Nvidia Maxwell Architecture changes things up a bit


Nvidia’s Maxwell architecture is coming very, very soon in the form of the mid-range GTX750 Ti, supposedly as soon as next week! It won’t be on a new silicon process, instead Nvidia will be testing out Maxwell on the same 28nm process as current Geforce 600 and 700 series cards. That’s similar to AMD’s current plans with GCN 1.1 and the Hawaii, Bonaire and Oland families but Maxwell is a little different because it includes an architecture change that’s very important in Nvidia’s mobile plans.

Let’s first refresh by understanding the following: Kepler-based Geforce graphics cards usually shipped with 192 CUDA cores inside each shader module. This was a functional doubling of the graphics units from the Fermi architecture but Kepler’s was different in that Nvidia dropped individual units that could be either emulated or replaced outright by CUDA cores that could do the job better.

Shoving 192 CUDA cores into a single shader module was also beneficial to Nvidia because, together with a die shrink to the 28nm process, it was able to get Kepler graphics into more devices and offer, at the very low end, performance that was equivalent to a Geforce GT630 graphics card.

Later on, Nvidia created the Tegra project and used custom ARM cores along with custom graphics hardware that was reminiscent of the Geforce 8000 era and at the time Tegra 4 was compared to similar graphics hardware found in the Sony Playstation 3. Nvidia’s next update to Tegra is K1 which, as the name suggests, is the first to implement Kepler-based CUDA cores and it includes exactly 192 CUDA cores. It’s taken Nvidia a very long time to do it, but they’ve finally gotten their graphics hardware down to being included in a full System-on-Chip (SOC). K1 implies that it boasts a single shader module but there’s no plans currently to add more.

The icing to this cake is that now you have the same graphics hardware in computers, laptops and mobile devices which makes it easier for developers to optimise their games for the different platforms (same argument as the PS4, Xbox One and the PC here) and Nvidia has been working on honing their gaming technology to allow it to run on anything from a K1 chip to a full-blown Titan.

Nvidia Maxwell architecture layout

According to leaked documents sent to VideoCardz, Maxwell changes this up a bit by rearranging the shader module assemblies and optimising the design further. Maxwell now has 128 CUDA cores to every shader and the upcoming Geforce GTX750 Ti features five of these shader modules slapped together. The diagram above is supposedly for the GTX750 Ti and belongs in the GM107 family. It shows 640 CUDA cores split into twenty smaller groups called Streaming Multiprocessors with eight texture units per shader module and 16 raster operators (ROPS) for the entire GPU. Two 64-bit memory controllers round out the offering, which means it’ll be a 128-bit memory bus.

Architecturally it’s not too dissimilar from Kepler and this is probably why Nvidia went ahead with Tegra K1 anyway despite the fact that Maxwell may be more efficient overall – transitioning from one to the other should be easy because the main designs don’t change too much. However, where the GTX750/Ti differ from the Geforce 800 series is the production processes, where 20nm will be reserved for those instead. For Nvidia, the first Maxwell cards will be a test of how efficient the architecture is and the next round will be to see how much in power and heat savings they can squeeze out of it.

AMD's Compute Core layout for Hawaii, with four sets of 64 shader cores in each one.
AMD’s Compute Core layout for Hawaii, with four sets of 64 shader cores in each one.

With regards to AMD, Maxwell is similar to the make-up of Hawaii as well. Where Hawaii’s single Compute Core holds a total of 256 shader cores, Maxwell has 128 per shader module. Logically, though, they’re organised in a similar fashion – Hawaii in four groups of 64 shaders, Maxwell in four groups of 32 shaders. While one can’t exactly compare the two architectures in terms of unit counts and layout, they are very, very similar to each other.

This means that we could see some interesting offerings in the mobile and desktop market in the future. Instead of using a single shader module in the case of the GT630, it would be possible to shove two shader modules into a slightly larger die area and have overall better performance thanks to the extra 64 CUDA cores, allowing gamers to crank up the detail settings a bit. According to VideoCardz, Maxwell is also more efficient than Fermi, with the GTX750 Ti delivering the same performance as the GTX480 while using less than four times as much energy for the same tasks.

With just these leaks, I think I can make a few predictions as to where this is going to go:

1) Fermi and Kepler-based low-end GPUs get the boot to make space for new 800 series GPUs in the low-end. This will probably start off with the GT810 with 256 CUDA cores, 16 texture units and four ROPs, but I think that will only pop up in the second half of 2014.

2) Compared to Kepler, we’re going to see a big drop in power requirements, to the tune of 50% or more. The equivalent of the GTX660 Ti with Maxwell hardware will probably only require a single 6-pin PEG power connector.

3) Memory buses will top out at 512-bit buses. It makes sense, considering that AMD’s Hawaii cards top out at this level and Nvidia needs to gear itself up for the UltraHD 4K era.

Source:  Videocardz

Discuss this in the forums: Linky