Discovered i had been asleep, the 6970 already released meanwhile, of course forget the 5870 chip when you can get a new generation chip.
But i guess it's simply a radical improved design.
Small quick question: Radeon HD 6970 has 2 GB ram, which is completely superior to 1 GB for the default 5870; will there very soon be versions with more than 2GB ram?
Let me get one tomorrow (if they already shipped).
Unconfirmed gossipping allowed!
I am not aware of any hardware variant of 6970 with more memory. Although
a new GPU 6990 is also scheduled to be launched soon.
That's most interesting.
The 6970 is 28 CM tall says reviewer, so it fits in the box here.
Will the 6990 fit on the mainboard?
The 5970 is 33 CM long, even the hugest mainboards, the 4 socket boards,
are not that tall. 30 CM might just fit, longer hits at the right side of the card the PSU connectors.
Also release date is interesting, is there a date for it?
In meantime i ordered a 6970, should be here in a week or maybe sooner.
It's eating 33% more power i read at review site meanwhile not really 33% faster than 5870, we shall see, 250 watt for a single chip
Just for curiosity: HD6990 will be a dual GPU card? And in that case, OpenCl support will be in single gpu mode again?
Alfonso could you explain what you mean with 'single gpu mode'?
You mean at a card like 5970 openCL can only adress 50% of the card and the other 50% is idle?
AFAIK there was two problems. first that second GPU core return garbage results. this was/(is?) because of internal crossfire.
second serialization of execution on two GPUs. this can be solved with setting GPU_USE_SYNC_OBJECT=1.
Look at the supported boards, you will see that HD5970 has OpenCL support only if you use only one GPU. If you search for HD5970 in the forums, you will find that me and other people managed to have multiGPU with HD5970, but this isn't officially supported (and it hasn't for years).
In short: if you create a context for each GPU, work in LinuxX86_64 and play with some enviromental variables, you can use both cores and 50% of the board memory (25% for each context/GPU).