Amd kaveri features: gpu and compatibility (part ii)

And we come to the second part of this interesting article, where we are going to focus on the third and last great novelty of Kaveri, its integrated GPU.

With the announcement of AMD “Berlin” (upper image), its server environment Apu, the number of Shaders and therefore the GPU that the desktop versions will inherit, such as Kaveri, is practically confirmed.

Unlike current Apus, this is going to be the first to integrate GCN architecture that, as you know, is the one used in the 7000 series of desktop graphics. The first Apu “Llano” was made up of 400 VLIW5 architecture shaders, an architecture that is very old and seen from the HD2000 series to the 5000 series that gave us such good results. Trinity and later Richland, integrated the improvement of the previous architecture, now called VLIW4 that have up to 384 Shaders and that we also saw in the high-end graphics of the last generation, HD6900 series.

We leave you a drawing to see the difference in size that the integrated graphics of these Apus have been suffering.

Leaving behind a brief historical mention, we are going to explain a little more carefully what this new architecture consists of and how Kaveri will integrate it.

Unlike VLIW4 / 5, GCN is a modular architecture made up of Compute Units (CUs) and in each CU we find 64 Shaders, 4 Tmus (texture units) and a certain cache memory for computing.

CUs form groups of up to 4, thus forming a Compute Unit Array. If we have multiple Arrays in conjunction with other units such as the UTDP (Ultra Threaded Dispatch Processor), ACE (Asynchronous Compute Engine), GCP (Graphics Command Processor) together with the memory controller and rendering blocks that make up 4 Rops and 8 Pixel Pipelines, this is how we obtain a graph based on GCN architecture.

The Kaveri GPU is going to be a little different since it will be based on Sea Islands which is the second version of Southern Islands (the first one based on GCN), and we will explain the differences that we are going to find.

Now, Compute Unit Arrays are no longer used, but have been changed by DDP Arrays (Data-Parallel Processor). These are computational units made up of multiple CUs, which have their own memory interface and which work with the UTDP (Ultra-Threaded Dispatch Processor) to be more efficient in the simultaneous execution of different types of operations and workloads.

DDP Arrays are capable of executing multiple intensive general-purpose calculations, such as graphics or computation, simultaneously and independently.

Each Data-Parallel Processor Array is capable of running multiple intensive general-purpose calculations (computational, graphical, Boolean - represents binary logic values ​​- among others) simultaneously and completely independently.

The GCP (Graphic Command Processor), which has been replaced by the Command Processor, has also been removed. This CP is a unit that is in charge of managing the commands sent to the GPU via hardware interruptions, that is, IRQ, to ensure its operation and execution speed. Now we leave you the scheme with the mentioned novelties.

Also this new evolution of the GCN architecture brings with it other changes (standard operation for HSA, bi-directional access with coherence…), but we are going to focus on the GPU of Kaveri, which not many days ago was finally practically unveiled.

Codenamed " Specter ", this new integrated GPU will be made up of 2 Data-Parallel Processor Arrays, and that each Array will have 256 Shaders distributed in 4 SIMDs, and that will finally give the final amount of 512 GCN Shaders. In turn, it will have 32 texture units (Tmus), a unit for tessellation and figures that have not yet been filtered or confirmed in terms of number of rendering blocks, although it is speculated that it could have 2, thus leaving 8 Rops and 16 Pixel Pipelines.

Specter is the code name that makes up Kaveri's most powerful GPU, although as it happens in current and previous versions of Apus, there will be more trimmed GPUs of which its name is also known, Spooky (which will have either 256 or 384 Shaders).

WE RECOMMEND YOU The new HDR10 + imaging standard will debut this month

It really looks a lot, in distribution, number of Shaders and others, to the desktop version 7750, a graphic that also integrates 512 GCN Shaders although based on the first generation, Southern Islands.

New socket, chipsets and other curiosities

* Kaveri will have PCI Express 3.0, made up of 24 PCIE 3.0 lines and also has a Unified Media Interface bus, made up of 4 PCIE 3.0 lines to communicate directly with the chipset.

* It will also release new chipsets (FCHs) with the name A88X and A78 (called Bolton D4) and the only thing that is known to date is that it will have in the case of the A88X up to 8 Satas 6Gbs (Sata 3), unlike the A78 that will integrate only up to 6 Satas 6Gbs. Of course both will have an integrated USB 3.0 controller.

* Unfortunately not everything is gold what shines, and it is that it will be necessary to change which socket, again, to be able to enjoy Kaveri, leaving socket FM2 with a processor support until Richland, since physically it is impossible to mount a Kaveri on FM2 (pin location). However, the new socket FM2 +, will be compatible with Richland and earlier since it can be.

As we saw in the transition from AM3 to AM3 + for the FX series, it inherits the characteristic black color of this socket. In the previous Computex we could see the new Asus board, with A88X chipset, tremendously similar to the F2A85M-Pro and which was the first to be seen.

And here we have come, naming each and every one of the new features that surround Kaveri, the new Apu and the true coexistence between a CPU and a GPU.

It is expected to be ready for the last quarter of the year, between October and December if it is not delayed or if there are no changes to its Roadmap. Until then and for when we have reliable data on their frequencies, final code names and models, we say goodbye for now.

Thanks for paying attention to this reading!