During its Game24 event that kicked off earlier today, Nvidia has finally unveiled its new GM204 Maxwell GPU and two new graphics cards based on that same GPU, the Geforce GTX 980 and the Geforce 970. The new GM204 Maxwell GPU will bring quite a few new features and offer twice the number of CUDA cores, double the amount of ROPs and four times more cache compared to the Kepler GPU.
Based on the 28nm manufacturing process and featuring a staggering 5.2 billion transistors on a 398 mm² die size, the GM204 Maxwell GPU brings previously seen power efficiency as well as a couple of new features.
Just like previous Fermi and Kepler GPU, the GM204 is also composed of an array of Graphics Processing Clusters (GPCs). In the Geforce GTX 980, which is apparently based on a fully enabled GM204 GPU, each GPC features a dedicated raster engine and four Streaming Maxwell Multiprocessors (SMMs). This adds up 16 SMMs, each with 128 CUDA cores, a PolyMorph Engine and eight texture units, adding up to 2048 CUDA cores and 128 Texture Units (TMUs). Each GPC also features a memory controller, or in this case four 64-bit memory controllers for a 256-bit wide memory interface.
Each memory controller is connected to 16 ROP units and 512KB of L2 cache, which adds up to a total of 2048KB of L2 cache and 64 ROPs. This is quite an improvement compared to the the Kepler GK104 GPU which features 512KB of L2 cache and 32 ROPs.
The memory subsystem on the GM204 has been reworked in order to reduce DRAM bandwidth demands. With this 3rd generation delta color compression with multiple layers of compression algorithms as well as certain caching improvements, Nvidia claims that the GPU is able to significantly reduce the number of bytes that have to be fetched from memory per frame, by up to 25 percent less compared to the Kepler GPU.
What is a bit more important regarding the new GM204 Maxwell GPU are the new features, and there are quite a few interesting ones including Direct3D 11.3, Voxel Global Illumination (VXGI) lighting, Dynamic Super Resolution, new Multi-Frame sampled Anti-Aliasing (MFAA) and the virtual reality VR Direct Technology.
For starters, the new GM204 Maxwell GPU comes with full Direct3D 11.2/11.3 compatibility, including features like Rasterizer Ordered Views, Typed UAV Load, Volume Tiled Resources, and Conservative Rasterization. The Voxel Global Illumination (VXGI) lighting is a new way of global illumination by using voxels. The VXGI lighting uses a voxel grid and cone tracing technique to calculate an approximation of global illumination as well as ambient occlusion on the GPU in real time. The GPU uses support for a number of new graphics features that dramatically speed up the voxelization process where we convert triangles into voxels and dramatically speeds up the Global Illumination (GI). The new VXGI will be available for Unreal Engine 4 and other major game engines in Q4 2014.
The new GM204 Maxwell GPU also brings the new Multi-Frame sampled Anti-Aliasing (MFAA). While MSAA was great, it was also a big hit on the performance. The MFAA is actually a new anti-aliasing technology that alternates AA sample patterns both temporally and spatially to produce the best image quality while still offering a performance advantage compared to traditional MSAA. According to Nvidia, the result is an image quality which is close to 8xAA on MSAA with performance hit of about 4xAA MSAA, or about 30 percent faster than MSAA.
In addition to the new MFAA anti-aliasing technology, the GM204 will also bring the Dynamic Super Resolution, described by Nvidia as "a new technology that is designed to provide gamers with better looking games." Basically, Dynamic Super Resolution renders the game at higher resolution and then scales the image down to the native resolution, thus providing a better level of detail. In order to address the issue that happens when downsampling is concerned the Dynamic Super Resolution uses a 13-tap Gaussian filter during the conversion to display resolution, which reduces or eliminates the aliasing artifacts.
The GM204 also brings a couple of display related features, including a total of five display outputs, four of which can be used at the same time, including three DisplayPort outputs, dual-link DVI output and the new HDMI 2.0 connector for 4K@60Hz output. this display output configuration now allows to connect three G-Sync displays from a single GTX 980 graphics card. The GM204 also features a new NVENC video encoder that brings full support for HEVC (H.265) encoding as well as hybrid HEVC decode support seen earlier with the GM107 Maxwell GPU and Kepler GPUs.
One of the new features is VR Direct, Nvidia's pledge to virtual reality. With the increasing popularity of virtual reality, mostly thanks to Oculus Rift, Nvidia has introduced the new VR Direct, a set of technologies and features that will improve experience and performance of virtual reality headsets by lowering latency/input lag. Nvidia worked on reducing the lag from the GPU as well as with asynchronous warp features, new MFAA anti-aliasing and VR SLI , which uses two GPUs to each eye, rather than using alternate frame rendering (AFR).
As noted earlier, the GM204 is the GPU behind the newly launched Geforce GTX 980 and GTX 970 graphics cards so all these features will be available on both graphics cards. While the Geforce GTX 980 is based on a fully enabled GM204 GPU with 2048 CUDA cores and 128 Texture Units (TMUs), the GTX 970 is a cut down version with 1664 CUDA cores and 104 TMUs, while the rest of the specifications are the same on both graphics cards.
The GM204 will be the foundation for future Geforce 900 series graphics cards and the gap in price between the US $329 priced GTX 970 and the US $549 GTX 980 graphics card clearly shows that we will soon see some other graphics cards with different CUDA core number, possibly the GTX 970 Ti.