GM107 it's the memory bus. It's a miserable 128-bit item. That added on-die cache I mentioned might help offset that a bit. And in reality nobody is going to buy either of these new Maxwells to game beyond 1080p. But I still instinctively gag at the mention of a 128-bit bus on any GPU. Memory bandwidth is simply too important.
One final point to note re the specs is that all the effort on power efficiency means neither the 750 nor 750 Ti require a supplementary six-pin PCI Express power connector as standard. They get sufficient power from the PCI Express bus. That's pretty significant in the context of building a low-power, super-silent game box for the living room. Just be aware most if not all of the factory overclocked boards already being offered do have a six-pin power port.
Anyway, what should you make of these new boards and Maxwell in general? The 750 and 750 Ti themselves aren't going to change your world much unless you're hoping to build a small form factor PC and seek the best possible performance per watt.
Early benchmarks from the usual suspects around the web agree that Maxwell sets new standards for power efficiency – the 750 Ti looks to be getting on for three times quicker than the GT 640, fastest 600-series board without a supplementary power connector.
As a pure performance proposition, they're less immediately compelling. But that's thanks to how Nvidia has positioned them price-wise. In simple terms, AMD's Radeon R7 260X is faster than the GTX 750 and the R7 265 is faster than the 750 Ti. The end.
Oh, yeah, did I mention AMD recently tweaked its product range in preparation for the new Maxwell? No surprise there.
When you start looking beyond pure performance, it all gets hideously complicated, of course. Suddenly, it's Mantle vs G-Sync or whatever and you tumble down the proprietary-tech rabbit hole. The way things are set up in the graphics market right now, you can't have it all. You must pick sides.
But Nvidia has definitely done some pretty staggering things with Maxwell regards efficiency and power density. Even on existing 28nm chip manufacturing tech, Maxwell is a bit special in that regard. When higher-end Maxwells based on TSMC's 20nm node appear – presumably later this year – I've a feeling the results are going to be truly spectacular.
Quite what Nvidia is going to call the more powerful Maxwell chips is hard to say, of course. I doubt there will be a Maxwell with better raw performance than a 780 Ti any time soon – GK110 is still one hell of a chip - which makes launching, say a Geforce GTX 880 that's slower than a 700-series board a bit tricky. But that's the way GPU branding has gone these days. It's a mess.