Amd 'polaris' architecture in detail

We are not going to roll up excessively and we are going to break this article down into several parts. The first and most important we will see the new GCN 4.0 scheme with the diagram of the RX480, we will talk about its front-end and the modifications it has undergone, the intermediate parts such as Shaders, memory controllers etc and finally the less important parts but which have its relevance. Let's start…

RX480 diagram. 'Polaris' in all its glory.

Just by taking a look, we cannot avoid having in mind one of the relevant gpus of the previous mid-range, the R9 380 / 380X, because this scheme is tremendously similar in composition and location to the elements that make up the new 'Polaris'.

Much has been discussed about asynchronous computing but everything indicates that it will become even more prominent with virtual reality (VR onwards) as well as low-level apis (Directx 12 and Vulkan), but this is another issue that we did not come to discuss. but news that 'Polaris' brings and that we will have with them.

We recommend reading our guide to the best graphics cards.

The new front-end scheme has 4 asynchronous computing engines (ACEs) and two new HWS units (Hardware Scheduler), or translated into our language, hardware programmers.

HWS will have Shaders available at all times for a specific task, being this the highest priority to access those Shaders. This complete and complex operation is of vital importance for the new low-level apis (DX12 and Vulkan) or for the VR since it was very difficult to guarantee the available resources complicating the computation tasks, such as for audio processing, planning processes and manage the balance of resources between calculation tasks and graphs reducing the dependence on the CPU. Each of these units can access the entire gpu individually.

These new units are accessed via microcode, that is, they are programmable and AMD can update their operation. Gradually, games or software that support it will arrive and the programmers will take advantage of its features. These units are available in the 480, 470 and even the 460 without cutouts.

Compute Units and Geometry Engines.

The Shader system by Compute Unit remains the same, 64 Shaders for each of them. In the RX480 we have a scheme of 36 CUs giving a total of 2304 Shaders in total.

The improvements we have are mainly in the caches and in the prefetch (prereads) making the storage of instructions more efficient. The level 2 cache has increased from 768Ks to 2Mb and in addition to being more efficient to access now it can also be grouped.

The buffer for saving instruction waves is much larger, and is the one waiting for tasks. And as a novelty we have the native ability to execute Fp16 and Int16 operations. Regarding the Hawaii architecture, AMD tells us that we have up to 15% more performance per block over it with Polaris.

Finally, one of the most anticipated improvements comes from the geometry engines. Bring along the new ' Primitive Discard Accelerator '. This unit has the simple task of not loading the geometry that is behind an object or that is small enough not to be visible, using its pipeline to quickly discard tasks that will not be useful for others that will always benefit user experience, that is, gaining in efficiency and performance.

Looking at the image above, we see how an 'Index Cache' has been added, which is basically a small amount of memory for instance geometry, that is, for objects or things that are repeated on screen over and over again, it prevents L2 cache memory is used and this information can be stored locally, gaining again in memory bandwidth and efficiency.

Memory and Delta Color Compression.

As with the veteran R9 285, a color compression system was released to save memory bandwidth, an ideal technique for gpus with 'little' bandwidth or not as large as in higher models.

The new RX 480 has a total of 256Gb / s of bandwidth, a figure much higher than those found within its same segment in past generations, such as the 380X that is around 180Gb / s but which in turn is a figure lower than 290 where AMD indicates that 'Polaris' is more efficient and has more effective bandwidth in practice due to increased bandwidth and color compression. This difference would reach up to 40% energy savings according to AMD.

This is due to the increase in L2 cache memory and its new manufacturing process in 14nm Fin Fet, but it really is due to its new compression system, having a ratio of 2/4/8: 1. Every time the data can be compressed, it is cached to a lesser extent saving energy, in addition to being compatible with high amounts of memory such as the 8Gb gDDR5 it has.

WE RECOMMEND AMD XConnect announced, desktop GPUs on your laptop

Connectivity and video.

Both the Nvidia 'Pascal' architecture and the new AMD 'Polaris' have updated all the connectivity of their cards, having up to 3 DisplayPort ports and an HDMi 2.0 rev.B compatible with HDR (High Dynamic Range) video, an available bandwidth of up to 18Gbps and with a maximum resolution of 4K @ 60Hz, resolution 4 times higher than the current standard.

As for audio, it is capable of taking out 32 channels for greater spatial immersion.

The DisplayPorts are updated to version 1.4 to guarantee better color using Display Stream Compression and thus bring very high resolutions such as 8K with a refresh rate of up to 60Hz and bringing 120Hz to 4K displays with HDR support.

The entire 'Polaris' range supports HDR, something really important to show better pixels, a wider color range and contrast. Throughout 2016 and especially 2017, monitors and TVs with support for this technology will come to the market, which is important especially for movie lovers since the Ultra HD Bluray standard will come along with them, also being compatible with HDR.

Not only moviegoers will enjoy this technology, but it also comes for the most demanding gamers!

In summary, we have an update of the more than well-known and off-road Graphic Core Next architecture, updating the most critical parts such as the geometry units that handle the tessellation, the most accurate HWS units for asynchronous computation instead of putting as many ACEs as well as saving in bandwidth. It is a gpu with a lot of potential and an affordable price for what are usually these types of cards, which we will soon begin to see the true performance of it with mature drivers that can take advantage of the technologies described here and the massive arrival of models' custom '.

These are the sections that we liked the most and we hope that many of the unknowns will be solved, but we leave you an image gallery with all the presentation of the architecture so that you have more detail than what has been analyzed so far.

Remember you can give your opinion to our review of the AMD Radeon RX 480, the new bombshell from AMD in mid-range and high-end graphics cards.

See you soon!