GDDR6 Memory

GDDR6 DRAM is the direct successor to GDDR5X, which in turn was only used by the Pascal family of NVIDIA graphics cards. As with last time, NVIDIA has teamed up with the DRAM industry to help move forward the capabilities of graphics memory without investing much more into HBM2, which is a more expensive alternative, especially when it comes to the consumer market.

While the differences between GDDR5X and GDDR6 are not as big as the change from GDDR5 to GDDR5X, there are still some notable ones. To begin with, clock frequencies have been increased to 10–16 Gbps, with NVIDIA opting to use 14 Gbps for the current offerings. Secondly, GDDR6 splits memory into two channels per chip, allowing, for instance, a 32-bit chip to now output two independent 16-bit channels, which should in turn help further with power and performance metrics in highly parallelized scenarios (which GPUs tend to operate at). Then there is the part where GDDR6 can have an operating voltage of 1.35 V or 1.25 V (the latter in cases where a reduced data rate is not critical) compared to GDDR5X that has a specified operating voltage of 1.35 V only. This means that the adoption of GDDR6 provides many options for more heat- and power-efficient solutions, though it remains to be seen if this comes with a tangible deficit to performance potential.

Finally, perhaps the most impactful difference is that unlike GDDR5X, all major DRAM vendors are expected to adopt GDDR6. We have already seen announcements from Micron, Hynix, and Samsung, which means that GDDR6 will become the memory workhorse of new graphics card generations from both AMD and NVIDIA—just like GDDR5 has been in the past.

Lossless Color Compression

For several generations now, NVIDIA GPUs have had a lossless memory compression technique that reduces memory bandwidth requirements. In a nutshell, the GPU will look for repeating pixels in textures and, instead of storing every single repeated pixel, it will store just "repeat blue, 4x4 block". Similarly, if the colors of neighboring pixels are similar, it will store how to calculate the next pixel's color from the previous pixel's color using a handful of bits instead of the full 32-bit RGB color value. The key to improving this algorithm is to find a pattern in textures that are used in games, and then add dedicated circuitry to the chip to enable compression and decompression of this one specific pattern. I wouldn't be surprised if NVIDIA used machine learning to optimize their collection of such recognized patterns by feeding it all textures in all games ever created, which is an extensive task with the potential to yield fantastic results almost immediately.

Effectively, color compression means less data to write to, and transfer from VRAM to the L2 cache. This reduced traffic propagates down the line with fewer data transfers between clients, including the aforementioned game texture, and the memory framebuffer. With Turing, NVIDIA has developed the 5th generation of the color compression technique which improves the compression ratio by over 25% relative to the Pascal microarchitecture, which in turn already had a 20% improvement over Maxwell for further context.