Just asking, wouldn't the 1mb cache(if it exists) be used to to help with, not just graphics, but the HD Sound and Physics that AMD are gonna tout?

US

Too many variables and not enough known about the vertex buffer fetch, texture buffer fetch and filtered fetch of both architectures, I'm afraid.

For purely bilinear texture fetches, R600 has half the rate of G80 per clock (16 versus 32). In theory G80 should never fall below this rate.

For trilinear filtering, it's not possible to say what R600's rate is. There's every likelihood G80 is twice as fast. Same applies to anisotropic filtering.

Because these are unified architectures, the shaders can request vertex fetches, interleaved with filtered texture fetches. Additionally, non-filtered fetches may also be interleaved. I don't remember seeing any information comparing how G80 handles these 3 types of fetches (rusty memory alert). Consequently I don't know if they are all performed by the single group of TA and TF pipelines.

Whether they are or not, trying to theorise about their performance is bloody hard, because there are so many momentary factors at play (the interleaving I mentioned is just one).

If G80 does indeed send vertex and texture buffer requests through its TF pipelines, then it will indeed suffer in terms of absolute throughput. But the workload of VFs (at least in DX9 games) shouldn't be very dramatic, while non-filtered fetches have historically always clogged up the filtering units...

Jawed

Well, that said, a cache of any reasonable size is helpful, of course, considering the physical restrictions of the design. The other thing is how the cache array is implemented and organized.
Take a look at the Cell/B.E. (a.k.a. raytracing monster), although not everything there is in the category of caching. ;)

A lot of what I've read of the R600 reminds me of what I've read about the physx PPU.

40 is for a crossbar. 200 is for DDR memory. So the baseline in our little "ring bus is doomed" scenario is 240 clocks, apparently.

But the penalty for DDR is 5x bigger than that for a ring-bus compared with a crossbar (200 compared with 80-40=40). Sure, it's an extra mile, that 16%, but look how many miles have had to be travelled to build the GPU in the first place. With the architecture in place, surely sizing buffers is, relatively, trivial. The metrics for all these buffers still have to be in place, scaling for worst-case scenarios is still needed, etc.

Jawed

Mao , this score not is in DX 9, not is with HDR .... in one 8800 GTX the average is 30- 40 fps in this point - in DX 9 - . And sorry ;)

As far as I know, this could be complete BS...

If 8pin PEG power connector had something else than 2 more grounds (+12v/sense + ground for example, allowing a significantly lower voltage drop) it would sound ok, but as is I strongly doubt the card's driver could know what you're using.

Maybe so, but until the Official release I guess we are in the dark.

US

Of course this could be Fud
http://www.fudzilla.com/index.php?op...d=791&Itemid=1

US

I've got OCZ GameXStream 850w with 70A on 12v rail, you think it'll work with x-fire?

What PSU did the guys who benched use?

I'm Sure...

So it does seem like R600 is going to be heavily texture limited..sigh when is ATI(AMD) gonna wake up..

I hope they at least put 24 texture units into the R600 refresh.

One thing you knew Nvidia would do is make sure they have tons of texture throughput with G80..and sure enough they did, with 32 units.

Does it have a 8pin connector?

I was gonna get the 700W version then realised that I might be making a mistake and cancelled.
OCZ GameXStream 700W Power Supply

I do see that the 850W has two PCI-e 2x6 pin which is good, and I also see a 2x4/1x8pin(although they say it's for the CPU/Mobo?), so it looks good although people might argue differently.

http://www.ocia.net/reviews/ocz850/page4.shtml
http://www.cluboverclocker.com/revie...m850/index.htm

I wonder if there is any chance for X-Fire (Dual) R600's working on the 700W version?

US

How many amps does it have on 12V? That's all that matters there.
And yes, those on the first link, where there's 2x4 pin stuff, they are for mobo, not video cards.
I'm not sure if there's any PSU's out there yet with 8pin PCIe plug

Ye, was wondering .. so it won't do hey?

US

Well, my PSU isn't very old, hope they send a few cables in the x-fire package, if not where can you get such cables?

http://www.tweaknews.net/reviews/the...ply/index3.php

Thermaltake's ToughPower 1000w supports 3x8pin and 3x6pin

US

A person i trust received a R600 XT sample for testing, and found out that he couldn't overclock the core more than 2 to 3MHz, even with improved air cooling (he's an experienced and unbiased source, so i tend to believe him).
It's not yet known why are the power and pin requirements are so restrictive, but he did test with two high-end PSU's (an OCZ GameXtream 850W and a PC Power & Cooling 850W).

Is the 8pin plug absolutely needed for even basic overclocking or is it an intentional driver lock out ?
If that's so, then the price advantage of the R600 XT might be negated by the need to purchase a new -and certainly expensive- power supply.
Where's the 8800 GTS 320MB competitor ? A R600 XL would sound terrific, if true.

Actually with HDR ON but Dynamic Light OFF the 8800 GTX averages much higher than that. Techreport has the 8800 GTX at stock averaging 137.4 with HDR ON at 2560x1600 res.

http://techreport.com/reviews/2007q2...a/index.x?pg=4

So, the numbers that Mao posted seem extremely reasonable with HDR ON and Dynamic Lighting OFF.

Regards,
SB

You may not be able to overclock but it should run it just fine. 450 watts max for 2xR600's. Leaves roughly 250 for the rest of the system. Assuming you aren't running 20x15k RPM SCSI drives in your computer you should be more than fine. That is, of course, assuming you have a quality 700 watt powersupply which should be the case with OCZ, as FSP group manufactures most of their power supplies.

Or with a 2x6pin -> 1x8pin converter you should be able to overclock with the same powersupply as it has 4x6pin connectors. Of course, then you wouldn't be able to do Crossfire without using 1x4 pin -> 1x6pin converters. Boy would that end up being mess. Not to mention Overclocking a Crossfire setup on a 700 watt PSU might be pushing it. :D

Regards,
SB

That's what mao5 said and it's seems intentional as he said drivers older than 8.361 worked with 2x6pins and the 8.361 drivers only worked with 1x6 & 1x8 .. so seems pretty intentional.

Hopefully ATI will sort it out with 2x6pins again in future driver releases.

US

From the above link:

"For this test, we set the game to its "maximum" quality settings at 2560x1600 resolution. Unfortunately, the game crashed on both GeForce and Radeon cards when we set it to use dynamic lighting, so we had to stick with its static lighting option."

LLB

Which is basically just repeating what I said. :)

Regards,
SB

Well, Bob was hinting at 100 cycles to cross the die. If that's the case, the difference is 100 cycles. Sure it's trivial to increase buffers and a ring is not the harbinger of Doom. Just more expensive, especially taking into account the overdimensioning, and reducing perf/W, which hasn't exactly been a forte of ATI lately.

Anyway, I guess I made my point. :wink:

So, then all of ATI's engineers were blowing up our arses by claiming the Ringbus scales better with high speed and a wider bus?

I'm a bit puzzled as to why they implemented it then if there is absolutely no benefit considering they already had a crossbar architecture that presumably worked well.

Regards,
SB