AMD RDNA4 in LLVM project The next-generation gaming graphics architecture known as GFX12 or RDNA4 is being prepared by AMD. It appears that AMD has now decided which GPUs from the RDNA4 series will be the first to deploy. The GFX1200 and GFX1201 are two unnamed graphics processors which have now been spotted in Linux […]
The only hint people had for that was the bus width, which was indeed 256-bit. It’s just that people couldn’t fathom that AMD had a fast GPU with so little memory bandwidth.
I’m not sure this leak shares anything that would tell us what performance target this GPU could belong to.
It wasn’t about bus width, it was about nVidia’s fictitious CUDA core counts with Ampere.
At the last minute, the 4352 CUDA cores of the 3080 (same as the 2080 Ti) was changed to 8704 “CUDA cores”, because the INT32 ALU was replaced with a dual-function INT32/FP32 ALU. People who didn’t understand that (i.e. basically everyone who didn’t call out nVidia’s dishonesty in marketing those figures) thought, from the leaks, that it’d be 8704 shaders against 4608 shaders. It wasn’t. It was more like ~5200-5400 shaders, depending on resolution, against 4608, with the latter running at a substantially higher clock speed.
Ironically, the reverse happened with RDNA 3, as the values leaked were incorrect - they said 12,288 ALU’s for Navi 31, without mentioning that it was really 6144 FP32 ALU’s with 6144 INT32/FP32 ALU’s that could be partially used. So people thought it was 12,288 on the side of Navi 31 versus 16,384 on the side of the 4090, with those numbers meaning the same as they did with 5120 for the 6900 XT versus 10,496 for the 3090. But they didn’t mean the same thing at all. It was ~7400 effective shaders for Navi 31 versus ~10,240 effective shaders for the 4090. With no real clock speed advantage.
As it turns out, the 4090 scales pretty poorly though, so it’s not as far ahead of the 7900 XTX as it should be base on raw compute.
Maybe you’re right, but I’m not sure the timelines agree with this. While people were trying to figure out whether or not the Navi 21 was going to be competitive, Ampere was already a known quantity regardless of shader count. The cards released a couple of weeks before rdna2 specs were announced.
That’s kind of my point. Ampere was out, with known specs. RDNA 2 specs were leaked before AMD announced the cards. People who got this leaked information compared number A to number B without understanding that number A was manipulated by dishonest marketing. So they drew the wrong conclusions about performance, saying AMD would be lucky to match the 3070.
Which made it pretty amusing when every one of the first three RDNA 2 cards that AMD released was faster than the 3070, from the 6800 to the 6900 XT.
On the other hand we had leaks for Rdna to be competitive at the high end and take the performance crown due to chiplets.