backtop

Blog: Hardware DailyTech Digest: Making Sense of "R600"
Kristopher Kubicki (Blog) - April 26, 2007 6:19 PM
Print E-mail del.icio.us 78 comment(s) - last by Proteusza.. on May 14 at 9:05 AM
My personal thoughts on next generation Radeon

Well unless you've been living under a rock since Monday, you've probably noticed we have access to a few R600 cards. 

One thing that concerns me is that many of our readers may think the 750 MHz core clock on the Radeon HD 2900 XTX, if it even comes to market, will be the final clock.  I can almost guarantee you that if anyone decides to bring this to channel, the core clock will get a bump.  Keep in mind that Sven managed to overclock the core to 845 MHz on the XT card.

In addition, I was quite pleased with the HD 2900 XT tests.  I do believe when overclocked -- and this card seems to have a bit of headroom -- the card does well against the 8800 GTS and starts to encroach on the 8800 GTX territory.  In our testing, it doesn't beat a GTX, but then again it won't be a $500 graphics card either.  That task was reserved for the XTX and I think it’s quite obvious that increasing the memory size and frequencies on the R600 GPU won't usurp the 8800 GTX.

I won't pretend that the tests we did were exhaustive.  Since DailyTech does not sign embargos, we sort of get what we can take when it comes to publishing early tests.  Scott Wasson, Ryszard Sommefeldt and Anand Shimpi are true experts when it comes to GPU performance, and if you have any reservations about our preliminary examinations, certainly wait for their benchmarks. 

In conclusion, what did the HD 2900 XTX benchmarks really show?  I think the most obvious answer is that the difference between 1GB GDDR4 and 512MB GDDR3 is certainly not going to be a viable option for R600.  Like we had mentioned before, this card is at least partially scrapped at this point, save a few OEMs who will be putting it into workstations.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
By EastCoast on 4/26/2007 10:24:16 PM , Rating: 2
I understand you guys like to be thorough but are you sure that is the final retail version? I just can't seem to get out of my head how those results look like an OC XT and not a XTX. Come on, spell the beans...




By KristopherKubicki (blog) on 4/26/2007 10:40:42 PM , Rating: 2
The XTX and the XT are identical except for the footprint and memory type. It's likely the core clock will change as well, but the core is physically the same.


By Snowy on 4/26/2007 10:49:31 PM , Rating: 2
This is kinda off the "retail" part, but do you have to use the 8-pin power connector? I've heard you can just use the 6 pin, then if you want to over-clock, you need to plug in another 6 or 8 pin. Can you confirm this?


By Armorize on 4/26/2007 10:52:14 PM , Rating: 2
I posted this in the other article u meantioned but id like to get something out there.

No one has mentioned crossfire vs sli yet. do you only have 1 of each ati card? I'm just curious if their performance will be differant with crossfire or not. Also if these cards have other neat little packages on the cards that are bogging things down, I read on the INQ that it has hdmi sound capabilities I believe it could be that because these arent the real cards that will be released to the public and only referance cards. Just a thought.


By KristopherKubicki (blog) on 4/27/2007 12:35:32 AM , Rating: 6
I was curious about the lack of HDMI as well, though one partner told me on the XT the only HDMI capabilities would be via a dongle (so no sound obviously).

Someone "reached out and touched us" today and gave us 2 Radeon XT cards for use in Crossfire and a different driver (350MB ?!?!). Expect benchmarks this week.


By CyborgTMT on 4/27/2007 1:56:04 AM , Rating: 2
If you can get your hands on that XTX again, I'd be interested in seeing another round of benches with the 'new' driver.


By nkumar2 on 4/27/2007 3:06:25 AM , Rating: 2
as a reader i really like to know the best performance for future references. is there no way u guys could actually get some numbers for these cards under crysis, i would love to know how it holds up in crysis. i dont know i would assume that someone should have the darn demo to play with.


By CyborgTMT on 4/27/2007 5:15:36 AM , Rating: 2
Crysis demo won't be released until at least June, with a limited number beta keys given out (20,000). There are claims of a 'leaked' version floating around the net, but since I have not looked for it, I don't know if that is true or a rumor. Either way it would be an early build not suitable for testing and very unethical to use.


By jazkat on 4/30/2007 8:57:56 AM , Rating: 1
the r600 is built for future games not dx9, in dx10 the 8800 gtx will get a trouncing because the xtx will be 40% faster :) and im only updating for dx10 if i want to play dx9 games il buy a dx9 card now for those who says they get the 8800gtx for both will come unstuck becuase in dx10 the 8800gtx wont be performing as you all think, heres some numbers

The Radeon HD 2000 series supports Superscalar marchitecture and it means that with Vec5 or Superscalar, ATI can process 5 scalar instructions per clock. It has 64 Unified Shaders and by multiplying these two numbers you end up with the amazing number of 320 Stream processors.

The Nvidia G80 GTX can only handle Vec4 instruction in a scalable way, or four independent instructions, and it looks like ATI might have an advantage here. Nvidia can process 128 Stream processor instructions, but at the much higher clock speed of 1.35 GHz, while the R600 can do 320 Stream instructions per clock, but at close to half the speed or 750 MHz.

In raw numbers, the G80 at GTX clocks can handle 172.800 Millions of instructions or Shaders if you like, while the R600 can handle 240.000 Million Shaders per clock. If this turns to be right, ATI could run Shader instructions - especially the unified ones - up to forty per cent faster.

this will only be noticeable in next generation games and benchmarks,

ahhahaha wont be long till we see the dx10 3dmark so when can all witness the gtx gettin a total hammering.


By Proteusza on 5/14/2007 9:05:18 AM , Rating: 2
Supporting lots of shaders is all well and good, but lets face it, current titles do use pixel shaders in fairly large amounts, and can require lots of pixel pushing power and a high memory bandwidth. With its 512 bit bus, the card has the highest memory bandwidth ever seen, the most shading power ever seen, and not a bad fillrate, yet it still isnt great. I think its the fillrate that kills it, that and drivers that dont fully utilize its VLIW architecture yet.

Any future games will still rely on fillrate to some degree. Yes shaders will get more intensive, and maybe then the card will do better. But it seems that it just doesnt have the fill rate to compete with the 8800 cards.

Someone mentioned somwewhere that what nvidia did right is the balancing of the 8800 series. Not too heavy in any one direction - good shader performance for todays and tomorrows games, good fillrate, good memory bandwidth. ATI got the balance wrong and it looks like they are suffering for it.

Why buy a card built for tomorrows games when you can (or could have bought one 6 months ago even) buy one that performs well in todays and tomorrows games?


By Zoomer on 4/27/2007 8:03:10 PM , Rating: 2
That sounds suspiciously like a large, reputable company that also manufactures top-end motherboards, laptops, etc. It was the same for the x1950 series.

Don't post the driver set size!


By AlabamaMan on 4/28/2007 10:18:00 PM , Rating: 2
So Kristopher, are you still on for those tests? The week is almost over.


why i think dt didnt have the latest revision
By nkumar2 on 4/27/2007 12:11:29 AM , Rating: 2
why i think that they didnt have the final revision is this, the card is clocked at 745 and memory at 2040 or something, but the final specs are 800core and 2200 for the memory so dailytech doesnt have the final retail card, thats why i think they their card revision is not final either. shoot call me a fanboy but i flip sides every 6 months whoever is faster.




By KristopherKubicki (blog) on 4/27/2007 12:32:49 AM , Rating: 4
That is certainly a good point that we brought up in another post - the card can overclock to 845 fairly stable. It would be ludicris to think ATI would ever release the XTX with the clocks that its at right now.

That being said, is a 50MHz bump going to do it? I dont think so, and neither does the majority of OEM partners that have turned down this card.


RE: why i think dt didnt have the latest revision
By johnsonx on 4/27/2007 3:11:35 AM , Rating: 4
quote:
.....the majority of OEM partners that have turned down this card.


I think that's the real issue here. If many OEMs are saying that the XTX isn't worth their bother, then it's quite clear no amount of tweaking, OC'ing, final versions, driver optimizations, etc. are going to bring the XTX close to competing with the 880GTX.

It appears to me that the ONLY way the XTX could be marketable now is if they cut the memory size down to 512MB (to lower the cost), and boost the clocks as high as possible. If they can split the price difference between the 8800GTS 640 and the 8800GTX ($450 or so), then they might have something. Sure they won't have the top performing card this time around, but at $450 they'll outperform all the GTS's and nicely undercut the GTX's they can't match. Then the regular XT could split the prices between the 320Mb and 640Mb GTS's, and outperform both. The perfect final piece would be a HD2900Pro with slightly lower clocks and 256Mb RAM, priced between the 8600GTS and 320MB 8800GTS.

In the above scenario then, NVidia would get the fastest video card claim, while AMD would enjoy victory at every price point below $500.

AMD could coin a new variation on the old phrase: If you can't beat 'em, then go between 'em!


RE: why i think dt didnt have the latest revision
By FITCamaro on 4/27/2007 11:43:13 AM , Rating: 2
Correct me if I'm wrong but with a 512-bit memory bus, doesn't the card need at least 512MB of RAM on it? I've never seen a card have a certain size memory bus in bits and have less memory than that in MB. I assumed it was because you need that much memory or the card will run out of memory due to the bus being able to provide far more data than there is room for.


By johnsonx on 4/27/2007 1:19:08 PM , Rating: 2
No, I don't think that has to be true. It depends on the density and layout of the memory chips.

That said though, I'd imagine the mythical 2900Pro I proposed would have a 256-bit memory bus. Just lowering the clocks a little and reducing the amount of memory wouldn't be enough to make it a proper $200 mid-range part, either from a performance or a manufacturing cost standpoint.


RE: why i think dt didnt have the latest revision
By Martimus on 4/27/2007 2:12:27 PM , Rating: 2
I had a 128 bit video card from Tseng labs (the ET6000 VPU, PCI Interface, 128-bit, 4 MB (MDRAM)) and it only had 4MB of RAM. Having 1MB per 1 bit memory bus isn't a need, it is really just an arbitrary number.


By johnsonx on 4/27/2007 2:58:32 PM , Rating: 3
It's been a long while since I had my hands on an ET6000, but I doubt the 128-bit spec referred to the external memory bus width. There was a time when various vendors claimed all kinds of large 'bitness' that had little basis in fact. I could be wrong of course, but I don't think so.

But I do agree there's no need to have as many megabytes as there are bits in the interface; you just need the right number of chips with the right bit width.


By svenkesd on 4/29/2007 6:57:34 PM , Rating: 3
512 bit memory bus means it can read a 512 bit line of memory at one time. 512MB of RAM is the total amount of memory (which can be accessed 512 bits at a time).