AMD RV770 refresh -> RV790
- Thread starter w0mbat
- Start date
Coding for vec4 always as the default case is hard. There are rarely any apps in the HPC world that I know of where vec4 is a perfect match. Even in graphics world, it isn't the best fit as G80 shows. Getting max perf from AMD hardware is going to be mighty hard(impossible?) if you want to write portable code and AMD tools don't do some kind of vec4 packing.
It's to make it easier for you (the human) to read / write shaders. I think you'll find that even on vectored hardware, the compiler is doing far more magic than you imply is going on.Jawed said:
It is, relatively speaking, easy to auto-vectorize languages that don't have pointers, nor side-effects, and where programs are ridiculously short.
Seems like you're assuming R600 to be a vec4, when it's vliw (Jawed is also talking cell etc which is different ofcourse). At least for shaders the compiler does a pretty good job of packing scalar code, when possible for branching. A bit of unrolling will usually be a good idea, but it would be stupid to vectorize yourself..
I knew that i should not have deleted my already typed text. It said something about "quite serial in a transcendental, not technical way".That's parallel as occasionally fallen blocks not at the leading edge shift
But as I said, let's please move this discussion to another thread.
Jawed
Legend
When they announced the 4860, also released board shots which showed "Qimonda IDGV1G-05A1F1C-50X" chips. The shots later turned up on the amd site with the memory chips details photoshopped out. Sincerely hope someone didn't lose their job over that
From qimondas gddr5 page that's 5.0Gbp/s. They don't list the current 4870 gddr5(3.6Gbp/s), too slow obviously....in the pdf they also list 5.5Gbps.
The leak on chiphell last week suggested 4.0Gbps, the current lowest listed qimonda sku. I suppose the possibility also exists that the overclocked version at the end of april may get 5.0Gbps to differentiate it enough from the regular RV790...also i think samsung is shipping 3.6 and 4.0 and has announced 5.0Gbps, just in case qimonda falls to pieces.
Edit: for completeness hynix here have listed 3.6, 4.0 and 4.5Gbps in mass production
Last edited by a moderator:
Jawed
Legend
I posted this:
http://forum.beyond3d.com/showthread.php?t=53170
as a continuation of my thoughts on scalar as the starting point.
Jawed
http://forum.beyond3d.com/showthread.php?t=53170
as a continuation of my thoughts on scalar as the starting point.
Jawed
Yes, they are.
About every board out there with GDDR3 or GDDR4 has memory clocks slower than rated speeds. Maybe it has something to do with reliability? I honestly don't know, but it's a common practice.
This is often the case, and there are multiple factors why. Its less so with GDDR3 chips because the characteristics of them are fairly well known these days and they are running relatively low data rate in comparison to what the ASIC's memory interface can handle. In the case of GDDR5 remember that we are making a one time jump of the memory interface running in the 2.4-2.8GHz (GDDR4) to 3.6-4.0GHz range and it was never a guarantee that it would be achieved even close to the max capable data rates.
In fact, though, HD 4870 could have been set free at a slightly higher mem speed, but A12 silicon had nicer engine timing characteristics than A11 so we traded off a little memory speed for a higher engine speed to still stay in the same thermal/power envelope while giving more performance overall.
why cant we get ram running like that
Uhm... he just explained why?
I think he means RAM for the whole system, not gfx cards
Because system ram is better with lower latencies than graphics ram.