The specifications of the entire Navi 2x lineup have been leaked. These include Navi 21 (Big Navi), Navi 22 (RX 5700 XT successor), and Navi 23 which will be the budget offering. Starting from the top, we have Navi 21 which will feature 80 CUs or 5,120 shaders as already reported with sixteen L2 cache tiles meaning a bus width of 256-bit. The raster backend (ROPs) count has been cut from eight in Navi 10 to four in Navi 21.
| Property | Navi 10 | Navi 14 | Navi 12 | Navi 21 | Navi 22 | Navi 23 | Navi 31 |
|---|---|---|---|---|---|---|---|
| num_se | 2 | 1 | 2 | 4 | 2 | 2 | 4 |
| num_cu_per_sh | 10 | 12 | 10 | 10 | 10 | 8 | 10 |
| num_sh_per_se | 2 | 2 | 2 | 2 | 2 | 2 | 2 |
| num_rb_per_se | 8 | 8 | 8 | 4 | 4 | 4 | 4 |
| num_tccs | 16 | 8 | 16 | 16 | 12 | 8 | 16 |
| num_gprs | 1024 | 1024 | 1024 | 1024 | 1024 | 1024 | 1024 |
| num_max_gs_thds | 32 | 32 | 32 | 32 | 32 | 32 | 32 |
| gs_table_depth | 32 | 32 | 32 | 32 | 32 | 32 | 32 |
| gsprim_buff_depth | 1792 | 1792 | 1792 | 1792 | 1792 | 1792 | 1792 |
| parameter_cache_depth | 1024 | 1024 | 1024 | 1024 | 1024 | 1024 | 1024 |
| double_offchip_lds_buffer | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
| wave_size | 32 | 32 | 32 | 32 | 32 | 32 | 32 |
| max_waves_per_simd | 20 | 20 | 20 | 16 | 16 | 16 | 16 |
| max_scratch_slots_per_cu | 32 | 32 | 32 | 32 | 32 | 32 | 32 |
| lds_size | 64 | 64 | 64 | 64 | 64 | 64 | 64 |
| num_sc_per_sh | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
| num_packer_per_sc | 2 | 2 | 2 | 4 | 4 | 4 | 4 |
| num_gl2a | N/A | N/A | N/A | 4 | 2 | 2 | 4 |
| unknown0 | N/A | N/A | N/A | 10 | 10 | 8 | 10 |
| unknown1 | N/A | N/A | N/A | 16 | 12 | 8 | 16 |
| unknown2 | N/A | N/A | N/A | 80 | 40 | 32 | 80 |
| num_cus (computed) | 40 | 24 | 40 | 80 | 40 | 32 | 80 |
The flagship Navi 21 will supposedly have a clock speed of up to 2.2GHz, resulting in a single-precision compute performance of 22.5 TFLOPs, nearly twice as much as the Xbox Series X. All this at just 238W. It is still lower than the 29.8 TFLOPs rating of the RTX 3080, but the latter does consume more more than 350W under load (100W more). Regardless, I wouldn’t put too much faith in these figures as they primarily represent the performance in purely compute oriented workloads. This GPU will supposedly have a VRAM of 16GB. By all calculations, this would mean that Big Navi will sit between the RTX 3070 and the 3080, with the 20GB variant of the latter further cementing its position. Howevber, it should still be a capable 4K graphics card.
Then we have Navi 22 which will likely succeed the Radeon RX 5700 XT with the same CU count of 40. However, it looks like it’ll either have a lower L2 cache size or a reduced bus width of 192-bit or both. As per data from other outlets, this GPU may reach boost clocks as high as 2.5GHz. However, I seriously doubt that as you’d need water-based cooling to keep such as GPU from throttling. That is especially true since the 7nm node from TSMC is quite dense, and that this part has a max TDP of 170W. Navi 22 has a rated SP performance of 12.8 TFLOPs, just above the XSX. While many of my fellow journalists claim that this GPU should be enough for gaming, I wouldn’t recommend it for anything higher than QHD.
Lastly, there’s the Navi 23 GPU which will have 32 CUs or 2048 shaders. This will likely be the successor to the RX 5500 XT. There’s no info about the clock speeds and memory but I reckon it’ll feature a 128-bit bus and a boost clock close to 2GHz.
There’s also a Navi 31 GPU which looks identical to Navi 21. You may think that it’s a next-gen GPU but there’s a good chance that it’s either an OEM exclusive product or a professional PRO series graphics card.
| Property | Renoir | Cezanne | Van Gogh | Rembrandt |
|---|---|---|---|---|
| num_se | 1 | 1 | 1 | 1 |
| num_cu_per_sh | 8 | 8 | 8 | 6 |
| num_sh_per_se | 1 | 1 | 1 | 2 |
| num_rb_per_se | 2 | 2 | 2 | 4 |
| num_tccs | 4 | 4 | 4 | 4 |
| num_gprs | 256 | 256 | 1024 | 1024 |
| num_max_gs_thds | 32 | 32 | 32 | 32 |
| gs_table_depth | 32 | 32 | 32 | 32 |
| gsprim_buff_depth | 1792 | 1792 | 1792 | 1792 |
| parameter_cache_depth | 1024 | 1024 | 512 | 256 |
| double_offchip_lds_buffer | 1 | 1 | 1 | 1 |
| wave_size | 64 | 64 | 32 | 32 |
| max_waves_per_simd | 10 | 10 | 16 | 16 |
| max_scratch_slots_per_cu | 32 | 32 | 32 | 32 |
| lds_size | 64 | 64 | 64 | 64 |
| num_sc_per_sh | N/A | 1 | 1 | 1 |
| num_packer_per_sc | N/A | 2 | 2 | 4 |
| num_gl2a | N/A | N/A | 4 | 4 |
| unknown0 | N/A | N/A | 8 | 6 |
| unknown1 | N/A | N/A | 4 | 4 |
| unknown2 | N/A | N/A | 8 | 12 |
| num_cus (computed) | 8 | 8 | 8 | 12 |
The specs for the next 3-4 generations of APUs are also present. From a quick look, you can see that the CU count is going to stay unchanged with a max figure of 8. As we move from GCN to RDNA in VG, the wave config reduces from 64 to 32. but the max waves fetched per SIMD grows from 10 to 16, but the overall cache size seems to have been reduced going from Cezanne to VG and then again to Rembrandt.
Furthermore, the compute unit count in Rembrandt has been increased from eight to 16, thanks to an increase in the number of CUs per shader array. Overall, it looks like AMD will depend on the clock speed gains till Cezanne, then leverage the RDNA design to boost GPU performance with VG. Finally, Rembrandt will increase the execution units to significantly improve overall throughput.
