Its absence in the Android world has been taken advantage of by other players, such as ARM itself with Mali or Qualcomm with its Adreno. This prompted them to turn to other markets, such as the Chinese manufacturer Innosilicon, famous for its ASICs for mining, which not so long ago presented its Fantasy 1. This is the first card graphics based on a PowerVR since the early 2000s Kyro, but can they compete with NVIDIA and AMD in the PC space?
What is mosaic rendering?
In the late 1990s, graphics card designers had to struggle for performance with a common problem, lack of bandwidth. Graphics processors compared to what they are today were very simple. The first part of the 3D pipeline, before rasterization, was calculated by the CPU. The second part of the change was achieved by the graphics card, which required vast amounts of bandwidth that memory at the time could not provide without skyrocketing costs.
The solution offered by Imagination was rendering by Tiles, which still remains the basis of its architecture, so even today the Fantasy I once the geometry is calculated in the GPU itself, additional steps are added compared to a GPU classic. A Tile Renderer sorts the position of geometry in RAM based on its position in the scene just before rasterization to create individual display lists for each tile which it will then solve one by one during the rendering process.
Due to the small size of each block or tile, this allows it to be solved without having to access VRAM, since they use internal memory for this. This also makes it ideal for lazy rendering which often uses multiple frame buffers to calculate scene lighting. Its other advantage is that knowing the position of elements in the scene being essential to generate the spatial data structure for Ray Tracing, it is easier to implement Ray Tracing in this type of architecture.
However, this has two drawbacks. The first is that it requires more complex hardware than a regular GPU to achieve the same performance and therefore we will always get lower performance for a chip of the same size, the second is that the existence of high speed memory like GDDR or HBM eliminates its advantage in a gaming PC. This is why this type of architecture has become standard in handheld devices, where memory bandwidth for consumption reasons is limited.
PowerVR B-Series, the graphical architecture of Fantasy I
To understand the architecture of Innosilicon’s Fantasy I graphics cards, and incidentally also what’s inside Apple’s processors for its devices, we need to take a tour of the current Imagination architecture and although we know that it was recently presented C-series, also known as Photon, at present the most advanced devices use Imagination’s B-series as their architecture.
The heart of the B-series
The organization of each of these cores is as follows:
- Four USC blocks, Unified Shader Cluster, where each has up to 128 ALUs in FP32 for a total of 512 per core. Given the ability to execute an addition and multiplication instruction in a single clock cycle, it is capable of performing 1024 operations per clock cycle.
- 8 texture units, each capable of producing 4 texels, for a total of 32.
- 16 ROPS.
- 1 paving unit.
- 1 raster unit.
Each of the cores is exclusively responsible for one tile or block on the screen independent of the rest. Therefore, each of them has its own raster and tessellation units. In addition to carrying a small internal memory to resolve the frame buffer inside and reduce the impact on system RAM. However, this memory is used exclusively for ROPS and despite the advantages of the GPU, due to the huge texture maps in use today, it is necessary to access VRAM to obtain the texture data.
Fantasy I, the first chiplet GPU
The great novelty of the Imagination B-Series used in Fantasy I is the fact that it is the first GPU composed of chiplets, that is to say different chips that work together as a single processor. To do this, the list of screens is sent to the first of the four chiplets that make up the GPU, while the other three are subordinate. It is a solution very similar to the one that AMD has proposed in the patents with RDNA 3 and which will surely be common to all GPUs of this type in the future.
However, this solution differs on one specific point, the use of tile rendering to do what is pre-rendering and to be able to have several lists of screens not before rasterization, but from the beginning of the 3D pipeline. The concept is none other than rendering the scene without shaders or textures of any kind and from the computer pipeline and not from the graphics. This allows you to organize multiple lists of commands and not just one that will allow you to exploit the large number of cores when pre-rendering. This process happens automatically after the first GPU’s command processor reads the screen list.
This allows us to have multiple screenlists for the same scene that can be organized by the different cores. This is how it is achieved that with a configuration of 2 chiplets each is in charge of half of the screen, with 4 of them they are distributed over a quarter.
What has Innosilicon brought to your graphics card?
However, not all of the work was done by the folks at Imagination, but Innosilicon was the one to design the rest of the graphics card, adding the PCB design and choosing the rest of the materials. Where what stands out the most is the use of GDDR6 or GDDR6X memories depending on the model to be used, the support for DisplayPort 1.5 and HDMI 2.1, but above all the use of its Innolink technology, which has been designed to internally communicate the four chips that are part of the GPU.
Specifically, we have two different variants, the so-called Type A can reach 5 TFLOPS of power in FP32, it has a memory interface with 128-bit GDDR6X VRAM at 19 Gbps with a bandwidth of 304 GB/s. Type B, on the other hand, has two full GPUs and hence is made up of 8 chiplets in total and doubles the number
The Innosilicon Fantasy I are not for your PC
The reality is that you won’t be able to buy Innosilicon’s Fantasy I graphics cards to use in your gaming PC, nor would you be interested in that, because Imagination designs its architectures for handheld devices where Windows isn’t not the dominant operating system and neither is it. is DirectX, because we find a series of shortcomings. It makes no sense to add features to your hardware that your customer won’t use and the biggest customer for these GPUs, albeit secretly, is Apple and more specifically its Metal API.
Ironically, PowerVR is so tied to Metal, the API used in iOS, macOS, and the rest of Apple’s operating systems, that in the end the folks at Tim Cook ended up signing a deal with Imagination. so that they can continue to develop the integrated GPU in their processors. So in the current Apple A15, M1 and its Pro and Max variants, what’s inside is a PowerVR. The counterpart of this is that those in Cupertino have created the general idea that they are so powerful that they can create all the hardware in a system and compete for resources against the entire world. The reality is very different.
It may surprise us that a 4-chip GPU can hit 6 TFLOPS when the PC input range is already hitting that, but keep in mind that this is a design built for mobile processors, but for the purpose of achieving cloud computing and not for use in a gaming PC.
Designed for data centers and cloud computing
Let’s not forget that in servers it is normal to use several processors and that we have more and more servers based on smartphone processors. Let’s also not forget the tendency to virtualize a graphics card in the cloud for several customers, by its nature the Fantasy I does not require virtualization, each of the chiplets that compose it can function as a small GPU.
So we have an architecture that derives from mobiles and evolves to datacenters, but without going through the neighborhood that is the PC. This means that a series of features that are essential for PC games today are missing. That’s why, despite the Fantasy I looking like it might be reminiscent of a Gaming GPU or not looking serious with these colors, they’re really for cloud computing, even though this is a first generation. Are we facing a future where the graphics card is not in the hands of the user, but in the server?
In any case, China as a rival superpower of the United States must be completely independent from a technological point of view and that means creating its own solutions outside the classic solutions of NVIDIA, Intel and AMD, which we remember are US companies.