AMD: R7xx Speculation

[AlphaWolf]

So far 4850 does not seem such a great overclocker though. Some sites got 700 no problem, which is the max the current OC software will allow. But another site only got 675/682 on two different brands of 4850, and concluded most of the better chips must be binned for 4870. Also the chip seems to run really hot, could be a negative on OC.

I am wondering how this 9800GTX+ will OC? If it's truly a 55nm part, maybe it will have more headroom. But it wont even be available for three-four weeks, a pretty big negative itself.

It's really close on 4850/9800GTX, depending on where final prices settle out with rebates, OCing capabilities, future driver improvements, etc.

it probably runs hot because the aibs suck at installing the hsf.

Some notes before the testing. The fan speed very low on the factory bios, I haven't been able to use any fan control to raise the speed so far. The temp is most likely fine even with an Idle of 79c but Being an XS'er I removed the factory thermal goop and replaced it with A.S. Ceramique. the idle temps are now 55c most of the time.

from here

That shouldn't be too big an impediment for anyone who wants to get the most out of the card.

 

[Silent_Buddha]

http://www.hexus.net/content/item.php?item=13935

If we assume that's actually a working proto-type or reference R700...

Look closely at the power connectors there's a 1x6 and 1x8. However, there are only 2x6 pin power connectors connected to the board. Meaning it's possible that R700 will consume less than 225 watts.

Regards,
SB

 

[Wirmish]

1. The 9800GTX retails for $229 so that would be $30 more.
2. It is also done at 55 so it is not just an OC 9800GTX.
3. It is cheaper to produce.
4. I think it is actually a little smaller than the 4850/70s

1. 9800GTX+ is expected to cost $229... in july or later.
2. So it's a smaller 9800GTX... hum... a smaller OC 8800 GTS... no ?
3. How do you know that ? 65nm yields vs 55nm ?
4. ---> link

EDIT: Good scaling no ? -> CoD4

 

[willardjuice]

Dont knock it till you've tried it. If the 4870's dont have dual duallink, theyve just lost a customer in me.

Well according to TweakTown even the 4850 has two dual-link ports:

Finishing up the tour of the HD 4850, we find ourselves staring at two Dual Link DVI connectors and a single TV-Out port

 

[randomhack]

any thoughts about a possible 4850 gddr5? or at least a higher memory clock?
Its a bit strange that 4850 has such a low memory bandwidth.

 

[AlphaWolf]

any thoughts about a possible 4850 gddr5? or at least a higher memory clock?
Its a bit strange that 4850 has such a low memory bandwidth.

I don't think you'll see AiB's using GDDR5 on lower margin parts for now (assuming it's allowed). If you want the bandwidth you'll probably have to buy a 4870 or overclock the memory.

 

[Blazkowicz]

well it's called the 4870!
anyway AMD seems to do just fine with the "low" bandwith and quantity of ram of a 4850, matching a 8800 ultra is nice.

 

[armchair_architect]

So, RV770's 1200GFLOPs take 104mm2, while GT200's 1000GFLOPs take 153mm2:

• RV770 is 11.5 GFLOPs/mm2 on 55nm
• GT200 is 6.6 GFLOPs/mm2 on 65nm

For double-precision:

• RV770 is 2.3 GFLOPs/mm2 on 55nm
• GT200 is 0.5 GFLOPs/mm2 on 65nm

Jawed

So? If all you cared about was theoretical GFLOPs/mm^2, you could do FAR better than RV770. What matters is achievable GFLOPs/mm^2 on real-world (not contrived) workloads. RV770 wins by that metric too, of course, but not nearly to the same degree.

 

[Kaotik]

I don't think you'll see AiB's using GDDR5 on lower margin parts for now (assuming it's allowed). If you want the bandwidth you'll probably have to buy a 4870 or overclock the memory.

Of course GDDR4 and 5 are different things, but there was 3850's with GDDR4 too, IIRC

 

[randomhack]

well the card does do very well with the current bandwidth. but I was wondering if this bandwidth will be enough to give a correspondingly large boost for GPGPU apps.
most GPGPU apps appear to be bandwidth constrained IMO?
perhaps AMD tweaked the cache or the register file .. but that will soon be known i hope.

 

[digitalwanderer]

I think they just packed it chock full of driver bunnies and fairy dust.

 

[Miksu]

It seems that one has to pay attention to the used motherboard (chipset) when reading reviews about 4850's crossfire performance. Tweakdown has tested 2x4850 configuration with P45 and X48 and the results are really interesting.

 

[dizietsma]

So far 4850 does not seem such a great overclocker though. Some sites got 700 no problem, which is the max the current OC software will allow. But another site only got 675/682 on two different brands of 4850, and concluded most of the better chips must be binned for 4870. Also the chip seems to run really hot, could be a negative on OC.

I am wondering how this 9800GTX+ will OC? If it's truly a 55nm part, maybe it will have more headroom. But it wont even be available for three-four weeks, a pretty big negative itself.

I think that is the crux of the matter, how will a 700MHz 4850 stand against the AIB overclocked versions of the 9800GTX + which you would sensibly assume to be out about 800MHz or so given the normal + 70MHz overclocks the AIB typically do.

I think it will be pretty close.

 

[AlphaWolf]

It seems that one has to pay attention to the used motherboard (chipset) when reading reviews about 4850's crossfire performance. Tweakdown has tested 2x4850 configuration with P45 and X48 and the results are really interesting.

That has to be some driver issue, I really can't see that being the extra lanes.

 

[Kaotik]

It seems that one has to pay attention to the used motherboard (chipset) when reading reviews about 4850's crossfire performance. Tweakdown has tested 2x4850 configuration with P45 and X48 and the results are really interesting.

Indeed, the differences between x16/x4 (P35) vs x16/x16 (X38) on last gen cards were a joke compared to this - they actually seem to saturate the x16 PCIE 2.0 buses for good now

 

[dizietsma]

(1) When AMD was burning the midnight oil to push the first Opterons out of the door and to bring x86-64 into the mainstream, Intel's PR response was to mount a nearly year-long, massive negative PR campaign that often amounted to precisely the following quote:

"You don't need 64-bits on the desktop." Short, sweet, to the point and incontrovertibly wrong.

Incontrovertibly right. The first opterons where released back in 2003 or so and 5 years later the vast majority of desktops still uses 32bits. As for milking and innovators, AMD milked the K8 for far far too long.

 

[Sunday]

Even 290 mm2 is pretty close, IMO. Arun is also suggesting that 65nm wasn't as small as it could have been for NVidia.

apparently they've managed to shrink G92 for only 11%!!

http://www.bit-tech.net/news/2008/06/19/nvidia-has-9800-gtx-55nm-g92/1

"On the back the new 9800GTX+ core components look slightly smaller and the new 55nm G92 is 17mm square versus 65nm G92 which is 18mm square! Whoooo!"

 

[Arun]

There's already a thread for G92b now, please don't pollute this one... (and I already saw that some time ago and my estimate based on the die shot and some skewing was 255-275mm², similar to what is now being discussed in the other thread)

 

[Dio]

(2) When ATi struck out of the blue with R300 to leapfrog nVidia's discrete 3d gpu technology a few years ago, nVidia responded with a knee-jerk spate of negative advertising that was so pedantic, so intense, so acrimonious, and so erroneous that few of us who lived through it have forgotten it. And that negative advertising campaign lasted only so long as it took nVidia to stop talking about how "wrong" ATi's approach to 3d was when nVidia was finally able to bring its own similar products to market.

Neither of these negative advertising campaigns was successful, and in my opinion, both of these campaigns did far more to bolster the competition than they did to bolster the the companies that created and funded the negative PR.

You think? I disagree. If you look around at a modern game nowadays, the minimum spec is typically a 9700 or 9600, and the entire FX series is nowhere to be seen. But:
- it still sold
- it took a good six months before the web realised the mud was being slung
- the paper press never really picked up on it at all

Negative press generates a lot of anti-source light and heat on the web, but even there the bad feeling doesn't really last beyond when a decent product actually tips up. You'd like to think that their reputation would be intracontravertibly soiled, but that's just not the case.

It's like negative political advertising. Just about everyone involved in politics frowns on it and tut-tuts, but it works.

Sad but true.

 

[Neb]

That has to be some driver issue, I really can't see that being the extra lanes.

How about it being the cards running out of VRAM (its high res with AA and AF). So they have to swap to RAM and the 2x16 lanes then beats the 2x8 lane configuration in throughoutput GPU <-> system RAM.