Normally there's more than one final number in data parallel applications! For folding, I guess the result is a set of positions and motion vectors for the atoms in the molecule.

Additionally, folding@home uses the two GPUs for two distinct work units. Work is not shared between the two GPUs.

You could design an application to spread itself over two or more GPUs. So, both cards will be computing equal shares of the application. The best performance comes when data does not need to be generated by one GPU and read by the other - or the sharing is minimal.

Jawed

Kyle Bennet seems to be absolutely sure about that 300w:
http://www.hardforum.com/showpost.ph...0&postcount=35
http://www.hardforum.com/showpost.ph...6&postcount=39

Err, I was actually asking about the MAD test :oops:

Not to bring up old news, but wasn't he "absolutely sure" that NV30 would stomp all over R300 too?

Just sayin, EETimes and InformationWeek reported 200W. Not saying that those publications can't be incorrect, but I'd take their reporting as gospel LONG before I take Kyle's reporting as gospel.

I still think it requires 300w of connectors...but not 300w. IOW >225W and < 300W.

There is no way it's going to use any ounce of electricity from those connections, hence, 300w is BS.

Still, greater than 225W is a lot of juice, don't get me wrong, and having 2x8pin and 2x6pin connectors on a psu (or some mangled combination of connectors/adapters adding up to 450w of connections from the PSU not counting mobo power) is a tough pill to swallow.

None-the-less, we've been expecting this for some time AFAIK.

They also suggest that the final R600 board will be 13" long, which doesnt jive either.

It doesn't...Yet those boards in the pic don't look like the OEM parts which are black.

*Hits forehead*

The Stream Processor version of R600 probably does look like that even for retail.

Who the hell knows...At least we've got a better idea of what's going on now. :razz:

I think it's pretty clear that they are assigning two independent kernels to two different cards and that the MAD test is the absolute optimal workload, so I doubt crossfire has anything to do with performance. ALU tests are generally one big uber shader that writes out junk data at the end (or throws it away if you can stop the compiler from dead-code elimination). There is no sharing, no texturing, and framebuffer bandwidth is irrelevent. No z-writes either. No AA. No AF. It's just an ALU test, period.

So IMHO, the combined 1018 GFlop figure probably implies 509GFlop peak on a single card, ergo, it's probably 512GFlop theoretical.

http://www.extremetech.com/article2/...2099613,00.asp
"Wednesday's meeting was also used to clear up what Richard characterized as a swath of misconceptions and rumors concerning the delayed R600 —or, what is to be AMD/ATI's first DirectX 10-capable graphics card.

"We pushed out the launch of the R600 and people thought is must be a silicon or software problem…it's got to be a bug," said Dave Orton, president and chief executiveof ATI. "In fact, our mainstream chips are in 65nm and are coming out extremely fast. Because of that configuration, we have an interesting opportunity to come to market with a broader range of products," he explained.

"Instead of having them separate, we thought, lets line that up, so we delayed for several weeks," Orton continued, referring to the R600 family as a whole, which AMD now says will come out at the same time (a matter of weeks as opposed to months, according to Richard) instead of just the high-end version."


Not a whisper about Barcelona...

They probably got left out of something and are going on an evil tirade because of it. From an engineering standpoint their comments seem to hold little weight. I've yet to see a picture that actually has 3 power connectors on it and even then pulling 100% of the potential power is unlikely. Besides with the 200W reports from the event I'm assuming someone from AMD quoted that as the power draw.

It's possible one of the true SPs might get up there with 2GB of ram and the fan that sucks in small children but I think it would be rather obvious what you're looking at when you see those specs. Not to mention probably hear it in the next room.

Thanks all for the clarification on that matter! :)

Being wrong doesn't take confidence.
edit: That was meant to imply that just because Kyle sometimes shoots streamers out of his ass doesn't mean hes a firework.

That's incorrect. That's like saying there's a MADD and ADD for every component of each shader. There is actually one MADD in four components (8), plus 1 scalar ADD. If you used your math, the Xenos would have 360 GFLOPs peak. In reality it has 216.

So 64*9*0.8=512 like DemoCoder said.

As far as the 300W figure is concerned, that's calculated by adding together the maximum amount of power one can draw through all of the sources available to the card: slot, 6-pin connector, 8-pin connector. The fact that the card has this configuration suggests that it will, occasionally, use more power than a pair of 6-pin connectors plus the slot could provide - in other words that the peak power usage will be over 225W. However, if the peak usage occasionally hits 225W, the typical power usage will be lower.

There's also a strong rumour that the card will function quite happily with a pair of 6-pin connectors in all respects other than that it will disable overclocking in the driver settings. If that's correct then it means the card will achieve default clock-speeds and never exceed 225W even at peak consumption, but that overclocking may push the peak power up to 225W.

All in all, a typical power output of 200W fits quite well with that, I think.

Ss far as the delay is concerned, ATI/|AMD's official line is that R600 was delayed so that it could be launched alongside the mid-range equivalents. I've read nothing from AMD or ATI that suggests a launch with Barcelona - it's simply about launching all R600 variants at the same time so as not to annoy their OEM partners. At least, that's the official story.

If this is true then ATI has insulted its enthusiast customers by holding back an enthusiast-level product when there was no technical need to. If they're lying - well, that's bad too. Either way, ATI has nothing to be proud of, here.

As far as the 1 Teraflops figure is concerned, this is clearly a marketing number. It is therefore possible (IMO) that it includes not just the GPUs but also the CPUs in the box they were demonstrating.

Minor nitpick: 64*9*0.8 = 461 and NOT 512.

It's either more FLOPs per ALU or the frequency is a lot higher than 800MHz.

I thought the same, but the pics clearly state that brooks only running two clients on the two R600.

888 MHz or 10 operations per cycle ... maybe a 5D-MADD? (if the 64 are right ...)

I was saying there is a MADD and ADD for every component. I already changed my assumption that it's using a MADD+ADD like they have in the past and went to 5 pure MADDs. Besides how do we know they aren't actually using a MADD+ADD configuration like the the R5x0 series. If they're all MADDs aren't we just looking at a really big xenos with almost no other changes besides the ADD->MADD? They quoted 95% efficiency with Xenos, no clue why they'd set out to actually try and improve that. R580 is already capable of a good deal of the DX10 requirements so they wouldn't need very many radical changes to make that concept work. I'm still not entirely convinced they aren't using a Vec4+1 MADD+ADD setup. If they ran code that didn't use the ADDs we'd see exactly what we are now.

64*5*MADD*.8=512GFLOPs

Also HardOCP said R600 REQUIRED 300W. That's like saying a computer with a 1KW supply will suck 1000W out of your wall socket.

BS :smile:

If it was vec4+1 MADD+ADD and they ran a MADD test you would get 64*4*MADD*.8 ~ 410 GFLOPs.

IIRC G80s measured MADD issue rate was near peak but it's only 345 GFLOPS worth. So I'm assuming that R600 is MADD capable on all 320 thing-a-magigs.

[MADD+ADD][MADD+ADD][MADD+ADD][MADD+ADD] + [MADD+ADD]

This setup where all of the ADDs essentially drop out because they aren't being used. So it would effectively look like 64*5*MADD on the benckmark, in reality it would be 64*5*[MADD+ADD]. Doubt that's what they're doing but I'm not going to rule it out yet.

Its not the retail cards, the motherboard not ATX style, has 2 cpu socket and very long, check the PSU is long too, looks like as a enermax galaxy http://www.enermax.com/english/produ...y1.asp?PrID=60

I don't know, it seems to me that it would be more insulting to customers do the usual 'soft' release where just a few thousand of cards go on the market initially with most consumers having to wait a month or two to get their hands on one. NVidia has been in a strong position for their past couple of launches so they've been able to delay launch long enough for a proper 'hard' launch with plenty of cards on the market.

If R600 is a compelling product which is 'better' than G80 and equal to the expected G80 refresh and if the R600 family mainstream/low-end parts outperform their G80 equivalents, the delay might work out very well indeed. A couple of 'ifs' in there of course!

I must say that I've always been somewhat bemused by the way the real money-makers (i.e. mainstream and low-end chips) are released so long after the high-end parts. I can see why the IHVs do it that way but surely it is better to release the whole family around the same time?

Yeah that certainly wouldn't be indicative of Xenos heritage. It would also put a single R600 at > 750Gfops. Pretty unlikely IMO.

Wouldn't it have been better to just beat up on G80?