Maybe you're right, who knows.. but they must use that very wide ring bus for something, right? :)

Other than 512bit huge pee-pee bragging rights?

Guess we'll see week.

just because people are interested does not mean they consider these revelations the be-all-end-all of gpu testaments. no matter how you slice it, it IS relevant. if you don't think so, then feel free to ignore them and keep to yourself. in an info-obsessed community, calling other people ignorant for considering these synthetic results as any indication of some new tech's capabilities when there is a nda in effect and very little official info to go on is pretty judgmental.

i mean the title of this thread is 'Rumours & Speculation', so cut people some slack. if you want concrete thorough analysis based on a gauntlet of the latest pc game fraps, then what are you doing in this thread anyway?

besides, we already get it: real world > synthetic. obviously. but pissing on 3dmark as 'meaningless' is pretty 'ignorant' in itself, if that is the term you wish to use. i'm no hardware professor, but when i looked at the latest vga charts on tomshardware last, it took only a cursory glance to discover that the 'overall games fps' average of 7 unique 3d intensive games scaled almost identically with the '3dmark 06 [v 1.0.2.]' results. reckon this is only coincidental?

Personally, it's mostly in top-end parts that I think default 3DM scores are worse than useless. I can see where they'd have some thumbnail usefullness in midrange and lowend parts. I mean, clearly they *do*, as the OEMs and IHVs use them that way. But default for high-end? No thanks.

I've looked at the CTM documentation, sections 3.2 and 3.3 indicate 3 instruction co-issue (ignoring flow control and texturing):

• RGB
• A
• RGBA "presubtract"

Three operands are allowed. The presubtract operates on operand 0 or operands 0 and 1:

• SRCP_OP_BIAS - 1-2*src0
• SRCP_OP_SUB - src1-src0
• SRCP_OP_ADD - src1+src0
• SRCP_OP_INV - 1-src0

and the result can be used as one of the three operands for either or both of the "main" instructions.

No, I was claiming that a vector instruction devolves into a trivial co-issue in R600.

Jawed

Out of curiosity, did anyone manage to get all of the images from that thread at OCworkbench?

http://my.ocworkbench.com/bbs/showth...421#post412421

It's a rumors and speculation thread...going ballistic over 3dmark06 scores supposedly pertaining to R600 is what we have to do here! To do otherwise would be blasphemy! BLAS-PHE-MYYYYYY! :twisted:

My reading of it seems to imply that presubtract cannot freely choose the two operands, so it's not like you can arbitrarily co-issue a X+Y or X-Y for any two X and Y. Instead, it just seems to be a DX8 source modifier unit with 2 extra ops.

Should I call _bias, _sat, _x2, _d2, etc in DX8 "co-issue"?

If you read horizontally, everything comes out just fine. These diagrams aren't consistent in the way they orientate the ALUs (if you go back to R5xx versions of these diagrams you'll see what I mean):

http://www.digit-life.com/articles2/...520-part1.html

e.g. X1600 should show the ALUs organised vertically if it follows the X1800 convention - which we know is 1:1 ALU:TEX. Or maybe X1800 is the odd one out, lol - except I've got an X1900 diagram here that agrees with the X1800 orientation.

:lol:

Jawed

I wonder if there'll be "cache coherency" traffic too.

One of the things that's got me puzzled is the separate cache that's shown at the left, labelled "memory read/write cache". If you interpret that as streamout for written data, then that could imply it's the source of vertex data when its read back in again. But that cache isn't connected to the samplers...

Also, you'd expect a cache to be too small to hold a streamed-out buffer - unless R600 is consuming the vertex streams at the same time as they're being written (dunno if D3D10 allows that). So if the cache is too small, when's it used as a read cache? Is it really distinct from the L2 cache for texture/vertex data?

It could be functioning as post-transform cache, I suppose.

Jawed

Well, I ran down all the performance kings (as in, the fastest card when launched) from the last few years, in cronologically discending order

8800 GTX - $650
7950 G2 - $599
X1950 XTX -$450
X1900 XTX - $650
7800 GTX 512- $650 (Heh.... I don't think any of the 43 people who got it paid less than $700)
X1800 XT - $550
7800 GTX - $600
X850 XT PE - $550
6800 Ultra - $500
9800 XT - $500
9700 Pro - $400

So, to answer Geo's question.. ;) ... August 2002.

It'll be interesting that after the middle of the month presumably you can add 8800 GTX Ultra 849(?) and HD 2900 XT 399 to that list. :P

From just a casual glance one would almost think that the companies are going in opposite directions with one going for higher prices and one going for lower prices. :D

It has absolutely no bearing on actual price trends as I'd imagine an XTX type card will go for 499 or 599. But it's still amusing to see either way.

Regards,
SB

http://http://img523.imageshack.us/i...pvfr600td6.jpg

what's the difference between GPU TDP and Board TDP ?


so the R600's(GDDR3) power consumption is 160 or 225 ??


and 80's Board TDP is 225W too , no ?

No - it definitely makes compilation more interesting within the constraint of the 3 source operands available :lol: The documentation also suggests there's a pipeline hazard there.

The independence of the alpha channel in the main ALU allows things like this in one instruction:

MAD r0,rgb, r1, r2, r3
ADD r0.a, r1, r2
MUL r0.a, r0.a, r0.a

That ADD being quite useful, according to Mike Houston when talking about folding@home.

As far as the ALU pipeline is concerned that's an implementation detail. We're talking about low-level co-issue here, aren't we, at the pipeline level? We're talking about a compiler using the units fully? How did NVidia implement the mini-ALUs for these modifiers? Was there an additional payload in NV4x or G7x?...

I wonder if those DX8 modifiers consume additional instruction slots in a D3D10 GPU. Sort of similar to the way that the fog unit has disappeared (thus requiring shader code)?

I don't see why that ADD raises your hackles. It's there for the compiler to use.

Makes R600 co-issue seem less fraught, actually :lol: Isn't that where we came in?

Jawed

Enter crackpot theory mode: (which can otherwise be completely disregarded because the 9700 pro was an all time high price point when it came out)

ATI was rather down and out (in terms of sales and debatably performance/features) when their performance champ came out, and had no choice but to price it low in order to get sales! Similar to the current situation with the X2900xtx!

^^;

That 16K+ 3dm06 is impressive , but that cpu getting over 2fps on both scores also seems super-duper clocked ? No ?

you forgot the X800XT PE @ $499 (i bought one at launch for $525 which seemed to be the new price. The XT just sorta popped up and took over the 475-499 price area.

It looks likely that some DX9 games will perform markedly worse on R600, seemingly due to a lack of texturing rate. (Bear in mind I think 7900GTX performs markedly worse than X1950XTX, if you want a reference point.)

I suppose there'll be some exceptions. From all the benchmarking mess we've seen so far, it seems that the ROPs are pretty much equivalent in capability to G80's (e.g. at 4xMSAA with HDR) so it's the texturing that'll show the biggest difference.

I think D3D10 is a different ballgame. I think there's a decent likelihood that this will favour R600, prolly quite heavily - because I think ATI spent a lot of transistors on stuff that's new for D3D10 whereas NVidia seems to have spent a lot of transistors on texturing and ROPs.

If R600 is a D3D10 beast, how many games is that going to be relevant in? How long before you replace the card you buy this summer?

The best thing is to wait until the games you want to play are benchtested on the cards you're considering - hey there's less than a week to go now.

The expected bandwidth for R600 once looked like a probable indicator of performance, but now if it's any indication of performance it seems like it could only be for D3D10 games or better than 4xMSAA.

Maybe Call of Juarez's D3D10 patch (coming in the next week or so?) will make the crowds go "Woo!" but I'm pretty dubious we'll see any compelling D3D10 game content this year.

I personally found a big disconnect between the pre-launch G80 benchmarks and those that appeared in launch reviews, and those that have appeared since. I'm finding benchmarks less and less informative... Hell, I'm in a minority round here and think 8600GTS is a decent performer in DX9 games for its specification (not its price) so what do I know?

Jawed

Thats because the dollar is dropping against the Euro. Do not panic.

If the r600 is heavily geared for dx10, then perhaps ati tried to duplicate the success of the r300, which had exceptional dx9 performance. Even the cut down versions sold well, despite that the competing nvidia cards to things like the 9600pro were faster most of the time, they just couldn't compete in dx9 performance.
Of course, with the r300 ati also had a beastly top end card that won in everything, sometimes by a large margin, and much better AA performance. It's possible ATI could have much better AA performance (due to their larger memory bus) this time, but the desperately need a high end yet mass market priced card if they want that halo effect again. (perhaps the xt will be priced low enough and perform well enough to give them that)

Does DX10 somehow reduce the need for texturing?

Although I guess, if the bottleneck shifts to other parts of the card in DX10, parts that ATI would do better in based on this presumption, it's somewhat irrelevant...

9700 pro was probably just cheaper because of inflation. That was a while ago.

The rest of the cases typically Nvidia launched first, months before ATI, so they could charge a premium early. By the time ATI came in prices had already dropped.

I think that you also need to look at the margins the cards and boards were bringing in to make a proper analysis. Nvidia is trying to ever increase it's margins, and at the same time the retail prices are going up. It's quite easy to get better margins by increasing your retail price (assuming of course that you are still getting the sales). This is an over simplification but it still needs to be considered. How much money is everyone making from the sales now compared to how much they were making 4 years ago as percentage of the board cost.

You mean like the fact that the $399 9700 Pro was driven by the petite 218mm^2 R300?

I think those assertions might be in need for re-consideration in light of actual DX10 competition coming in. Say, a few weeks from now...