AMD: R8xx Speculation
- Thread starter Shtal
- Start date
Wow, 400 double-clocked SPs via Fast14, hyper-transport based MCM design, return of ring-bus and FP16 TFUs and some EDRAM for free MSAA 16x. Missing these old-time crazy teories
nice!
640 is more than enough for lower mid-range, right? as a replacement of RV740 it would be a bit bigger and a good 15% faster running the same amount of shaders.
Well, with the i7 @ 4Ghz dragging up P scores like the top thrill dragster, I would count on most 4770s to get into the P6500-7500 range.
(Assumption here: the Performance preset wouldn't be particularly bandwidth limited, as you're getting a 28% increase in GPU score with a 25% clock increase with the 4850/70. I sense something similar being talked about way before.)
If JuniperXT wasn't using a 4Ghz i7 for the P9500, it would be moderately impressive (real GPU score would be 8300-8900), esp for 60-75W. A bad case 4Ghz i7 would mean a 7900~8000 GPU score, which is a nice step up in terms of SKU/price-perf but not cheery for the die size increase.
(Assumption here: the Performance preset wouldn't be particularly bandwidth limited, as you're getting a 28% increase in GPU score with a 25% clock increase with the 4850/70. I sense something similar being talked about way before.
)If JuniperXT wasn't using a 4Ghz i7 for the P9500, it would be moderately impressive (real GPU score would be 8300-8900), esp for 60-75W. A bad case 4Ghz i7 would mean a 7900~8000 GPU score, which is a nice step up in terms of SKU/price-perf but not cheery for the die size increase.
Hmm, that more than 40mm² difference to rv730 indeed looks like quite a lot "just" for dx11, if that's really 640SP part. Would seem more plausible with 800SP. Of course it would be possible there are other changes using die space which improve performance, but it had to improve performance quite a bit if AMD is willing to sacrifice shader clusters for it. The 3dmark score (if legit) though doesn't really suggest a 800SP part - might be limited by memory bandwidth (what's the memory clock?) however.
Oh and I thought those 60W TDP were for the high-end mobile part. For the desktop part that would be seriously impressive but I fail to see why it would need less power than the HD4770 (unless those 60W are chip not board power?). Hopefully though they figured out how to downclock GDDR5 for this generation...
Oh and I thought those 60W TDP were for the high-end mobile part. For the desktop part that would be seriously impressive but I fail to see why it would need less power than the HD4770 (unless those 60W are chip not board power?). Hopefully though they figured out how to downclock GDDR5 for this generation...
+29% is quite a lot for today, but many SMx to SMx+1 steps in past were even more demanding:
R420 -> R520 = +100% (160M -> 321M) [both 16 TMUs, 16 ROPs, 16 PS ALUs]
R580 -> R600 = +88% (384M -> 720M) [both 16 TMUs, 16 ROPs, however, both of them more capable, 56 ALUs -> 64 ALUs (48 PS vec3+1 + vec3+1 ALUs + 8 VS ALUs -> 64 unified 1+1+1+1+1 ALUs)]
R420 -> R520 = +100% (160M -> 321M) [both 16 TMUs, 16 ROPs, 16 PS ALUs]
R580 -> R600 = +88% (384M -> 720M) [both 16 TMUs, 16 ROPs, however, both of them more capable, 56 ALUs -> 64 ALUs (48 PS vec3+1 + vec3+1 ALUs + 8 VS ALUs -> 64 unified 1+1+1+1+1 ALUs)]
This is true but imho not really comparable. R580 vs R600 I don't even dare to try to figure out what changed, it's a very different chip.
R420 to R520 is more similar. R520 has 2 vertex shaders more, which should account for a small fraction of the difference. Pixel shader though is really quite different, that's probably quite a large difference in transistor count (it not only has more capabilities but needed to boost precision which alone should boost the basic MAD's transistor count by a factor of 2 or so).
I just don't see these differences requiring such fundamental hw changes with DX11.
RV870 cards already hit reviewers desks
At least AnandTech has them
If you wonder about my source it's Bingo13 post at XS ... (I hope everyone knows who he is
Link: http://www.xtremesystems.org/forums/showpost.php?p=3957884&postcount=29
Is that Derek? w0mbat, why the pseudo-moral-policing?
Link: http://www.xtremesystems.org/forums/showpost.php?p=3957884&postcount=29
Is that Derek? w0mbat, why the pseudo-moral-policing?
Bingo13 is Gary Keys, their chief motherboard guy or something, AFAIR.
R420->R520: ring bus, new MC, new dispatcher, > DX support, anyone? way more going on with that generational change than can be directly compared
flippin_waffles
Newcomer
Some pretty good information in this thread that seems believable enough.
http://www.semiaccurate.com/forums/showthread.php?t=536
http://www.semiaccurate.com/forums/showthread.php?t=536
mczak: Yes, that's what I'm talking about. They made changes, which wasn't necessary to achieve SM compatibility, but which they considered as important for other reasons. We could be in similar situation now. It's obvious, that unnecessary (from SM compatibility standpoint) chandes won't consume as much tranzistors as in previous architectures, but I think there will be some.
Also some local members doubted, that R8xx is just R7xx + DX11. Huddy mentioned more twice as big DX11 scheduler. If I'm not mistaken, scheduler of RV770 consumes only a few percent of the entire die, so it would enlarge the chip by (maybe) 5%. 25% are still a question. Number of functional units didn't changed - I'd expect some change in their capabilities (mainly TMUs) and maybe some change of memory controller(?).
Also some local members doubted, that R8xx is just R7xx + DX11. Huddy mentioned more twice as big DX11 scheduler. If I'm not mistaken, scheduler of RV770 consumes only a few percent of the entire die, so it would enlarge the chip by (maybe) 5%. 25% are still a question. Number of functional units didn't changed - I'd expect some change in their capabilities (mainly TMUs) and maybe some change of memory controller(?).
Which parts would you consider "information"?
As far as i've seen most of the stuff in that thread is just plain wrong
If true, then I'd wonder how press should do proper reviews for a new arch apart from running your average 3DMark/Quake3 benches.
flippin_waffles
Newcomer
All of it. I believe every single letter is written in stone, and is "100% accurate".
Impressive, at a 100%, it goes well beyond the minimum indicated by the site's name. That thread may well be too accurate and will have to be deleted.
Well, ull have up to 8 days. Ull be lucky if it wont play out like with the HD4850s
To be honest, the launchdates is just a guess based on the pr event. Maybe theyll give u more time, i dont know...
We'll see. In any case, they seem to have a large headstart on the other green guys - so IMHO no need to hurry that much.
Similar threads
