AMD Polaris – New GPU Architecture with Smaller Transistors
Last 4 years were quite unusual in GPU development history. Before 2011, Nvidia and AMD could rely on the regular improvements in transistor shrinking, to create GPUs with more transistors in the same size and at a lower power consumption. However, since HD7970 release in late 2011, high performance GPUs have remained on 28 nm process. Both GPU companies put a lot of effort into making large GPUs and improving their efficiency to get the most performance, while constrained by the power consumption and the physical size of chips they could produce. This year, we should finally see a new jump in the manufacturing process and a large one at that.
Samsung and TSMC are finally rolling out their 16 nm/14 nm FinFET manufacturing processes, which are suitable for powerful GPUs. A decade ago, such ratio between the old and new process would have meant 4 times reduction in chip area and somewhat similar reduction in power consumption for the same performance. These days, it is not so simple anymore and these numbers do not mean the same thing. While multiple chip foundries manufacture 14 nm chips, the technology behind the same name varies a lot in performance.
So, with the technological background in mind, let us look at AMD’s future GPU architecture. It is named Polaris (North Star) and AMD calls it the 4th generation GCN architecture. We can expect it to be an evolution of the earlier designs, with the additional benefit of long awaited node shrink. In terms of memory, Polaris cards will come with GDDR5 and stacked HBM. Most likely, HBM will be limited to the higher end GPUs, while most of the new line-up will continue using GDDR5. AMD plans to launch first Polaris GPU in the middle of this year, most likely starting with lower power models and going upwards after that. AMD has demonstrated a low-end Polaris prototype card at CES, comparing it against Nvidia’s GTX 950. Both were running the latest Battlefront at medium settings at 60 FPS. The PC with GTX 950 consumed 140 W, while PC with the Polaris prototype achieved the same performance using just 86 W of power. Considering the rest of the system and GTX 950 TDP, it is reasonable to assume around 2 times improvement in performance per watt compared to the latest 28 nm GPUs.
Polaris based graphics cards will also introduce some other features not directly related to 3D graphics rendering. While Fury did not support HDMI 2.0, Polaris will support both HDMI 2.0a and DisplayPort 1.3, utilising 4K 60 Hz TVs and giving the ability to connect to 5K 60Hz monitors or being ready for 4K 120 Hz monitors. For people who like to record their games, Polaris will have hardware encoding and decoding for the new h.265 video codec up to 4K 60 FPS.
Overall, it is quite exciting to see that we can look forward to a certain jump in GPU performance and efficiency this year. Nvidia is also preparing their Pascal architecture for the same node shrink. It will be an interesting year for PC gamers in terms of new hardware and ways to utilise the increased computational power.