In order to make graphics cards more and more powerful, there are essentially two pillars: more computing cores and (or or) higher clock rates. We leave out advantages such as increased performance per clock cycle based on improvements in the general architecture.
Both the number of processing cores and the clock rate depend very much on the reduction in the structure width. The former because simply more arithmetic units can be accommodated per unit area, the second because these arithmetic units switch faster due to the reduction in size.
Hardware guides on GameStar Plus
You don’t want to wait for a graphics card revolution, but want maximum performance now? We’ll tell you what you have to consider before upgrading to an Nvidia Geforce RTX 3000 or AMD Radeon RX 6000.
Still, graphics chips tend to get bigger and therefore more and more expensive. This is also based on two pillars: larger chips reduce the yield on a silicon wafer and are at the same time more susceptible to damage.
The solution to these problems is chiplet design. A new AMD patent now gives hope that we may see modular GPUs in the near future. But chiplet designs are still very complex and therefore costly.
What are chiplets?
Chiplets are small chips that are interconnected to form a large chip array, i.e. a large chip. AMD is already doing this with its Ryzen processors (CPU) and is known to be very successful with it. However, this step is much more difficult to take with graphics cards.
What are the problems with the chiplet design?
In short: the signal transit times, also called latencies. Since GPUs, unlike CPUs, do not work sequentially but in parallel, ideally all cores must be able to access the memory at the same time. In the case of chiplets, however, this is made more difficult because the individual modules also have to be connected to one another.
How does AMD intend to solve the chiplet problems?
So that applications, i.e. games, can communicate with the graphics card as usual in the form of a single, monolithic GPU, only one chiplet should be addressed by the CPU.
For information: The CPU issues so-called draw calls (drawing commands) to the GPU, which, by the way, can often be the limiting factor with regard to the frame rate, especially at low resolutions – keyword CPU limit.
According to the patent, all other modules should be connected to the first chiplet (primary chiplet) via crosslink with a high bandwidth. Specifically, the HBX (High Bandwidth Crosslink) is supposed to connect the L3, also called last-level caches, of the individual chip sets.
When can we expect chiplet-based graphics cards?
AMD has not yet made any announcements in this direction. At Nvidia, the graphics chips based on the hopper architecture are to be based on a chiplet design. According to current information, the Lovelace architecture will appear before Hopper.
As of now, it looks like we’ll be seeing chiplet GPUs within the next two generations of graphics cards, i.e. the next four to five years.
The editors-in-chief of GameStar, GamePro and Mein-MMO are looking less far into the future. What we expect for the game year 2021:
Our forecasts for 2021 – what do the editors-in-chief of GameStar, GamePro and Mein-MMO expect?