HD2900XT Q&A

[f61]

So whats the deal with these... does the 2900XT have too many streams for its memory/bandwidth? Had it been say 160 streams, would it have performed better? Trying to get some facts about this. THanks

f61

 

[Xazax310]

160/320 streams doesnt matter as shown i mean with 96 streams the 8800GTS competes with the 2900XT, doesnt matter as in current performance

I expect though that with its 320 streams and 512/512mb it will tear a new one into DX10 games and the ability to use those all 320 Shader Processors, current games arent exactly "DX10" sadly the first games we get are pretty bad at being DX10 graphics wise and gameplay(lost planet was horrible!) but with the expection of COH which is more of a last minute tacked on API.

Wait till we see second, to third generation titles based entirely on DX10 that can actually use these Graphics cards effectively and the way there meant to be used.

if you ask me 8800GTX/GTS are just the best DX9 cards money can buy with DX10 support(though not well)

Give it a year, my bet is on 2900XT winning out completely(without AA/AF turned on :roll: )

 

[jamesgoddard]

The main problem is all to do with texture units, the HD2900XT has 16 running at 800MHz ish where the 8800GTX has 64 (I believe) running at 575MHz.

The ATI unit has much more shader performance than the Nvidia units, just they have dropped the ball big time with textures.

Anyway - this is where I see the main problem for the HD2900 units..

 

[TheGreatGrapeApe]

well the texture units are an issue, but I'm not sure if they dropped the ball on that as big as on the lack of hardware AA resolve. Straight up without AA the HD2900XT can compete with the GTX, then turn on AA and you have a different story. To me the weakest link is the lack of hardware AA which is the HD2900's primary weakness. Sure it could use more texture power, but no one complains that it does poorly in texture rich situations so much as complain that when AA is enabled it gets clobbered.

I'm sure it would benefit from more texture units especially for AF, but it's interesting when you enable HQ AF on the G80 it takes a nose dive so I wonder what would happen to the HD2900 if it had the option to run the same AF level as the X1900 since it rarely got complaints last generation. I don' disagree that the HQ AF on the HD2900 will never be as good as that on the G80, but it would be interesting when comparing to the X1950 what it would do if given the standard default AF and not forced into HQAF alll the time. Would the TU deficit be as noticeable? I think if they didn't go HQAF they would've been OK for performance, and it would simply accentuate the lack of hardware AA which would still be the biggest sticking point.

I hope they improve both in a refresh, but the more time goes one, the more it doesn't seem that the TUs are the true sticking point after the initial shock of terrible performance, if it was still the TUs alone then the HD2900 shouldn't be able to keep up without AA, when most people complain most about AA performance.

Compared to the GTX the HD2900 needs more TUs and improved ROPs, but of the two I'd say the ROPs and their lack of AA resolve is the biggest issue to most people.

 

[Ancalagon_uk]

The texturing and ROP are an issue, but with the clockspeed being a full 50% higher than a 8800 GTS, this is reduced from a crippling flaw to a power drain.

Its biggest problem actually is that it doesnt have anti aliasing resolve units in its back end.

What this means is that anti aliasing needs to be done by the shaders. Remember that this card is all about its 320 shader processors? now they have to do anti aliasing as well, whereas the 8800 shaders are just shaders.

In my humble opinion, the 320 streaming processors using a VLIW architecture is too complex for a graphics card. They eat up too many transistors, which is why it doesnt have enough texture units or AA resolve units. to compensate for this, they boosted its clockspeed, which makes it run hotter and use more power, and thus necessitate a louder fan.

I like nvidias idea of having the shaders running at a higher clock than the rest of the chip, because then you get extra performance without eating into your transistor count and die size on a chip.

The reason using lots of transistors is bad, is because it increases the size and complexity of the chip. The wafers on which chips are made are fixed in size, thus if you have a chip with lots of transistors, it takes up lots of space, and you cant make so many of them from one wafer. a certain percentage of chips made will be bad (generally from 20% - 70% I think), so having big complex chips can really hurt how much they cost to make.

thats why ATI could have added resolve units etc, but it would have cost a fortune. same reason G80 is two separate chips (one much much smaller than the other).

 

[TheGreatGrapeApe]

but it's interesting when you enable HQ AF on the G80 it takes a nose dive

You obviously do not own a G80 because that is about the dumbest and most inaccurate statement I have ever heard :roll:

Sure Rob, whatever buddy, that you own it doesn't mean you know anything about it, as proved by many Robs before you.

I have a 8800GTX along with about 20 games and the difference in performance going from HQ af to Quality af is pretty much NONE.

That's not what other sites observed, the difference in performance is quite large, and it actually flips the performance deltas in many cases. The IQ difference isn't worth activating it since at default the G80 matches the HD2900's AF level, but it does take a hit when switching, and for you to say the impact is none, proves to me you don't know what you're talking about. It may be minimal in your opinion, but it's obviously not going to be none when forced through a different algo at higher quality.

Pretty dramatic performance drop in fact;
http://www.ixbt.com/video3/r600-part3-d.shtml#p18

I have tested both methods back to back in many games and most of the time the difference is only around 1-2fps at the most, that is a far cry from taking a nose dive as you so claim :roll: I would be interested as to know where you got this info from because it certainly doesn't pertain to what I have seen with my own 2 eyes.

Well your two eyes, may just not be good enough. Just like people who say 16XAA versus 8XAA on an LCD makes a big difference when they don't even have the display depth to notice the corrected pixels.

The G80 has proven to be very efficient with both AA and AF.

It's fine at default settings, but the G80 is neither efficent at shader based AA or HQ AF.

Gawd the things some people say :roll:

Yeah especially by those who use the name 'rob'.

 

[realibrad]

so who is this rob guy that i keep reading about. is it just one guy who always puts rob in his user name?

 

[f61]

I see... lack of 2 AA units forced double duty on the shader processors. An ignorance of Tesselation processing bit their arse big time and forced over-complexity.

But on the other hand NV didn't hit the mark either with the 8600 series, which given 256 bit and 64 streams might have stolen the X1950 market.

Both look a little loyal-greedy at the appearence of being, uummm, stupid.

f61

 

[JAYDEEJOHN]

Ive read alot of benches that show HQ AF kills the 8800 series. I know better, and dont have to own A GTX to know it either. rob ignores facts such as this, and goes by experience, but just like the blind man and the elephant, experience doesnt always let you see the whole picture. Tho there once was a rob that had GTX in sli, and maybe using HQ AF with that might make the impact less severe. Tho, that couldnt possibly be this rob, as he only has 1 GTX.

 

[JAYDEEJOHN]

The 8800 series doesnt do shaders at all like the 2900, thats a fact, and Ill just take the Apes link, as this is old news. And fact as well. This doesnt mean that the GTX is one hella kick card, just some facts about it and the 2900, oh and also the GTS as well

 

[f61]

rob:
Post your links to your data. I'm sure its validated on the internet.
f61

 

[realibrad]

rob is starting to sound like that phrozt guy. altho, rob seems to have been penetrated much deeper and with something much larget than phrozt.

 

[JAYDEEJOHN]

When I have the time, Ill test it on a GTS. 2 or 3 %? OK, we will see

 

[JAYDEEJOHN]

Youre the one sounding fanboyish. Do you still own your 1900? Just because I like a product, I dont dump it or the company if something new comes out. This aint the Jones forum, its Toms, and it aint about keepin up either. When I get to the benches Ill get to em. If youre right Ill know, and even let you know. If youre wrong, well... we will see

 

[f61]

rob:
post your data, lets see what you're talking about
f61

 

[robspimpin8800]

Youre the one sounding fanboyish. Do you still own your 1900? Just because I like a product, I dont dump it or the company if something new comes out. This aint the Jones forum, its Toms, and it aint about keepin up either. When I get to the benches Ill get to em. If youre right Ill know, and even let you know. If youre wrong, well... we will see

So im a fanboy because I sold my X1900XT after owning it for nearly a year in favor of a upgrade to the next best card? :roll:

Come one now, you can do better than that jonnyboy :roll:


I have one computer and typically I upgrade my GPU about 1 time a year to whatever is currently the best rather it be NV or ATI. I have no need to hang onto every GPU I have ever owned instead of being smart and selling them on ebay when I am done with them which helps me to recoup some $$$ for my next upgrade. And when the G80 was released I dont think anyone in their right mind would dispute that it wasnt the best card when it was released and hell it still is. 😀

 

[robspimpin8800]

rob:
post your data, lets see what you're talking about
f61

Well to be fair im not the one making the ludicrous claims which speak out against what every other forum or benchmarking website has ever shown but here I am yet again the only one that can speak from exp who actually owns an 8800GTX while also being the only one offering any sort of evidence.

http://www.penstarsys.com/reviews/video/gigabyte/8800gts_640/gb88gts64_7.htm




The difference here is absolutely minimal. The G80 is a texture filtering machine. It makes no sense not to use HQ filtering. The next test shows how much of a hit going from no AF to 16X AF imposes.

 

[robspimpin8800]

Crickets

Good night all. :lol:

 

[TheGreatGrapeApe]

What this means is that anti aliasing needs to be done by the shaders. Remember that this card is all about its 320 shader processors? now they have to do anti aliasing as well, whereas the 8800 shaders are just shaders.

Yes, and that's the mid-way point that ATi abandoned and why I mention it. Shader based AA is a very DX10 thing to do and is a required option now (as in support it if the app calls for it) and a requirement of D3D10.1, so while this definitely hurts the R600 on legacy titles, it may be less of an issue in the future as shown by the G80's ability ot handle shader based AA in COJ and nV's vociferous complaints about it's use (despite being a D3D feature).

thats why ATI could have added resolve units etc, but it would have cost a fortune. same reason G80 is two separate chips (one much much smaller than the other).

Clocking the shaders at a higher speed is great idea, but it hasn't shown itself to be incredibly efficient compared to previous architecturres, especially at idle. Also the added cost of the NVIO is something that though though of a 'smaller' also requires additional board space, wirign, and still costs extra compared to the slightly bigger chip itself. The issue would be yields that would determine the efficacy. If the failure rate were much higher due to the addition of the simple part of the chip. And of all the things on a chip to worry about the NVIO portion (RAMDACs , TMDS, etc) is an interesting part to make modular if you are going to add things to it (like dual-link HDCP support), but if not then it makes the chip look like an afterthought because they ran out of room to include the I/O. Although if you're short on space it's the easiest to also turn into it's own chip because it's pretty distinct from the rest of the process and not required for feedbacketc, and just accepts output from the ROPs.

The layout of the shaders has been debated back and forth and both show their benefits under different situations , but sofar neither has been given a ruely good test like Crysis and UT3 should do. Evenn FSX in DX10 should offer some interesting differences from what we've seen sofar.
\

 

[TheGreatGrapeApe]

Well to be fair im not the one making the ludicrous claims which speak out against what every other forum or benchmarking website has ever shown...

http://www.penstarsys.com/reviews/video/gigabyte/8800gts_640/gb88gts64_7.htm

That is your 'every other forum/benchmark' info? You'll need something better than that, as Penstarsys is more of a personal site, like Sudhain's than a quality review site like Digit-Life, Xbit, THG, B3D, etc.

Like B3D says, the quality and horsepower are there, but it's up to nV as to whether they optimize or not when HQ isn't forced on, and from the digit-life tests it's pretty obvious there's still some optimizations there.

http://www.beyond3d.com/content/reviews/3/4

"While NVIDIA are certainly able to do per-app detection to pare that back and go back to older variant level selection (the choice to do so still exists for the hardware), we're hoping that's not a choice made for many applications given the filtering horsepower G80 has.

We'll also pipe up about turning off all filtering optimisations out of the box by the driver, at least for SKUs like GeForce 8800 GTX and GTS (and the same was said of certain GeForce 7-series products) where the chip's configuration is so performant that it's something we urge users to do. Quality mode still applies the ages old trilinear opt, which we'd like to see disappear (by making High Quality the default mode in the driver and letting the user apply tradeoffs, not the IHV)."

Like I said, quality mode CAN still be better than ATi/AMD's HQ, but if there are optimizations then you would notice it when turning on HQ AF, if there was no difference in performance, then there wouldn't be optimizations. Ati by default has the optimizations off because it's default is HQ mode. Whereas you never know with the G80 when it's set to default quality mode. So more than just a single test in a single game at a single point is required to find that out and Penstarsys didn't do much to check that now did they?

 

[f61]

Sorry about my mid-flame post. I get tired of personal attacks disguised as flame-wars. Its a bit over my head, but it looks like Ati focused more on what DX10 is supposed to be, but really needed AA processing. NV read the market right now, and built a high end card for today's rather poor offerings of DX10, but allowed a big improvement to the 7xxx series, which is really what NV fans want. I did point out that perhaps fewer streams might have helped, and thanks to ancalagon_uk for confirming my suspicions. Ring-core is a nice idea that worked OK in the 1950 series but it IS rather complex and does limit stream processing due to mgmt overhead. Maybe the HD2800 offering(s) will balance this better as roadmapped at 256bit/256-512Mb/96 streams. Maybe the AA can squeeze in there. Still, the 88xx series is aimed higher, but the ATi implementation might do it some justice. We'll see.

In my humble opinion, the 320 streaming processors using a VLIW architecture is too complex for a graphics card. They eat up too many transistors, which is why it doesnt have enough texture units or AA resolve units. to compensate for this, they boosted its clockspeed, which makes it run hotter and use more power, and thus necessitate a louder fan.

One other thing strikes me about the detailed testing at the ixbt link: The HD2900 started to pull ahead further at increased rez. Frankly, it was nearly amusing at 1600x1200. I note this for an important reason, most fans are clinging to the highest FPS #. As plainly indicated that occurs at 1024x768 or 1280x1024, rez's that wil be obsoleted in 12-18 months. So if one looks forward in time to a screen rez of 1600x1200, the ATi system actually looks better: after all one does turn off AA/AF to achieve smoother frame-rates, ahem, at least nowadays. And most newer sites don't test at 1680x1050 or 1920x1200. I see a LOT of 1680x1050's going out the door at e-tailers, and think reviews should incorporate wide-angle when possible. Thats part of the Vista experience, that like-it-or-not will also creep in during the next 12-18 months.

Reviews would do JQP a great service by also testing, or revisiting tests using the Vista/DX10/wide-screen POV. Its getting busy these days, but, IMHO there will a need for this shortly. Toss in an E6600 on a 1333 MHz BW as a 'mainstream test'. Full tilt to the bench facility QX6xxx DDR3 etc also for peak performance, whatever.

For the record, I am quite dissapointed in BOTH ATi and NV so-called mainstream attempts using 128bit. These are non-events to the DIY, and specifically targeted to the box-with-a-badge retailers. Ask for an X1950 or 7900 series instead and get more outta the box.

f61

 

[clue69less]

Sure Rob, whatever buddy, that you own it doesn't mean you know anything about it, as proved by many Robs before you.

You mean like RobsX2? Did he get banned or something? How could that happen to such a nice guy...

Well your two eyes, may just not be good enough.

I think Kelso woulda called that a burn.

 

[Heyyou27]

I noticed a slight performance hit in Oblivion with High Quality anisotropic filtering in Oblivion.

 

[rammedstein]

well, at home i use 2x 2900XTs in crossfire, i'm hanging out for an x38 board because the pcie bandwidth isn't enough, even when overclocked.in this setup they perform in the same ballpark as the 8800gtx sli, now, people may say this is a waste but i use a 30" dell monitor (30007WFP, 2560x1600), and i don't use AA but i do use AF, the reason i don't use AA is because at that resolution it is hard to find jaggies when playing normal and the performance hit isn't worth the visual detail increase.

 

[r0x0r]

rob is starting to sound like that phrozt guy. altho, rob seems to have been penetrated much deeper and with something much larget than phrozt.

Not entriely true. Phrozt has *gasp!* admitted he was wrong right here:

http://forumz.tomshardware.com/hardware/Big-Price-Drop-Sapphire-HD2900XTftopic-242434-days0-orderasc-50.html

But yeah, rob is one of a kind.