Tomorrow, November 3, is the definitive presentation of the first graphics cards under the RDNA 3 architecture, which will supposedly come out under the name of RX 7000, not to be confused with Ryzen 7000. We do not know what models exist at the moment, only the top of the range through rumors and leaks. However, what is interesting about the new architecture is that it is going to have much more profound changes, even than when the jump from GCN to RDNA was made. So much so that it could be considered a new architecture and much superior to the current one. What do we know?
The GPUs with first-generation RDNA architecture were based on achieving the necessary jump in terms of performance per watt, for this the Radeon Technology Group remade a good part of the graphics chip in order to achieve an efficiency close to the GTX 10, but with much more raw power, although lacking the news of the RTX 20. For RDNA 2, the Infinity Cache, exclusive to PC, and support for DX12 Ultimate technologies such as Mesh Shaders, support for Ray Tracing in real time and the Variable Rate Shading. As for RDNA 3, we know that it will be the first disaggregated GPU in history, which does not mean multi-GPU, but it also hides other secrets, some of which we are going to reveal to you.
What news will we see in the RX 7000?
Since the cost of each wafer in TSMC’s 5nm node is very high, around US$18,000 each, AMD has decided to follow the same procedure that they started with the Ryzen 3000, separating a single chip into several different. In this case we have a Graphics Core Die in the center, being the largest chip and on the other several Memory Core Die placed on its periphery, the little we know is the following:
The Graphics Core Die it is the largest chip.
- It includes the entire GPU except the Infinity Cache and the memory controller with the GDDR6.
- Manufactured at 5nm.
- Each RDNA 3 family has its own GCD.
Memory Core Die.
- 16 MB of Infinity Cache per MCD, which is cumulative between all MCDs.
- It is said that there is a version with a larger amount using V-Cache in the future.
- Manufactured under TSMC’s 6nm node.
- The number of MCDs per chip is the length of the GDDR6 bus divided by 64. So an RDNA 3 graphics card with a 384-bit bus will have 6 MCDs.
This could be the range of graphics cards RX 7000
It is known whatand Navi 33 (32 Compute Units)the most modest chip in the known range, so far, is made of monolithic form under the 6 nm node. While his older brothers, Navi 31 (96 Compute Units) Y Navi 32 (60 Compute Units) they will be disaggregated chips.
|Chip||CDM number||buses||Amount of VRAM|
|Navi 31||6||384 bit.||24GB|
|Navi 31||5||320 bit||20GB|
|Navi 32||4||256 bit||16 GB|
|Navi 32||3||192 bit||12GB|
|Navi 33||Monolithic||128 bit||8GB|
The models in the table should be the ones we should see at least throughout the deployment of the generation, not counting other versions with fewer active Compute Units of each of them. From each segment of the range we should see the chip At the moment the images of a top-of-the-range model that will use the Navi 31 chip have been leaked, where you can see how they maintain the triple fan configuration and a power supply made up of 2 8 pin connectors. This marks a consumption of 375 W. What’s more, it is said that theThe maximum power used by the standard model would be 350 W with higher performance per watt than its NVIDIA rivals.
As with the RX 6000, AMD will start with the top-of-the-line models, so we should expect at least three models to start with. If what has been leaked is true, then we would be talking about the RX 7900 XTX with 24GBthe RX 7900 XT with 20GB and probably a third, obviously just the RX 7900. Of the three models we know the configuration of the most powerful, but not of the other two.
Changes in the Compute Units of the RX 7000
The most important change is in the cores of the GPU that presides over the graphics card, so much so that it can be said that AMD could give RDNA 3 a different name, due to the fact that the only time in the last decade when they made such a leap it went from the RX Vega to the RX 5000, that is, from GCN to RDNA. What news are we going to see? In RDNA 1 and 2 there is the WGP concept where two Compute Units are unified in some points, in RDNA 3 they have decided to go further and unify them into one. Which results in a new structure:
- 128 calculation units divided into 4 groups of 32 each. That is, twice as much as in the current generation of AMD.
- A single texture unit, capable of managing 4 texels per clock cycle, there is no change in this.
- Maintains capabilities for calculating Ray Tracing intersections of RDNA 2. Obviously, this part has been improved compared to the current generation.
- Double L0 data cachewe do not know if AMD has increased the number of access channels to reduce contention.
- A new unit called Ray Arbiter Unit to speed up Ray Tracing, which would add the ability to traverse the data structure without needing a shader program to do so. With this addition, the jump from RDNA 2 to RDNA 3 in terms of Ray Tracing is expected to be much higher than what we have seen from RTX 30 to RTX 40.
New Command Processor
We will not see this announced by AMD among the technical specifications of its RX 7000, but it is important for users because they are going to make the leap to a greater reliance on drivers for performance. AMD’s strategy with previous generations was to use an intelligent hardware manager, this meant wasting a lot of transistors to avoid having to make a complex driver to do the job.
However, hardware managers have a limit in terms of efficiency, they can not scale more and increase the consumption and cost of the GPU. If we add to this that from the RTG the batteries have been put on in terms of the quality of the Adrenalin graphics drivers, then we have the perfect excuse for said change. The idea is none other than to reduce the central area of the chip and have greater control of performance through the release of new drivers.