Maxwell Graphics Architecture
Maxwell Graphics Architecture
As you can understand, the massive memory partitions, bus-width and combination of GDDR5 memory (quad data rate) allows the GPU to work with a very high framebuffer bandwidth (effective). Let's again put most of the data in a chart to get an idea and better overview of changes:
|GeForce||GTX 780||GTX Titan||GTX 780 Ti||GTX Titan Black||GTX 970||GTX 980|
|Streaming Multiprocessors (SMX)||12||14||15||15||13||16|
|GPU Clock (Core/Boost)||863/900||836/876||875/928||889/980||1050/1178||1126/1216|
|Memory Clock / Data rate||1502/6008||1502/6008||1750/7000||1750/7000||1750/7000||1750/7000|
|Memory bandwidth||288 GB/s||288 GB/s||336 GB/s||336 GB/s||224 GB/s||224 GB/s|
|Power connectors||1x6-pin PEG, 1x8-pin PEG||1x6-pin PEG, 1x8-pin PEG||1x6-pin PEG, 1x8-pin PEG||1x6-pin PEG, 1x8-pin PEG||2x6-pin PEG||2x6-pin PEG|
|Max board power (TDP)||250 Watts||250 Watts||250 Watts||250 Watts||145 Watts||165Watts|
|Recommended Power supply||600 Watts||600 Watts||600 Watts||600 Watts||500 Watts||500 Watts|
|GPU Thermal Threshold||95 degrees C||95 degrees C||95 degrees C||95 degrees C||95 degrees C||95 degrees C|
So we talked about the core clocks, specifications and memory partitions. Obviously there's a lot more to talk through. We feel that to be able to understand a graphics processor, you simply need to break it down into small pieces to better understand it. Let's first look at the raw data that most of you can understand and grasp. This bit will be about the Maxwell GM204 architecture. NVIDIA’s “Maxwell” GPU architecture implements a number of architectural enhancements designed to extract even more performance and more power efficiency per watt consumed. The first Maxwell-based GPU was codenamed “GM107” (GTX 750 and 750 Ti). Overall the architecture was designed with power-limited environments in mind. These products are now faster than 780 cards, yet only consume 165 Watts. Actually, this small statement sums things up nicely, there is 40% more performance delivered per processor over the last gen architecture.
Right, so have a close look at the die as shown above. You'll notice the green clusters. These are the GPC engines; each GPC has 4 SMX/SMM (streaming multi processor) clusters with 32 streaming core processing blocks in total. You'll spot four 64-bit memory interfaces, bringing in a 256-bit path to the graphics memory. At 7 Gbps default that's instant extra memory bandwidth by the way, the cards can reach 224 GB/sec.
So above, we see the GM204 block diagram that entails the Maxwell architecture, Nvidia started developing the GPU around 2011 actually. Let's break it down into bits and pieces. The GM204 will have:
- 1664 (GTX 970) or 2048 (GTX 980) CUDA/Shader/Stream processors used
- There are 128 CUDA cores (shader processors) per cluster (SMM)
- 5.2 Billion Transistors
- 2x Performance of the GK104
- 16 SMM
- 16 Geometry units
- 104 / 128 Texture units
- 64 ROP units
- 256-bit GDDR5
One SMX: 128 single‐precision CUDA cores, double‐precision units, special function units (SFU), and load/store units.
So based on a full 16 SMM 2048 shader cores chip the SMX looks fairly familiar in design, in the pipeline we run into the ROP (Raster Operation) engine and the GM204 has a nice 64 engines for features like pixel blending and AA. The GPU has 64KB of L1 cache for each SMX plus a special 48KB texture unit memory that can be utilized as a read-only cache. L2 cache wise the GPU has 2 MB. The GPU’s Texture units are a valuable resource for compute programs with a need to sample or filter image data. The texture throughput is significantly decreased compared to Fermi – each SMX unit contains 8 texture filtering units.
- GeForce GTX 780 has 12 SMX x 16 Texture units = 192
- GeForce GTX 970 has 13 SMX x 8 Texture units = 104
- GeForce GTX 980 has 16 SMX x 8 Texture units = 128
So there's a total of up-to 16 SMX x8 TU = 128 texture filtering units available for the silicon itself (once all SMXes are enabled). Typically lower is worse, but these cards however require little voltage and can be clocked very high. And that's where performance kicks in at low power consumption.
To reduce DRAM bandwidth demands, NVIDIA GPUs make use of lossless compression techniques as data is written out to memory. The bandwidth savings from this compression is realized a second time when clients such as the Texture Unit later read the data. As illustrated in the preceding figure, our compression engine has multiple layers of compression algorithms.
Any block going out to memory will first be examined to see if 4x2 pixel regions within the block are constant, in which case the data will be compressed 8:1 (i.e., from 256B to 32B of data, for 32b color). If that fails, but 2x2 pixel regions are constant, they will compress the data 4:1. These modes are effective for AA surfaces, but less so for 1xAA rendering. Therefore, starting in Fermi Nvidia also implemented support for a “delta color compression” mode. In this mode, they calculate the difference between each pixel in the block and its neighbor, and then try to pack these different values together using the minimum number of bits. For example if pixel A’s red value is 253 (8 bits) and pixel B’s red value is 250 (also 8 bits), the difference is 3, which can be represented in only 2 bits. If the block cannot be compressed in any of these modes, then the GPU will write out data uncompressed, preserving the lossless rendering requirement.
The effectiveness of delta color compression depends on the specifics of which pixel ordering is chosen for the delta color calculation. Maxwell contains the third generation of delta color compression, which improves effectiveness by offering more choices of delta calculation to the compressor. Thanks to the improvements in caching and compression in Maxwell, the GPU is able to significantly reduce the number of bytes that have to be fetched from memory per frame. Maxwell uses roughly 25% fewer bytes per frame compared to Kepler. This means that from the perspective of the GPU core, a Kepler-style memory system running at 9.3Gbps would provide effective bandwidth similar to the bandwidth that Maxwell’s enhanced memory system provides.