The first Radeon RX 8000s will have memory capacities of 16GB, 16GB and 12GB
This week, the first leak of a Radeon RX 8000 in the Geekbench database appeared online. It could have been an ES sample of this GPU or a similar early prototype, so it’s unclear how close these GPUs are to launch. But this seems to have kicked off leaks for this new generation of AMD graphics cards. A bit more detail has now surfaced on the web about what to expect from these GPUs, especially the more powerful mainstream Navi 48 GPU.
The leaker with the handle Kepler_L2 commented on the Radeon RX 8000 on Twitter, and according to them, AMD is apparently planning three cards based on Navi 48, the larger of the two RDNA 4 chips, which has a total of 64 CUs (4096 shaders) and a 256-bit memory bus. It should also be the first of the GPUs based on the RDNA 4 architecture to come to market.
Three models with a larger RDNA 4?
AMD will offer two configurations with the full (i.e. not cut down) 256-bit memory bus. The first of these is supposed to use GDDR6 chips with an effective clock speed of 20.0GHz, which would give a bandwidth of 640GB/s (this is an improvement over previous rumors that talked about memory with a maximum of 18.0GHz). It’s likely that this card will also have 16GB of VRAM (and its Infinity Cache should be 64MB). And it should probably use a fully enabled version of the chip with all 4096 shaders / 64 CUs enabled.
However, according to what Kepler_L2 states, the 256-bit configuration will be retained for the second model as well, which should have a cut down number of compute units. This graphics card would use memory with a slower effective clock speed of 18.0 GHz , thanks to which it could be cheaper. That would give a bandwidth of 576 GB/s, and the model would retain the 64 MB Infinity Cache. It is possible that this is exactly the model that appeared in Geekbench. There, the leaked card did carry 16 GB of memory, but the GPU was cut to 3584 shaders (56 CUs). Thus, the other slower 256-bit/18.0GHz configuration could have 56 CUs.
The third configuration already has 192-bit memory, which means a capacity of 12 GB. Also, the Infinity Cache capacity should drop to 48 MB, in proportion to the bus width. To compensate for the drop in bandwidth a bit, a higher effective memory clock speed will be used compared tothe middle model: 19.0 GHz. This would give a bandwidth of 456 GB/s. The GPU will be cut-down as well, but we have no indication on the number of units yet – it could possibly be something between 3072 and 3584 shaders.
No one knows what names will these cards come to market with, at this time. It is possible that the first will be labeled Radeon RX 8800 XT and one of the other two RX 8700 XT, this designation would fit the current scheme. However, then there would be a problem with naming the third card (RX 8800 or 8700 without XT?). Perhaps, the first model could e called RX 8800 XTX, the second RX 8800 XT and the third card with 12GB memory would be Radeon RX 8700 XT. Alternatively, the middle card with 16GB capacity but slower memory could be something like the RX 8800 GRE (and the 12GB card the RX 8700 XT).
Cards with Navi 48 and Navi 44
As another source of information, BenchLife states that it knows of four configurations of RDNA 4 GPUs so far:
- R24D-E6
- R24D-E8
- R25D-P4
- R25D-P8
But of that list, only two items starting with R24D should probably represent graphics cards with the more powerful Navi 48 GPU. So either the codes for the two 256-bit versions are not different, or this list is incomplete.
The two entries starting with R25D should instead represent cards based on the Navi 44 chip, which is roughly half the size, containing 2048 shaders (32 CUs) of RDNA 4 architecture with a 128-bit bus and 32MB Infinity Cache. Graphics cards like the Radeon RX 8600 (and possibly the RX 8600 XT) could be based on this GPU.
- Tip: RX 8000 Radeons said to have all-new ray tracing accelerators
- Tip: RDNA 4 said to mostly be patched RDNA 3, RDNA 5 completely new
Better ray tracing performance, AI?
According to the video from Moore’s Law is Dead, which we mentioned in our last report, the TDPs should supposedly be similar to the last generation, and the company will officially list the fact that no one will have to change their PSU as one of the advantages (apparently the cards will be compared to the RX 7700 XT and RX 7800 XT cards). But exact TDP figures are still missing. If AMD presents the RX 8000 generation like that, it will probably still use power delivery via classic 8-pin cables instead of the trouble-ridden 12VHPWR/12V2x6 connectors.
Tip: 12-pins on GPUs melting again: Both fans and critics are wrong
According to Moore’s Law is Dead, RDNA 4 should vastly improve performance in raytracing effects. They say this could be done through the integration of multiple ray accelerators (i.e. dedicated ray tracing acceleration units) into each of the CU blocks, whereas until now AMD architectures (RDNA 2, RDNA 3, RDNA 3.5) always had a single ray accelerator per CU. This is also the case with Nvidia architectures, which similarly use a single RT core per SM block. According to MLID, the CU blocks of RDNA 4 architecture could even contain “matrix hardware”, i.e. dedicated units for AI acceleration.
However, it is questionable whether to trust this leak. Even if it is legitimate information, it may not be accurate in all respects. For example, MLID states that the Infinity Cache of the Navi 48 chip could be as high as 96 MB. However, due to the 256-bit bus width, as well as the GPU’s price point, 64 MB should be expected, and indeed that’s what other leaks are also stating.
The rumored increased performance in ray tracing could be achieved not by adding units, but by enhancing the capabilities and performance of each ray accelerator. Similarly, the presence of dedicated units for AI acceleration is still not confirmed for the RDNA 4 architecture. Until now, AMD has followed the philosophy of sharing the hardware for these operations with the main compute shaders. However, certain forms of execution unit sharing are not entirely mutually exclusive to the performance of AI matrix operations being increased.
Unfortunately, there is still uncertainty as to when these graphics cards will be released. It was mostly assumed recently that it wouldn’t be until 2025, but Moore’s Law Is Dead is now saying that the release could be as early as October or November (which is a bad news for Intel, who was likely looking forward to the opportunity to release its own Battlemage GPUs in advance before AMD’s new GPUs). It is not yet clear how realistic such an early launch is, though.
Sources: VideoCardz, Kepler_L2 (1, 2), Moore’s Law Is Dead
English translation and edit by Jozef Dudáš
⠀
