HD 4870 review thread.

Regarding the fan speed profile stuff:
I've been fiddling with that on my 3870, does anyone know if there are other algorithms than Manual & Automatic?
Not too keen on setting a Manual percent (only goes at that speed always = fried card if you're not careful!)

I would like to be able to set a target temperature & have the fan crank up as needed to keep it at that temp.

Or maybe somehow make the Automatic one more agressive in when it kicks in & the rate it ramps fan speed?
At the moment it only seems to kick in at 85C, I'd rather it kicked in somewhere in the 70s.

I have been using Rivatuner to manually set fan profiles/speeds but it only seems to be able to pick up 3 distinct speeds for some reason.
I'd prefer an ATI official & more smoothly ramping alternative.

 

Um a GX2 is much faster than HD4850 or the HD4870. The numbers for the 9800GTX is in between the HD4800 series cards.

 

I think they made an innocent mistake there. That result contradicts their own results a week ago. (RV770 tech preview)



From this article -> http://www.xbitlabs.com/articles/video/display/ati-radeon-hd4850.html

As you can see, HD 4850 numbers are consistent. It looks like they used a wrong data for 9800 GTX. (Possibly GX2, yeah)

 

Not always

No it isn't. The 4850 is faster than a 9800GTX in about 80+% of reviews and tests. Half-life 2 isn't one of the tests the 9800 generally wins and in this particular test they have it beating a 4870 /boggle.

 

So you really think that the reviewers go out of their way to make lopsided reviews and enjoy it so much that they have been doing it for 8 years? I have my doubts on that. Do you really think that someone would put together a comprehensive review like this with such a childish motivation? Before making such nice generalizations about them, I'd be really sure that it's not just a different testing methodology with different goals (the alpha AA) or a simple mistake.

 

Xbit have been quite reliable. I think its an exaggeration if you call them controversial over one mistake.

 

HIS RADEON HD 4870 512MB PCI-E@iXBT (in Russian, still no english version available).
All games results: http://www.ixbt.com/video3/rv770-2-part3-d.shtml#p18

 

I like that 'all graphs on one page' bit

 

I'd expect the benchmarkers to know what results that should be expected.

When a card's scores goes haywire in not 1, but several games on the same benchmark, shouldn't it be the reviewers picking it up first before the readers?

It's not been rectified yet.

p/s: What do you think about the Benchmark Reviews 4850 review instead? I'd like not to make generalizations, but you really should see that one.

 

Very true. Just as with NV's over-heating mobile Geforce situation, better to catch the problem before it becomes a problem.

Yeah... Sure the 9800 GTX is faster than the 4850 - if we observe the following conditions:
- TWIMTBP title that has shown an architectural preference for NV HW over time
- low resolution (biliinear fillrate bound)
- lack of AA/AF

Remind me never to read that site again.

 

Well, isn't that kind of "high utilization" ideally suited for DX10 support?

 

Indeed. Which has already borne out across multiple DX10 titles, observing performance scaling from 3870 to 4870. ATi is definitely holding their own even against NV's best, and it only gets better as the ALU:TEX instruction ratio increases

 

Errm, I wonder how things would change if ATI implemented proper AF. It's easy to be faster when you're underdsampling - if you're into texture flickering, by all means, go ATI. I'm not sure how valid a performance comparison is when switching off A.I. on ATI cards does not disable all optimizations detrimental to image quality while running HQ on nVidia cards does. Plus, most reviewers don't even bother to disable A.I. despite the horrendous degration in image quality A.I. causes. It's kinda like a reverse situation of what we had back in the days of the G7x, only this time ATI is the culprit - even A.I. off on the 48xx cards is still noticably worse than what we had on the R520.

 

I don't buy this "proper AF" principle. I have identical texture filtering quality (subjectively) with G92 @ 16x HQ in all titles (and 16x CSAA - thanks for that tip ChrisRay!). My old X1900 (R580) setup @ HQ looked the same.

I'm even playing the same games! Fear, Oblivion, many Source engine games from Valve, Prey, FS 2004...

Granted, I'm on a different screen (42" Panasonic Plasma 1080P @ 8' vs. 21+" Samsung LCD 1600x1200 @ ~ 2-3 feet), but I feel image quality has not changed drastically aside from what my eyes distinguish due to the screen change.

 

AF quality with A.I. enabled is worse than nvidia (G80 and higher) on Q, and much worse than HQ. With A.I. off (which also switches off all "legit" optimizations like shader replacement and at least on the R5xx cards butcheres performance) it's quite a bit better but still noticably worse than nvidia HQ.

Coda over at the 3dcenter.de forums did a A.I. on/off comparison (animated png, doesn't work in IE):
AI off
AI on

Looking at this shit, there seems to be no excuse for benchmarking ATI cards with any other setting than "AI off". Did anyone actually do that? The fact that ATI refuses to seperate filtering quality "optimizations" from other forms of performance optimizations is their problem, really. What makes things worse is that "AI off" on R6xx and newer ATI cards does not eliminate all filtering optimizations anymore, like it did on the R520. There is simply no counterpart to nvidia's HQ mode and there hasn't been for quite a while. I guess people just stopped looking and ATI got away with it.

Coda said he'd follow up with an article investigating the issue further and include G80 comparisons, so maybe we'll know more in a few weeks.

 

The Catalyst AI driver behaviour has not changed between R5xx and R6xx, so Catalyst AI off should still disable texture optimizations. I will investigate further to make sure.

The animated .png doesn't seem to be working for me in Firefox.

If someone reports an apparent image quality issue with Catalyst AI on or off then I will make sure that it gets investigated.

 

There's a .gif version:

off
on

The effect looks even more pronounced in the png for me but that might be because Opera plays the png frames faster than the animated gif.

The Computerbase.de test of the GTX 2xx also noticed worse texture flickering of high-frequency textures on the RV670 compared to the G92/G200 but they don't provide videos/animations. The only advantage gained from switching off AI was proper trilinear filtering. The flickering textures issue was not remedied by turning off AI and AFAIK nothing has changed with regard to filtering with the RV770.


Edit:
IIRC Terry Makedon once claimed in a PC Games Hardware interview that A.I. had no negative impact on image quality, which promted one of their writers to sarcasitically comment on the 3dcenter forums that we can all imagine what this sort of conviction will mean for R600 image quality. Looks like he was right.

 

I will look into this - as I said the driver behavior has not changed since r5xx. The anisotropic filtering and LOD calculations on R6xx/R7xx are similar to those on R5xx, but are calculated at a higher precision (so they should generally be more accurate now than they were on R5xx)

Anyway I will look at the articles in question and see if I can reproduce the results.

 

Bottom line:
HQ AF 16x on both IHVs HW (Cat AI to low/normal whatever it is) = no discernable difference in filtering quality. With 48x0 hardware using any other configuraton is a waste.

 

AI/on got worse with regard to texture flickering going from R420 -> R5xx but that wasn't as much of an issue because AIoff remedied the problem. With the R6xx, AI/off doesn't help with the texture flickering anymore. AI/on quality remained the same as on the R5xx. The real issue seems to be that AI/off quality has gotten worse compared to the R5xx.