RDNA 3 and chiplet design: AMD introduces first two Radeon RX 7000 cards

Under the motto “We move forward together”, AMD introduced Navi 3x GPUs based on the RDNA 3 architecture. AMD is currently in tightly packed weeks of new releases: The Ryzen 7000 processor (tested)Now RDNA 3 and the next Radeon RX generation, EPYC processors with Zen 4 cores next week and then mobile next-generation Ryzen processors in early January.

But now for the Radeon RX 7000 generation graphics card. The rDNA-3 generation should achieve at least 50% better performance per watt ratio. First, AMD addressed the fact that the Navi 3x series is a chiplet design. The package includes a central graphics chiplet die (GCD) and several memory chiplet die (MCD). GCD is produced in TSMC at 5 nm, MCD in 6 nm – also in TSMC. These are not just parallel Ryzen processors.

Each MCD has 16MB of Infinity Cache, giving the Navi 31 GPU a total of 96MB. For comparison: Navi 21 needs 128MB Infinity Cache. But AMD has developed the second generation Infinity Cache which is said to have a higher hit rate. This compensates for the slightly smaller capacitance. There are 58 billion transistors in one GCD and six MCDs. The GeForce RTX 4090’s AD102 GPU has 76 billion transistors and requires a chip area of ​​608.3 mm².

Between the central MCD and the six MCDs, the Infinity Fabric interconnect achieves a bandwidth of 5.3 TB/s. Memory controllers for GDDR6 memory are connected to the MCD. Radeon RX 7900 XTX has 24GB, Radeon RX 7900 XT 20GB.

Unlike NVIDIA’s current Ada Lovelace generation, the new Radeon cards support DisplayPort 2.1. AMD also never tire of stressing that the top model, the Radeon RX 7900 XTX, should only consume 355W. Also, no new electricity connection is required. The two new cards come with two 8-pins. The Radeon RX 7900 XT will settle for 305W.

The Radeon RX 7900 XTX and Radeon RX 7900 XT are the first two models of this new generation.

The dimensions for AMD’s reference versions are 287mm in length and 2.5 slots thick. AMD also wants to differentiate itself from the massive dimensions of the GeForce RTX 4090.

AMD was still cautious about performance. The Radeon RX 7900 XTX is up to 70% faster than the Radeon RX 6950 XT in UHD. In particular the ray tracing performance should improve. The level of the GeForce RTX 4090 will probably not be reached. But the target should be the GeForce RTX 4080, which NVIDIA will launch in two weeks. We’ll probably have to wait for independent and detailed tests.

rDNA-3 architecture

One of the most significant changes to the rDNA 3 architecture is related to the Workgroup Processor (WGP). FP32 arithmetic units can now perform a “dual problem”—two rather than just one arithmetic command. To do this, however, these computation orders must be the same, to a certain extent, in a GPU with associated computations as well. In the ideal case, a computing unit (CU) can thus execute 128 instead of the previous 64 FP32 arithmetic commands.

To be able to make better use of the processing units, AMD has increased the VGPR (Vector General-Purpose Register) by 50%. Each compute unit also has two unspecified AI units that can rapidly perform simple INT8 and BFLOAT16 calculations. These AI units are probably not comparable to Tensor Cores, and yet AMD is looking to increase the throughput of related computations by a factor of 2.7.

Improved ray tracing performance is achieved by allowing the relevant units to calculate 50% more rays per clock. The BVH structure of each ray tracing calculation is still calculated in shader units. Above all, ray box sorting and ray traversal are accelerated through RT units.

Separate clock for front and rear end

Rumors talked about GPU clock up to 3 GHz. The Radeon RX 7900 XTX didn’t achieve this and falls short of expectations in this regard, even at 2.3 GHz. One of the special features of Navi 3x GPU is that they can work with different clocks for front and back end. While shader units operate at 2.3 GHz, the front end can also run at 2.5 GHz.

In a way, the front-end determines the maximum possible computing power, while the shader units can operate at a lower clock speed to save power. AMD wants to be able to save 25% in power consumption with this.

Apart from the hardware, there is also the first announcement regarding the new software functions. The FSR 3 is said to offer better image quality and also uses a floating component. The exact details are still pending here. FSR 3 is to be officially launched in 2023.

In stores from 13 December

Both graphics cards should be available by December 13th. The Radeon RX 7900 XTX will cost $999 and the Radeon RX 7900 XT will cost $899. We don’t know any value of the euro yet. In addition to the reference design from AMD, custom models are also set to debut on this day. Then we will know what the GeForce RTX 4080 is capable of.

AMD’s program is still ongoing. We will update this notice continuously.