This morning once more we were talking power in TFLOPS FP32 (single precision) of the next NVIDIA RTX 4000 and AMD RX 7000 graphics cards. We are talking regarding more than 100 TFLOPS for the RTX 4090 and 92 TFLOPS for the Radeon RX 7900 XT. On the other hand, according to new information from a reliable leaker, this figure should be revised downwards. Indeed, we would no longer be on 92 TFLOPS, but 73 TFLOPS.
Technical specifications revised downwards for the next AMD Radeon RX 7000 cards
The reason for this is that the card would feature fewer stream processors than we thought. Indeed, there were an estimated 15,360 stream processors on the Navi 31 GPU, but there would actually be 12,288, or 20% less. As a result, the power in TFLOPS drops to 73, provided that we are still on a GPU frequency of 3 GHz. Note that this would impact the entire range of AMD GPUs and not just the RX 7900 XT. Similarly, the number of “working groups” would in fact increase from 60 to 48 and therefore from 120 to 96 Compute Units.
Thus, the mid-range chip represented by Navi 31 would not have 10,240 stream processors, but 8192. The number of Workgroups goes from 40 to 32 for 64 CUs instead of 80. As for Navi 33 for the entry of range, we would therefore have 4096 stream processors and not 5120. Workgroups go from 20 to 16 for 32 CUs and no longer 40.
Navi 31 old → new spec
60 WGP, 15’360 FP32 → 48 WGP, 12’288 FP32
on 3 GHz = 73 TFlopsNavi 32 old → new spec
40 WGP, 10’240 FP32 → 32 WGP, 8’192 FP32Navi 33 remains the same (16 WGP, 4096 FP32).
Thanks @greymon55 https://t.co/fdsrJT2aew
— 3DCenter.org (@3DCenter_org) May 2, 2022
Let’s finish by recalling that we will have to wait several more months before having official information that will fix the specifications once and for all. As always, even if this information comes from reliable leakers, some information may still change or be only partially correct, or even downright false.
