HD6970/6990 question

Discussion created by twisted_pi on Apr 9, 2011
Latest reply on Sep 14, 2011 by corry


At first sorry for my english.

I have some problem now with my clients who uses my software and GPU. My code works on 5870/5970 cards in CAL(code created in OpenCL, compiled to IL and a bit modified) with some modifications in ISA ASM(pressured use some patching coz of lack of all ASM commands in compiler).

Because is almost not possible to buy 5970, peoples buy 6970/6990 cards. To be still on market we must add new kernel for users, but here are problem. To be at TOP on market, out code must be fastest, it can be achieved by patching kernel for some ASM commands which are not supported by AMD compilet for unknown reason. Now kernels for 6970/6990 do not works correctly, so I assume that assmebler opcodes are changed(maybe all or maybe some), is this true? If yes, then is possible to get new Instruction Set info? If not, then how to fix that big problem if I cannot use 5970/6990 in OpenCL coz bug with 1 GPU working?

And last question, why latest update still do not support full potential of AMD cards? I was really 99% sure that it will be available in APP SDK 2.3, but not. Do AMD programming crew is anyhow informed about really important problems?? 1 GPU bug already exist for 1 year and 4 months, so same problem is with compiler which do not use full GPU potential.

We and our clients who mostly buy 2x5970 and they are thousands are really presured to change brand to NVIDIA, which show that their latest products are really close in integer numbers calculations(in real numbers they are faster always). We really love AMD , we can understand many things, we can understand that no one is perfect and make bugs, we can wait... but how long?

In final words, our 6970/6990 kernels must works slower than 5970(1 core), even if that cards have higher clocks, but 5-way VLIW changed to 4-way make no difference in speed - because lack of info, again not used full potential by compiler. Maybe they are faster in games(better tesselation etc), but not in OpenCL what is future for GPGPU.Now when I saw info that CAL support will be discontinued(is already in 11.4) in next releases, we have no choice to inform our clients to prepare to swap GPUs from more serious company, who listen community.

Sorry for sad words, but this is like we feel. We will not lose our customers, they will understand that is not our fault, AMD will lose.

I do not blame you supporters, coz you own job and I know that will get response similar to "we are working on it", just please inform correct persons who are responsible for it.