AMD: R7xx Speculation

Just thought I should point out something regarding non-power of 2 scaling of TMU/ROP/ALU clusters:
People seem to be forgetting the ring bus.

In R600/RV670 there are 4 main ring bus stops, each serving an ALU cluster, a ROP quad, a TMU quad and with either a 64 or 128bit memory controller connected to either 2 or 4 RAM chips.

Assuming that there haven't been major changes to this basic architecture:
-To have 5 ALU clusters there needs to be a 5th ringstop, which would mean a combined 320bit bus, 20 TMUs, 20 ROPs & 10 RAM chips.
-To have 24 TMUs there needs to be 2 more ringstops, which means 384bit bus, 24 ROPs & 12 RAM chips.
Both would also produce increased system latency on the ringbus.

The pics of 4850 card show only 8 RAM chips (& there has been no rumor of other than 256bit/512bit) so there must be either only 4 ringstops or its been power of 2ed to 8 ringstops.

 

Please, that's not my style. To engage in revisionism I'd somehow have to change all the accurate R580 analysis on the web, including my own, to inflate everyone's numbers and everyone's graphs and everyone's conclusions. The hardware doesn't need me to do that, it's fast enough on its own. Cherry picking one data point from one piece of analysis is never going to get you too far, whether you're looking to talk a piece of silicon up or down.

 

Right, so if it is still 256-bit with 4 main ring bus stops w/ 8 RAM chips (the only info that's really confirmed is 256-bit & 8 RAM spots so 4 main ring bus stops sounds good), then Jawed's image of what RV770 might be like if it has 800 SP's actually works because there are 40 ALU's (200SP's) to 8 TMUs to 4 ROPS per ring bus stop

So unless architecture has changed significantly from R600, we will be seeing either 480SP's arranged in 4 arrays of 120SP's (currently 4 clusters of 4 per array for 64 ALU's in RV670 to 6 clusters of 4 per array to reach 96 total ALU's) or if 800 SP's we have 4 arrays of 10 clusters of 4 ALU's for 160ALU's (800SP).

Unless of course they went to 8 ringbus stops in which case everything can be doubled including ROPs and in which case you'd see 8 ALU clusters of 12 ALU's or 8 ALU clusters of 20 ALUs...

 

BTW have they ever given actual R600/RV670 die shots?

I've wondered if they've been reluctant to present R600 die shots because they knew the architecture would be the basis for the next couple gens worth of projects (R700 at the least) in order to keep dimensions and sizes secret from the competition

 

I don't think ring stop count indicates anything more than the memory bus width. Look at RV570/560.

Also, while 1 ring stop per set of memory channels (1 to 4 32-bit channels) + 1 ring stop for PCI Express is generally accepted, there have been hints in the past that there are other ring stops.

Jawed

 

Agreed. Also, we know nothing about which ringstops are actually connected to what. The simplistic view that each tmu and each shader cluster is actually connected to a different ring stop for instance might be wrong (it certainly is for rv630).
There is probably a limit how many clients a ringstop can have, but it could be quite a few - from http://techreport.com/articles.x/12458/4: "The ring bus has about 84 read clients and 70 write clients inside the chip, and PCI Express is just one of the many ring stops, as are the eight 64-bit channels to local memory." If rv770 had too many clients, either the ringstops could be redesigned to allow for more, or probably simply even more ringstops (without memory chips attached to them) be added - though the latter might not be an efficient solution.

 

Hmm, yes RV570 is problematic.
I should have remembered that having had an X1900GT (unit disabled R580 version) myself

It would not be problematic if there was only 192bit memory bus...
Or if they disabled one quad but left the memory controller active which would give the right specs.

Far as I can see RV560 is simply half a R580 but with full 8 VS units?

 

Why fake?

The GTX280 number is in line with the leaks from CJ. They use a new driver not available for download (last time I checked).

The strange numbers are the scores of the 8800GTX and the 8800Ultra. The Ultra improved the score around 20% from the old driver to the new one ( roughly the same improvement as the GTX280 X4100 -> X4800 ) but the score of the 8800GTX did go down a little bit ( X2200 -> X2100 ). So the results could be true and the difference between the 8800GTX and the Ultra could have been a mistake from Nvidia. They crippled the GTX but forgot to do the same with the Ultra. Same procedure as last year (AFAIR NV40 -> G70).

 

I think most ppl are screaming fake out loud to soon. Was there ever a pic which wasnt fake? Someone always knows 100% for shure that its a fake, cause he can see it at the color, front, blablabla.

AFAIK GTX280 will be a fast card, but having troubles beating rv770 cf.

 

A fairly even distribution seems necessary to make the ring function at all well. Unless it has gobs of excess bandwidth...

It's interesting that RV770 would appear to have a 512-bit ring internally (same as RV670) that's capable of supporting ~128GB/s.

http://forum.beyond3d.com/showthread.php?p=1165766#post1165766

I suggested "To support ~128GB/s of memory a 512-bit ring bus is going to need to run at ~833MHz. Does that indicate HD4870's approximate clock?". Perhaps there's more headroom on the ring - or maybe it's wider then 512-bits internally.

Hmm, each quad of TUs has 20 samplers, for example. I'm not sure how to read those client counts. There should be quite a few different kinds of clients, e.g. command processor, instructions, constants, read-write cache, TUs, RBEs.

I expect there are ring stops with no memory attached. e.g. there's presumably a CrossFire dedicated stop.

Jawed

 

My first post here...

Hello!

I've been long time follower of B3D - a lot of respect to members knowledge



Back on topic I've got small hint regarding HD48xx CF performance form XS forum member...

by ORB here..


I know from XS he is not an AMD/ATi fan and his leaks are legit.

Any thoughts

 

Stop wondering. If the competition wants a die shot, it just has to buy a card in the local Fry's, decap the part in some lab and take a picture. A 007 job it is not.

 

If he's correct NV should be worried...

 

Oh, OBR. I know this guy, I used to work for the same website as he does. His opinions are sometimes a bit extremist and he has been many times accused of both fanATIsm and nVidiotism. But I believe he does have the cards and I believe him if he says CrossFire is better. Perhaps it's only better in some games and settings and by a small margin, but since these are just 4850's... man, I think nVidia's in trouble

 

You have to factor in the cost and TDP of each solution as well. They're very similar (assuming 4870s). You pretty much can buy 2x 4870s for the cost of one GTX 280 (when they're available, of course). As for which one will be the superior solution, that remains to be seen (and I'm sure will vary wildly/widely across apps and settings, as always).

edit: but if it's 2x 4850 we're talking about here, and it is indeed faster than GTX 280, NV is going to lose a lot of enthusiast marketshare this round.

 

Welcome to B3D.

My thoughts: If true, big win for AMD considering thats 4850 CF.. If ever Nvidia needed a shrink, now is the time.

$398 on AMD offers better performance than $649 on Nvidia.

Or

How about 4870X2 CF?

 

Ah yes, 3x GTX280 v 2x HD4870X2 should be amusing.

Jawed

 

Finally a use for those KW+ PSUs

 

Indeed, if you exclude for example that based on expected prices 3xGTX280 will cost almost twice the price of 2x HD4870X2