Lisa Su has appeared on stage to start the presentation of the new AMD RX 7000 based on RDNA 3 architecture, the presentation is focused on gaming hardware. A little jab at NVIDIA that has been focused on Artificial Intelligence for a long time. In addition, the AMD CEO is bragging about the number of platforms that have adopted the company’s technologies.
The new generation of AMD Radeons are based on the RDNA 3 architecture, where AMD has focused especially on improving performance per watt. We already knew that they had discarded the new 16-pin power connector to continue with the classic connectors. So for them it is important to increase this aspect compared to the previous generation, in total they say that the generation gap is 54%. On the other hand, the rumors about the graphics chip, or GPU, are confirmed in the sense that it is a disaggregated chip, which is made up of 7 different chiplets in its best configuration.
GCD (Graphics Compute Die)
It is the central chjp of the RX 7000 graphics cards, where the entire graphics chip is located, except for the memory controller and the Infinity Cache, manufactured at 5nm in the TSMC foundries. Its size is 300 mm² in case of the RX 7900 Navi 31 chip, but it will be smaller in lower configurations.
MCD (Memory Cache Die)
The Memory Cache Die, on the other hand, will be common in all configurations of the RX 7000 graphics card range, but in a different number, since it will depend on the memory configuration. Since in each of them there is un 64-bit GDDR6 controller and GPU L3 cache, dubbed Infinity Cache. In this aspect we have seen a cut in terms of capacity, from 32MB per memory controller to 16 MB. So instead of having 192 MB we have 96 MB with 6 CDM. Of course, with a much higher bandwidth, reaching 5.3 TB/s of bandwidth between GCD and MCD.
What are the early models of the RX 7000 graphics cards?
What is surprising this time is that AMD has not presented any graphics card under the name RX 7800, which means that they will use the Navi 32 chip, while those that have presented make use of the most powerful one under the new RDNA 3 architecture, the so-called Navi 31. The models presented are the AMD Radeon RX 7900 XTX and the AMD Radeon RX 7900 XT. Both are actually the same graphics card, but with some cuts in the second case.
If anything, the clock speed is lower than expected. And yes, we know that AMD has decided to give as a figure to calculate the power of its graphics card a more real and sustained over time, such as the Game Clock, but we are surprised that they have not used the speed of Boost. We believe that the failure to achieve a high clock speed is a consequence of separating the chip into several parts and not being monolithic. And no, it is not a problem of latency, but rather of increased consumption when intercommunicating the GCD with the MCD, especially the L2 cache with the L3 or Infinity Cache.
AMD Radeon RX 7900 XTX
Unlike NVIDIA, AMD has decided not to use the 12VHPWR connector for its new top-of-the-range graphics cards, this means that they have had to opt for a classic configuration with 2 x 8-pin connectors, which give it 300 W of power, while the other 55 W come from the PCI Express bus connector itself. This is the full version of the chip when using the 96 Compute Units of the Navi 31 chip and a 6 CDM setupwhich also gives you a 384-bit GDDR bus for 24 GB of memory
The fact that it is so close to the theoretical limit in terms of consumption makes us think that a future 7950 XTX is going to need a third connector to increase the clock speed in a future model. Its clock speed is 2.3 GHzwhich means that it reaches 61 TFLOPS of 32-bit floating point power compared to 23 TFLOPS of the RX 6950 XT. A more than considerable performance increase that all games will benefit from.
Their release date is December 13 and the price, in the absence of confirmation for Europe, is $999.
AMD Radeon RX 7900 XT
As for the second model, it is a shortened version of the first, since it does not use the extra power of the PCI Express bus, using only the power of the double 8-pin connector and remaining in 300W TDP. Which translates into a Game Clock of 2 GHz and a configuration of 84 Compute Units in total. so we talk about 43 TFLOPS, almost twice as much power as the RX 6950 XT. Another of the cuts is to have 5 CDM for a 320-bit GDDR6 bus Y 20GB of video memory.
As for its price, it is $899, so it would be, at least in the United States, at $899, which is the price of the canceled 12 GB RTX 4080, which it would surpass in all aspects. How much will it cost us in Europe? We do not know, for the moment we can only say that the launch date coincides with its older sister, the December 13th.
New graphics processing cores
The Compute Units are the equivalent of the processor cores and are in charge of executing the shader instructions that graphics programs have and that allow advanced visual effects to be made. In RDNA 3 AMD has made substantial changes, but has not revealed all its cards in the presentation, we believe that it is rather due to lack of time and the fact that these technical details are not usually given in these presentations.
Improved SIMD units
A trick that NVIDIA makes is to count the power of its floating point units as if they were working all the time performing calculations in floating point, which is unrealistic in real scenarios, since calculations are also performed with integers and both the RTX 30 and the RTX 40 have the second floating point unit switched with the integer one. What does this have to do with RX 7000 graphics cards? Well, AMD has decided to opt for a different solution.
For the RX 7000 graphics cards they have decided to go back to the pre-GCN era and make each of the 64 Stream Processors in a Compute Unit able to execute two instructions at the same time, so that’s like 128 units, but this depends on the how these are packaged, which is the job of the driver. Thus, there will be sets of instructions that can go together and others that cannot. Therefore, the maximum power in floating point, as in the RTX, is conditional.
Finally, units for the AI
Or what is the same, systolic arrays or tensor units. Which means that it is very likely that AMD will start applying Deep Learning algorithms in games like NVIDIA does. At the moment they have not revealed much information about these units. But it is something that we did not expect and honestly it has been a very pleasant surprise. We wonder if the also announced FSR 3 will capitalize on these units. In any case, there is no doubt that AMD could not be another generation without them.
Improved Ray Tracing
In this case we are observing how the trees do not allow us to see the entire forest, many are looking at the fact that the increase in performance is only 50% in terms of the number of rays per Compute Unit that the new Compute Units they can process, but the devil is in the details and we were delighted to see that they added data structure traversal capabilities that were missing from the RX 6000.
Unfortunately this improvement has not been reflected in the benchmarks they have shown, which are still games optimized for the RX 6000 in Ray Tracing. That is why, while waiting for the graphics card to be in our hands and we see a lesser dependence on the SIMD units, we refrain from making judgments about the performance of the RX 7000 graphics cards in terms of Ray Tracing.
Improved video codec and display driver
Outside of the Compute Units we have key elements in every graphics card such as the display controller and the video codec. Two parts in which AMD has taken a giant leap and has once again positively surprised us:
- The new display driver is intended to take advantage of the Display Port 2.1and its high bandwidth that allows high resolutions at huge refresh rates and with HDR. At the moment there are no monitors on the market that support them, but in this aspect they have been far ahead of NVIDIA.
- As for the video codec, we now have the added ability to encode and decode videos on the AV1 format up to 8K 60 FPSplus AMD has added cadvanced AI video acceleration capabilities. What is a forbidden way of talking about technologies that we have already seen on televisions such asframe interpolation or the automatic resolution increase.
AMD RX 7900 XT performance in games
What interests us about a gaming graphics card is its performance in games and it can be said that AMD has shown muscle with its RX 7900 XTX when it comes to showing its power under different resolutions, titles and game conditions. Obviously it doesn’t compare to the RTX 4090, but it’s normal to have a much simpler configuration and cost half the money. It’s a shame we don’t have the RTX 4080 on the market yet, which these early RX 7000 graphics cards compete with.
In any case, and as we have already mentioned before, the performance in terms of Ray Tracing is something that we are going to take with tweezers, more than anything due to the fact that in the face of the RX 7000 the games in their code optimized for AMD require be adapted to take advantage of the new unit for Ray Tracing. So the performance figures that we see on the slide are much lower than the final performance that we will see in a few months.
Puyitas of AMD made the NVIDIA DLSS 3
Those of the RTG have not been throwing poison darts against NVIDIA and its DLSS 3 with the ability to interpolate frames to generate new ones automatically. Especially in games designed for eSports with so little graphic load that they achieve hundreds of frames per second without any effort, only to later find that even the most powerful monitors cannot display them as they do not have an equivalent refresh rate.
Support software and drivers on RX 7000 graphics cards
AMD has greatly improved its driver support in recent years, which is why AMD is making a presentation on its Adrenalin hardware to configure the graphics card, games and applications. At the same time, there has been talk of the news that they are going to incorporate into their RX 7000 graphics cards.
- AMD HYPR-RX: a solution designed to reduce latency and improve frame rate. Dying Light 2 will be the first supported game.
- Integration of the hardware video codec of the GPUs under RDNA 3 architecture in OBS. Being one of the points for a long time that AMD had to solve and they have finally done it. Applause from here for them.
- AMD Smart Access Video,that allows to share the work between the integrated graphics of the Ryzen processor and that of the Radeon graphics card
2 thoughts on “AMD presents its new and powerful graphics cards”