AMD: Sea Islands R1100 (8*** series) Speculation/ Rumour Thread

Is 1GB in clamshell on 64bit actually cheaper than 1GB without clamshell on 128bit? (given same GPU). If not, then you're throwing half your bandwith away for nothing.
 
The point of clamshell is to hook up more GDDR5 chips for a given GPU with a fixed I/O width...
 
The point of clamshell is to hook up more GDDR5 chips for a given GPU with a fixed I/O width...
Well yes but the gpu in question here has a 128bit bus anyway, so chopping off half the interface (along with half the bandwidth of course) just to then use gddr5 in clamshell mode seems a bit awkward.
 
Nice article from TPU. However, people (and also me) don't agree with the inflation in numbers. 7400 to 8600, 7500/ 7600 to 8700, 7700 to 8800. :oops:

And Radeon 8870M compared to 650M where it shows light years superior performance. Why is it compared to it?

AMD Radeon HD 8000M Series Detailed, Performance Figures Released

What the hell, with such high number the performance doesn't make sense. 8800 to beat a 650? since when 8 means mid/low range? 4800M used 4800 desktop's chip, now 8800M using something comparable to a 7700's. Anticipating 9800M to use HD 8600's specs.
 
Well yes but the gpu in question here has a 128bit bus anyway, so chopping off half the interface (along with half the bandwidth of course) just to then use gddr5 in clamshell mode seems a bit awkward.

We're talking about the 384 SP 77 mm² chip here right? How small can you go and still comfortably fit a 128-bit memory bus on the chip along with all the other necessary buses for PCIe and video out etc.?
 
We're talking about the 384 SP 77 mm² chip here right? How small can you go and still comfortably fit a 128-bit memory bus on the chip along with all the other necessary buses for PCIe and video out etc.?
I guess it's a close fit, but I was purely going by hardware.fr which are saying 8500m/8600m are 64bit and 8700m is 128bit - and all of these are using the same chip.
 
hw.fr says 64bit for 8500/8600 and 128bit for 8700. TPU's database said 64bit for 8500 and 128bit for 8600/8700, but it was changed to 128bit for all of them :???:
 
I guess it's a close fit, but I was purely going by hardware.fr which are saying 8500m/8600m are 64bit and 8700m is 128bit - and all of these are using the same chip.

hw.fr says 64bit for 8500/8600 and 128bit for 8700. TPU's database said 64bit for 8500 and 128bit for 8600/8700, but it was changed to 128bit for all of them :???:

Confusing :???:

I guess we'll know for sure after CES.
 
We're talking about the 384 SP 77 mm² chip here right? How small can you go and still comfortably fit a 128-bit memory bus on the chip along with all the other necessary buses for PCIe and video out etc.?

You have the option of dropping the video output, as nvidia did on gf117. This would make it pretty much a laptop part only.
(I couldn't find a figure for the gf117's die size)
 
It's 64bit for 8500 and 8600, 128-bit for 8700. 8800 is 128bit.
I still don't really get it though. 64bit ddr3 is useless, can just barely beat intel current IGPs on a lucky day with that (unless that's an atom igp that is...) and probably losing to trinity igp. 64bit gddr5 in clamshell mode to get 1GB is ok but there seems to be very little point in using that over 128bit ddr3 (which will easily get you 2GB with nearly the same bandwidth, probably lower cost and maybe even lower power consumption as well, though maybe using more area).
So I believe it when I see it...
 
Well 8700M has "turbo" and 8600M has no turbo, a little difference in support circuitry; then 8500M looks like a lower speed bin.
So maybe there's just small difference between models and 128bit ddr3 vs gddr5 makes a huge difference enough (the latter is more power hungry too so fits better with the "turbo" variant that can underclock based on power use)

64bit ddr3 could be just used to fleece customers (but, you get AMD drivers instead of Intel drivers so it's still slightly different than using the Intel GPU). That would be some radeon 8530M variant
 
Looks like the naming is so confusing even amd couldn't figure out the names, judging by the pulled review of techreport of a 8770m which really was a 8790m :).

edit: techreport confirms the 77mm² and 128bit (for 8700 series). And their test samples definitely seems to use 128bit gddr5 in clamshell mode (they look like H5GQ2H24AFR-T2C parts to me, so there must be 4 more at the backside, all probably running at 1.35V).
 
Last edited by a moderator:
Looks like the naming is so confusing even amd couldn't figure out the names, judging by the pulled review of techreport of a 8770m which really was a 8790m :).
AMD also sent TechReport an Intel based motherboard to benchmark this GPU! Whoever came up with that idea likes to live dangerously. ;)
 
Well, I can't really see the point for the 384sp chips for the AMD platform... their APUs are already there. That leaves competing with Nvidia on Intel platforms, but I can't imagine that is very attractive.
 
Well yes and that is called clamshell mode :). I've never heard that term being used for ddr3, but it appears in gddr5 papers for instance this one from elpida: http://www.elpida.com/pdfs/E1600E10.pdf
Are you saying we'll see such configurations in the mobile space? I know they aren't that uncommon for some desktop cards (using same pcb with either one or both sides populated).
I think that is a hangover from GDDR3 documentation; I've never use that to describe such a GDDR5 configuration, we would use "x16 mode". If you take a look at a GDDR3 board of such a configuration IIRC you'll find the devices are mirroring each other on the PCB, hence the "clamshell", but there is no requirement for that on GDDR5 layouts.

I still don't really get it though. 64bit ddr3 is useless, can just barely beat intel current IGPs on a lucky day with that (unless that's an atom igp that is...) and probably losing to trinity igp.
They are tuned so they do. And thats at 3DMark, when it comes to games things get much better. 64b also comes in a smaller package that brings layout benefits as well.

If Mars is 128-bit they probably dropped the display pipelines and may reduced PCIe to x8.
This ASIC has display pipes/PHY's.

Well, I can't really see the point for the 384sp chips for the AMD platform... their APUs are already there. That leaves competing with Nvidia on Intel platforms, but I can't imagine that is very attractive.
The closest you are to 1-to1 performance between the APU and GPU the better things are for Dual-Graphics. However, the market for discrete notebook graphics is dominated Intel so the primary performances/configurations are chosen to address that market.

AMD also sent TechReport an Intel based motherboard to benchmark this GPU! Whoever came up with that idea likes to live dangerously. ;)
Actually, I'd suggest its more of an external pointer as to how the company is being structured and is able to run under the current stewardship. Things like "Never Settle" are more oblique pointers as well.
 
Actually, I'd suggest its more of an external pointer as to how the company is being structured and is able to run under the current stewardship. Things like "Never Settle" are more oblique pointers as well.
I don't doubt that this can seen as a positive from inside former ATI: not beholden to any CPU platform etc. But from outside, it really sends a terrible message: "they don't even like their own CPUs." It can't make the CPU group very happy either. And it's completely unnecessary: it's not as if AMD doesn't have fast enough CPUs to max out the performance of these GPUs?
 
Back
Top