AMD: R7xx Speculation

Status
Not open for further replies.
As was pointed out in another forum, this cooler actually seems to be made of aluminium instead of copper.


But that's giving me some trouble to figure out, which would essentially be better to get a lot of thermal energy off of a small contact area.

AFAIK cu heats up better/easier and so is better for a fast heat transfer whereas al gets rid of the heat better so that you wouldn't need as fast a fan speed. Complicating this matter are the heatpipes. There has to be some thermal transfer from the basic material to the heatpipes, too - but what would be better al or cu?
 
Are you talking per core? cause a 16 wide FP32 SIMD unit shouldn't take more than million tranies no matter how you are going to do it. RF will take roughly 4k transistors per vector register file.
I'm referring to a unit that encompasses ALUs (5-way VLIW x 16 lane), the sequencer unit that controls it (though I have my doubts about the way sequencing is architected), 256KB of effective register file (don't know how porting affects this), buffering used to re-order operands and connectivity with the rest of the GPU.

I'm thinking in terms of RV670 being 4 SIMDs and RV770 being 5. RV670 has 1MB, effective, of register file.

In R5xx's pixel shader hardware we know that each ALU lane + "corresponding" register file + connectivity amounts to almost 2M transistors - R580 and R520 differ almost solely in ALU lane count (48 v 16) and differ in transistor count by 63M (384M v 321M). I imagine some difference arises solely because of the "3:1 sharing" from ALUs to TMUs seen in R580 whereas there's no sharing in R520.

Jawed
 
AFAIK cu heats up better/easier and so is better for a fast heat transfer whereas al gets rid of the heat better so that you wouldn't need as fast a fan speed. Complicating this matter are the heatpipes. There has to be some thermal transfer from the basic material to the heatpipes, too - but what would be better al or cu?

I asked a mechanical engineer friend about this the other day. Copper is superior in every way to aluminum for convective heat transfer (i.e. heatsinks with airflow). Aluminum is better for radiative heat transfer, like in a vacuum. Radiative heat loss is minuscule in comparison to convection, so this really moot for us. Copper is better for our PC stuff in every way. It's just that aluminum usually is adequate for the job, and it is quite a bit cheaper and lighter.

What annoys me with some coolers is the anodized aluminum that they color to look like copper. You can see it in my heatsink images above. The fins and most of the base are Al, while the die contact area and heat pipes are copper. I couldn't stand this cooler in the end anyway. It got very loud whenever the GPU actually had to render anything. So I got a Arctic Cooling Accelero S1 instead. Fanless is best.

Which reminds me of the 8800GTX cooler. It is barely acceptable IMO. It's similar to this Gecube 3850, in that it has Al fins but a copper die contact area. The heat pipes are even Al on that one. The GPU runs right to 85C in most games and the fan is at 100%. Very loud not very cool. Quite a cheap, disappointing cooler if you ask me. Standard OEM-quality fare, I guess.
 
Last edited by a moderator:
Well I think we can see that it's no G80/R600-esque monster GPU. Considering the process is still 55nm (it is right?), I wouldn't think that there is going to be any epic performance leap happening here.
 
Well I think we can see that it's no G80/R600-esque monster GPU. Considering the process is still 55nm (it is right?), I wouldn't think that there is going to be any epic performance leap happening here.

What piqued my interest in terms of performance was the rumored GDDR5. Depending on circumstances, this could have given a significant boost and would have been interesting in its own right from a pure technology standpoint. (That is not to say that this won't be a worthwhile update regardless, even using the same process node.)
 
Well I think we can see that it's no G80/R600-esque monster GPU. Considering the process is still 55nm (it is right?), I wouldn't think that there is going to be any epic performance leap happening here.

You'd be shocked to see how little effort on the right aspects of RV670 goes a long way ;)
 
You'd be shocked to see how little effort on the right aspects of RV670 goes a long way ;)

oodles of fillrate and more bandwidth would work some magic. I get the vibe though that they are adding shaders, so we could just end up with yet another increase of the shader:texturing ratio and it might perform the similar to RV670 per-clock in games! :neutral:
 
I'm wondering if they ever thought of using fast-trilinear TMUs instead of plain bilinear. Granted they should be a tad more expensive than bi-TMUs, but at least they'd get single cycle trilinear which would speed up AF performance quite a bit IMHLO.
I'm afraid that this "tad" is essentially the same as doing 32 bilinear TMUs...
 
I wonder if this R700 is being made to compliment deferred shading, deferred rendered games more so then traditionally made games :?:
Tis it's true that most of the popular games are unreal engine based but few of them can properly use AA to date. No, I don't count going into the control panel and making changes.
 
oodles of fillrate and more bandwidth would work some magic. I get the vibe though that they are adding shaders, so we could just end up with yet another increase of the shader:texturing ratio and it might perform the similar to RV670 per-clock in games! :neutral:

I would think they would have learned by now that increasing the shader count is NOT the way to go, at least at this time or for the foreseeable future.

You're right, it's all about texture fill rate. Increasing the memory bus bit to 512 and adding at least 640MB+ of on-board memory per card will also help in their x2 configurations. The crossfire bandwidth limitations is a part of what is killing them in these situations, and x2 is what they seem to be focusing on right now to try and keep up... I use the word "try" very loosely by the way.
 
oodles of fillrate and more bandwidth would work some magic. I get the vibe though that they are adding shaders, so we could just end up with yet another increase of the shader:texturing ratio and it might perform the similar to RV670 per-clock in games! :neutral:

I'm gonna go with Jawed on this one- Z fill and AA samples are the major element.

Stock raw speed, the RV670 is quite formidable still, and although raising units (they've learned their lesson by now I reckon) in the right balance will help stock performance and to an extent make the RV770X2 a much better experience compared to RV670X2, Z and AA are bigger problems that need to be solved.


You could always design up and make an even more performance oriented variant of RV770, though it is unlikely, looking at AMD's finances.
 
If I were on my Mac now I'd be able to to some relative analysis in Photoshop, relative to the fan and then the PCIe (since we're pretty much too lazy to get a card out of a system for now :p)

Any accurate/semi-accurate measures on how long the PCIe X16 connectors are (in whole)?
 
2008041299a352fa5cdc00cvy1.jpg


20080412c081e23a80c02f4wo3.jpg


20080108a828732bfda38deri4.jpg


For me it looks like this grey thing on the possible RV770 pcb is too big and cannot be equal to the RV770's die size.
 
20080108a828732bfda38deri4.jpg


For me it looks like this grey thing on the possible RV770 pcb is too big and cannot be equal to the RV770's die size.

I thought that others here had labeled this as a mockup of a RV670 based product?
Though there are rumors that RV770 will be using the same PCB as the RV670, so who knows, only a few more months to wait.
 
The mounting holes in the cooler don't match the PCB. Additionally the "tape" or whatever the hell you call the strip that's supposed to contact the power-supply doobry-wotsits is in completely the wrong place to match up with that board.

As for the actual die size indicated by the face of the heatsink, well I can't be bothered frankly. It's clearly in the same ballpark as RV670, it's not a monster 400mm2 or something. So the rumour of a 250mm2 (I can't remember, to be honest) die seems correct.

Jawed
 
I'm afraid that this "tad" is essentially the same as doing 32 bilinear TMUs...

That would mean that a fast-tri TMU would cost as much as a pure trilinear TMU, whereby the first wouldn't make much sense. I'm not an engineer to know what a fast tri TMU would cost, but afaik PowerVR SGX uses fast-tri TMUs and it would be an at least weird design choice to use such costly TMUs when die size and transistor budgets are so critical in the markets they're addressing.
 
Status
Not open for further replies.
Back
Top