[B3D Analysis] R600 has been unleashed upon an unsuspecting enthusiast community

typo?

"such that board volume is higher than any of the board boards on test"

 

VLIW, specifically, is not really a change - its been this way for all the DX9 generations, R600 just further increases the granularity.

Optimisation is transparent to the developer - our shader compiler does this already, and its highly effective at its job even with R600's new structure.

Do the analysis on some recent titles, you'll see that the overall ALU:Tex ratio isn't that far away. The threading is there it minimise the areas where you are bound by one or the other.

 

This is going to sound like rubbing salt into the wounds, but the R600 is an utter let down, plain and simple. Saying it doesnt look good because the G80 is so amazing is a rather large copout. The 8800 series has proved fantastic, true, but its also, in many respects, what you'd expect from an entirely new core when done right. Infact its improvement over the G70/NV40 is what we expected the R600 to be from the R420/520/580 in respect to how much was improved. The R600 realistically came out looking like garbage in comparison, and the key issue i have with it is the simple fact that it isnt even a reliably huge performance leap over the X1950XT in some real world testing let alone the competitions new stuff, how does that happen? Couple this with the fact that its been delayed 6 months, the fact that the drivers are very immature to say the least, AA scaling is awful for the amount of bandwidth the card has, they didnt really improve filtering levels, the noise of the HSF has prompted reviewers to vocalize that they wouldnt use the card where audio mattered, AND it draws more power (substantially more compared to a GTS) and has issues with heat, and you have something that mirrors the opposite of what many of us were expecting.

R600 was hyped, no question, but this launch simply looks terrible with far too many glaring issues.

Well AMD has no midrange/Mainstream 6600GT-like part as it seems to be on a further delay, so they cant do that which means further Geforce 8 series monopoly. And as far as the 6800GT, ATI really had no competition for that part Geo, which is why it sold so well. The Pro was there to meet it price wise but couldnt match performance at all. Being able to OC the thing to near Ultra clocks helped too. This time round that isnt the case. Also Geo im not sure how well you are aware of how aggressively nVIdia and its AIBs are being, but eVGA and PNY have both dropped the price of their 8800GTS 640MB cards down to $330 and BFG is trailing at $340. As Driverheaven said, in order for the 2900 to even start to look good with this type of performance they'd also need to be right in that $325 price point area. But here again you have the result of the massive delay in their launch which i said would cause this kind of trouble. nVidia and partners have made a very pretty penny and can do these types of price cuts without issues, AMD on the otherhand really needs all the money they can get. In the end this is about as disastrous a situation as you can get.

 

You're actually some distance under it for a good few games (at least in their main shaders), too (so comparatively more matched than G80 in that sense). The chip can seem imbalanced, but then you can make any chip appear that way if you really want to.

There's some neat stuff in the works here that'll take a deeper look at just how clever the shader assembler is for R600 (it does some pretty wow things at times, with packing and register use). And hopefully we'll get Eric on the B3D mic to talk about it as well

 

Sorry, must have misunderstood this post by you:

Which was an answer to my comment that "biggest flaw of nv30 was that it was utterly worthless in dx9, and there's no sign of such on r600 & dx10 yet"

 

Nice selective editing there. Left out the bit where I said the engineering comparisons to NV30 had validity.

 

I find the R600 to be a "meh". I like the architecture, I think it has an interesting future, but pushing price/performance on a much delayed release day rather than trouncing the competition leaves something to be desired.

 

well it isnt just you who's said the G80 being so damn nice is causing some skewing of how people view the R600, so dont feel singled out just cause i quoted you. Your post was just best because you went into depth of what some others were saying .

 

Well, you and I are refighting my fight with Uttar re NV30 vs R300. I've always said that at least *part* of the story is that the other fellow pulled off a major positive surprise. He usually shows up at that point to say "Nope, NV30 suxxored". Tho typically he'll take 500 words or so to do it.

The reality is, in my view, don't get caught in the fallacy of the false choice. There is room for both to be true and for us to just argue (in the good sense of the word) over the degree each is responsible.

 

And my responce was to the later half of your statement. Was NV30 sucking in DX9, but the R600, as it stands right now, aint looking much better in DX10. But then again, the G80 is marginaly better, more so in another game. Then again, these are DX9 ported/patched to DX10. True, ground up DX10 games might be different.

 

There is some misterious problem with R600. In paper it is a beast. We only have to look at its father: xenos can move without problems DX9 games as Gears of war with much less shader power. What is the difference ? many transistors for enhance that shader power and allow dx10 stuff and a very different ROP system...

 

Well, we know by definition the ROP relationship is different than Xenos, so that's one place to look. I'm not sure what comparo benchies you've seen with Gears of War to show how Xenos does on it compared to, say, G80.

 

The pixel shader scheduling could be optimized for the applications, and we could and should see improvements in that field, but missing or defective MSAA hardware is simply overkill for overall performance at higher resolutions/with AA.
Xenos, AFAIK, misses the compression/decompression units in ROPS, but it has EDRAM and huge bandwidth to compensate, where MSAA hardware is in place.

Now the question is: R600 missing MSAA hardware is a design "feature "or there was some realization flaw (as Anand was guessing)?
In the first case, I cold call the decision almost suicidal.
In the second case, there's the possibility that RV630, R610 and R650 will not have this flaw, so we could see a different picture when they are out.

 

Something I was thinking about as some friends of mine called me up to join them in Vanguard.

It'd be interesting to see how R600 compares to G80 in this game. Granted the game is seriously CPU/Memory bound but it's also Graphics bound to a degree. And it is so far, the most graphically amazing game (except for the lack of AA which makes me not want to play or look at it) that I've seen. Oblivion is better in some areas, but there's just some things that make me go ga ga in VG. Volumetric clouds and really well done soft shadows are a couple.

I'm not so much disappointed at the R600 as I am about how long it will take them AMD to get the drivers in such a shape that it can show what it can do.

Then again, I had no real expectations about the card other than hoping for greater and better levels of AA and AF. And while AF has been a disappointment, I have renewed hope that the AA might help with some of the really nasty rendering artifacts that exist in many of todays 3D games.

I will say however, that if AMD had launched this at a higher price point. Woooo, boy would they have been asking for trouble.

Regards,
SB

 

However, issues with utilization on VLIW predate DirectX. R600 doesn't look right now to be a very good counterexample.
The proof of the pudding is in the eating.

Currently, an R600 with a significant advantage in peak ALU capability (which an effective optimizing compiler should be effectively scheduling for), performs at the level of a lesser speed grade of a chip that has much lower peak capabilities.

Perhaps I should amend my statement to say that
(complex threading - utilizing TLP)+ very wide superscalar+VLIW+SIMD+an unknown number of hardware flaws = unimpressive utilization.

EDIT:
If ALU scheduling is not at issue, then I am open to other input.
It does seem like a good number of other things surrounding the ALU arrays did not end up as polished.
I'm quite eager to see what R650 can do, since it will likely get the hardware surrounding the arrays that R600 should have had.

 

Can't wait

 

Making a rule of three ( is it said like that in english ? ; ) with Stalker...

 

At this point in time this is the only thing that is keeping me interested in the HD 2900 XT as an upgrade option...From Tech Report...

He starts off saying basically like everyone else he was quite prepared to not like CFAA and it's side effect of blurring things...

But then after using it, he goes on to say...

And this...

It's nice to know that 8xCFAA (4x+Wide Tent) is basically the same speed as 4xAA. B3D's IQ analysis just can't come soon enough. I'm almost willing to throw away 400 USD on this card just to see if it will help with many of the aliasing artifcats that can be seen in so many games.

I wish I could see it in action. I've already seen and played with the 8800 GTX, and it doesn't do much to address these situations.

If you can't tell, speed comes second to me to removing things that draw my attention away from what I'm playing.

Regards,
SB

 

I'm interested in whether G80 has the same sort of command processor, driver unloading, setup functionality as R600.

 

One thing has to be said though, even if the drivers are still in bad state based on some game results, it's great to see that there's practicly no difference if your run Vista or XP