We’ve got some new intel on AMD’s next-gen GPUs, and it looks like they’re going to be packing a serious punch. According to recent leaks, the company has managed to increase its performance per watt by an impressive 54%.
So what’s behind this sudden boost in power? Part of the answer lies in Infinity Fanout Links,a chiplet connection technology that allows for better communication between different parts of the GPU. This should result in smoother overall performance and higher speeds.
There’s still no word on when these bad boys will hit store shelves, but we’ll be sure to keep you updated as more information becomes available. In the meantime, check out our article on RDNA 3 architecture for everything we know so far about these upcoming cards.
We didn’t see as many MCM desktop architectures as we expected. But there are some interesting details, like routing or access times, that give us a better idea of where AMD is headed.
There’s a problem with the MCM architecture for AMD RDNA 3 and Navi 31
The main problem with taking the memory controllers and Infinity Cache out of the main chip is access times. With any MCM architecture, you’re going to have latency issues when you have a single chip (GCD) connected to many others.
Navi 31 is the main chip of AMD’s Infinity Fanout Links technology, and it boasts some impressive features. According to AMD, the Navi 31 offers significantly lower latency than on-chip MCDs, as well as a frequency boost. This results in a noticeable performance improvement across the board.
Lisa Su plans to reduce latency by removing two critical units from main chip
Lisa Su is planning to reduce the latency created by removing two critical units from the main chip. The key, she says, is increasing the frequency of the Infinity Fabric and GCD Clocks by 43% and 18%, respectively.
There’s no such thing as a little bit when it comes to performance. So AMD knew they had to reduce latency when they improved their previous generation of GPUs. They’ve succeeded, as we can see.
Even with doubled core clock speeds, the Infinity Fabric could not fully reduce latency across the board. But compared to Navi 21, there’s been a 10% reduction in latency, which is a significant improvement. If AMD wants to approach the performance and latency of RDNA 2, it’ll need to build on that.
If you’re looking for high-performance connectivity
In order to achieve all of this, physical connections need to be made in the high-performance substrate. Accordingly, AMD has 9.2Gb/s IFOP links running at full capacity, which is a huge improvement from past models Ryzen and EPYC.
RDNA 3 doesn’t provide power efficiency data for these IFOPs, but EPYC and Ryzen 2 do, depending on the chip and architecture. Due to this fact, RDNA 3’s consumption and performance have not decreased, but have actually increased by 54% because this data must be present between matrix and matrix 0.2 pJ/b, explaining why performance and consumption did not decrease.
When compared to your version 2.5D’s Elevated Fanout Bridge, the Glink 3D’s bandwidth is almost five times faster, as we already discussed, it was specifically chosen because it’s enough for this chip and above all because its price tag is lower compared to a more powerful variant.
The Infinity Fanout Link: Better Routing for AMD
The Infinity Fanout Link has half the number of wires than the competition, at 25 against 50. Due to its matrix-to-matrix structure, it is harder to achieve energy leakage within the substrate.
AMD has had to reinvent itself in order to stay competitive. This is no easy feat, and AMD has managed to do it without increasing costs. Your connection bandwidth increased by 10x when you had half the wires in your routing. The increase in IF clock by 43% Hardly any compromise on array-to-array latency is also very impressive. Let’s hope that AMD continues to innovate and improve its architecture. If we’ve seen what they’re capable of so far, then there’s a lot more to come from them and it looks very promising indeed.