HD 4870 @2560x1600

[lopri]

This is my initial impression on HD 4870. I purchased it to replace my beloved HD 3850 in my 'work' system (see sig #1 or a pic). My other system (self-designated 'gaming rig', see sig #2 or a pic) has two 8800 GT's in SLI configuration, and despite some frustration it has been fulfilling the duty of heavy graphical lifting. I think the roles might change starting now, though.

In short, the thing rocks. There are so many positives and only a few (but significant) drawbacks. I will start with the positives.

1) The drivers are surprisingly stable: I'm no fan of 'beta' or 'hot-fix', but the next Catalyst release is far and away, which left me no choice. But the outcome has been so far satisfying.

2) G80 taught me what 'texture' or 'AF' is. RV770 teaches me what 'AA' is. This is somewhat personal, though. Having a 2560x1600 panel means not being able to apply AA in most games. With HD 4870, I now can enable 4AA or even 8AA, and things seem to pop out towards my eyes.

3) Stutter-free gaming. I was so accustomed to various stutters (macro, micro, or what not) of G80/G92, I didn't know gaming could be this smooth. (Note that this doesn't always translate to FPS.) For instance, when I enter a new map or there is a texture swap in the game world I have/had to do a 360 degree turn with 8800 GTX and 8800 GT SLI. It became a sort of habit and I find myself doing it all the time. Enter a new room? Fast travel? Do a quick turn-around dance!
NOT ANY MORE with HD 4870. This difference feels so dramatic and I still habitually dance around whenever I feel like I have to, but to my disappoint there is no such hiccup with HD 4870. I'm sure some of you understand what I'm talking about.

The comparative screenshots could be biased, in that I wanted to see how HD 4870 performs where 8800 GT SLI sucks. It's entirely possible there are places where HD 4870 lacks, but I think there is a distinctive trend how/when G92 suffers and I believe anyone with a G92 can verify it. Check the Fraps counter to upper-right corner of the screenshots. Both systems are overclocked. (HD 4870 @790/990 | 8800 GT SLI @700/1000)

Call of Duty 4 | Max in-game setting including 4AA

HD 4870
8800 GT SLI

HD 4870
8800 GT SLI

HD 4870
8800 GT SLI

But benchmarks show 8800 GT SLI beating HD 4870? Here is why.

HD 4870
8800 GT SLI

What's happening is that 8800 GT SLI runs this game @60FPS+ like 90% of the time. The problem occurs in the other 10%. The FPS tanks to teens or even below teens out of nowhere. And you can't 1) predict when it's going to happen, or 2) avoid it even if you know where it happens.

Same thing happens to Oblivion as well. Without any AA, half of the time the 8800 GT SLI stays around 100 FPS (albeit stuttery). As a matter of fact, if you don't enable AA the FPS will stay above 30FPS like 98% of the time. But the moment you enable any kind of AA, you will experience game-stopping performance drops here and there. My favorite spot to test GPU performance in Oblivion is foilage around 'Fort Sutch' which is located above city of Anvil. (very west of the map)

Oblivion | Max in-game setting and tweakguides.com's tweaks applied.

HD 4870 @8AA
8800 GT SLI @4AA

HD 4870 @8AA
8800 GT SLI @4AA

The difference is so dramatic it's not even funny.

Gears of War | Max in-game setting as below

HD 4870
DX9 path: Totally smooth 62FPS (FPS cap). 'Buttery' is a perfect description.
DX10 path without AA : 30~50FPS very smooth
DX10 path with AA: 2~3FPS in the menu screen :laugh:

8800 GT SLI
DX9 path: Smooth 62FPS (FPS cap).
DX10 path without AA: 40~60FPS but quite a bit of stutter.
DX10 path with AA: ~20FPS -> Screenshot

HD 4870 can't handle DX10/AA in this game. Maybe we're finally seeing the shortcoming of 512MB on HD 4870. For a change G92 shows its strength here. Stutters occur often enough, but those might rather have to do with SLI. While the game play experience under DX10 without AA feels better on HD 4870, HD 4870 cannot handle the in-game DX10 with AA @2560x1600. I couldn't even navigate the menu screen. I haven't tried forcing AA via control panels for either setup but I've heard that it didn't work under DX10, so I didn't bother. I tried lower resolution with great success (thanks, n7) and @1920x1200 or below HD 4870 once again provides excellent experiences (50FPS+ without stutters).

Apparently the combination of DX10+AA @2560x1600 is too much for 512MB of frame buffer in this game, and it's one of those moments that could make you wish for a bigger/monolithic GPU with lots of RAM. Still this was somewhat strange because the next game is also extremely heavy on both GPU memory as well as system memory.

Company of Heroes | Max in-game setting as below

In-game settings screenshot

HD 4870
DX9 path: Very smooth 60 FPS (FPS cap).
DX10 path without AA: 20~50FPS depends on the maps and units.
DX10 path with 4AA: 20~40FPS depends on the maps and units. (I am not sure if AA worked, however. Couldn't tell it from screenshots)

8800 GT SLI
DX9 path: Very smooth 60 FPS (FPS cap).
DX10 path with or without AA: Unplayble.

Interestingly, this game behave opposite to Gears of War when it comes to DX10/AA. (although HD 4870 could handle GoW admirably in DX10 without AA) Geforce shows its usual collapse spectacularly. I've experienced this time and again: Up to a point, G92 performs great. But once it goes past the thresholds (either resolution, AA, or specific game scenes), things crawl all of a sudden. There is no middle ground or gradual performance loss. It just drops from great to unusable.

Anyway, here are some beautiful DX10 shots that Company of Heroes is know for.

Soft Lighting
Water Ripples & Reflections
Realistic Weather Effects


My quick take is this is a very impressive card. This card handles 2560x1600 resolution better than any other card/cards I've owned. This is not a small feat, especially for a card with 512MB of RAM. It beats cards with same or more amount of VRAM (G80/G92/RV670) Very satisfying, to say the least, and I can't help but wonder what it could have been if AMD had aimed higher (and were successful, of course). I would also love to hear from GTX 280 / 30" LCD owners.

Gaming isn't the only thing that this card excels at. We know the future computing will only be more graphics heavy. High definition contents will become a norm at some point, and next iteration of Windows is reportedly even more visual (and interactive). I've been interested in the use of GPU other than gaming on desktop for some time now. Adobe Acrobat is one of the applications that embraced GPU acceleration already. Here is an example (3D PDF sample). Change the background color to white (so that you can see) and try the different rendering modes. Make sure to have 3D GPU rendering enabled. (It is, by default) You will need Acrobat Reader 8.1.2 or above to view 3D contents.

RV770
G92

Both cards do great when it comes to shading, but for some reason G92 tanks whenever there are heavy polygons on-screen. (Granted I don't think this is an architectural defect but rather a driver issue)

I think 3D Acrobat will catch on pretty quickly in the future. Product catalogs, manuals, interactive portfolios, or even a simple Java games. It's easy to imagine, for example, a car brochure in 3D with interactive features. (let customers to change angles, colors, parts, accessories, etc.)

Now that I've lavishly praised HD 4870, I'll point out some of its negatives.

- Heat
- Heat
- More heat

- This thing is too hot. I appreciate that AMD designed the cooler such that it exhausts all hot air out the case, but 78C idle is NOT acceptable. HD 3850's 2D core speed was 300MHz, there is absolutely no reason this thing should run at 500MHz in 2D from end-users' point of view. There is a simple fan fix mod out there but RPM above 27~28% is definitely audible outside a case, and past 30% it becomes downright annoying. I'm keeping mine at 28% and while it keeps the card cool in idle, it goes up to 90C under load^$(#&@! Plus, the hot air pushed out of the card could almost heat up a small room. (and this is unavoidable unless a 3rd party measure is applied, I think) I happen to have my legs close to the back of my system, and I can feel my body temperature rising with this thing running.

- AMD should have fixed the angle dependent AF already. There are occasions where you can see 'moiré effect'. Obviously this depends on angles and surfaces, but it's quite visible when it shows.

- Despite inferior AF, performance hit from enabling AF seems higher than that of G80/G92 or sometimes even higher than enabling AA. WTF?

- HD playback acceleration only works for protected (i.e. DRM'ed) contents. WTF? It looks like only Blu-Ray movies from the discs are accelerated, meaning user-encoded H.264/VC1 or movie trailers, etc. are not accelerated. (it isn't like this with HD 3850, so hopefully it's a driver issue.)

- Why can't AMD incorporate profiles system in CCC? Is it that hard? I mean, it doesn't have to be pre-made profiles but at least give us something to work on.

So here I laid out my initial impression on HD 4870, which may sound overly positive. But I don't feel like I am exaggerating. It's performant (fast and SMOOOOTH), efficient (handles 2560x1600 quite well with only 512MB of RAM), versatile (both in 2D/3D), and affordable (considering.. ). It's without a question the best performing card I've owned to date. At least I can positively say that a single RV770 w/GDDR5 > G92 SLI. Now, the only question mark I have is vs. GT200, and hopefully someone will answer that.

 

[dennilfloss]

Great review, thanks! :thumbsup:

 

[mazeroth]

Absolute killer review! Thanks for doing this!

 

[bryanW1995]

thanks for the hard work lopri!

 

[n7]

Fantastic.

This is why "official reviews" don't get even close to explaining the whole story.


I am concerned about the Gears of War AA thing though...well more, the UE3 AA thing.
You have any other UE3 games you can test?

If not, what about GoW AA enabled @ say 1600x1200 or a lower resolution?
Does it still choke to death?

 

[Elcs]

Thank you for justifying my purchase.

I could have picked up a new SLI board and a second 8800GT for approximately the same price as an HD4870 but I went for the HD4870.

Anandtech's review sparked my interest in SLI'ing two 8800GT's but when reading about macro/microstuttering and then thinking that some games dont support it or scale with it, one more powerful card sounded the better option. This thread made me smile.

 

[Kuzi]

Great review lopri. I have a question, at the resolution you play at, does it make a difference in image quality if you turn on AA or not?

I would have thought it's not really needed at such a high res.

 

[chizow]

Well done. I especially like the comparison screenshots. Its posts like these from real-users that drive the point home about multi-GPU. One curious thing though was this bit:

I was so accustomed to various stutters (macro, micro, or what not) of G80/G92

What G80 part were you referencing here? I've read about some reports of jerkiness/stutter with G92 compared to G80 (due to less VRAM or bandwidth perhaps) and even saw some of it myself in the brief time I had a 8800GT, but was this done at 2560? Was it even occurring at high FPS or was it a low-FPS only thing? I think part of the reason you weren't seeing similar stutter on the 4870 was because you were running at much higher FPS.

 

[n7]

Originally posted by: Kuzi
Great review lopri. I have a question, at the resolution you play at, does it make a difference in image quality if you turn on AA or not?

I would have thought it's not really needed at such a high res.

You thought wrong

Pixel pitch isn't that much lower than what it is on a 17" 1280x1024 or 20" 1680x1050 or even 24" 1920x1200, so yes, just like you see jaggies on those displays, you see jaggies @ 2560x1600.

Lower pixel pitch essentially hides aliasing, so if you had 2560x1600 on a 24", yeah, it'd be hard to see the aliasing.
But since pixel pitch isn't really that much lower with a 30", it's still nice to run AA.

Here Anandtech lists pixel pitch for common display sizes: http://www.anandtech.com/displays/showdoc.aspx?i=2961

 

[DrBombcrater]

Originally posted by: lopri
- HD playback acceleration only works for protected (i.e. DRM'ed) contents. WTF? It looks like only Blu-Ray movies from the discs are accelerated, meaning user-encoded H.264/VC1 or movie trailers, etc. are not accelerated. (it isn't like this with HD 3850, so hopefully it's a driver issue.)

Some kind of issue between the driver and your software, for sure. HD acceleration works just fine for unprotected files on my system (4870/hotfix driver) under PowerDVD and DVBViewer.

 

[hemmy]

I had the exact same experience with stuttering, only when I went from X1950 to 8800GT it got a lot better.

strange

 

[AzN]

4870 is so efficient in fact it does more FP16 blending fillrate than 280gtx while G92 and GT200 does better bilinear fillrate with less theoretical fillrate. Considering FP16 blending rate is where the cards start choking in a game these days that's where you see the 4870 even beating down a pair of 8800gt sli in those situations. Anything higher quality or more things on screen that's where the RV770 seems to do better.

 

[lopri]

Thank you for kind words. I will reply to the questions.

Originally posted by: DrBombcrater
Some kind of issue between the driver and your software, for sure. HD acceleration works just fine for unprotected files on my system (4870/hotfix driver) under PowerDVD and DVBViewer.

Can you try playing any of the movies from links below? PowerDVD/Blu-Ray combo works on my system as well.

WMV HD Gallery
Quicktime HD Gallery

My CPU usage goes up to ~20% when playing those clips. That's equivalent to ~80% on a single core. When AVIVO/PureVideo work the CPU usage never goes up beyond 5% by the movie itself. And that is the exactly what happens when I play Blu-Ray discs with HD 4870 (both H.264 and VC1).

Originally posted by: chizow

I was so accustomed to various stutters (macro, micro, or what not) of G80/G92

What G80 part were you referencing here? I've read about some reports of jerkiness/stutter with G92 compared to G80 (due to less VRAM or bandwidth perhaps) and even saw some of it myself in the brief time I had a 8800GT, but was this done at 2560? Was it even occurring at high FPS or was it a low-FPS only thing? I think part of the reason you weren't seeing similar stutter on the 4870 was because you were running at much higher FPS.

8800 GTX is a whole lot better than 8800 GT, but I still did experience occasional hiccups. Those usually happened when there were big texture swaps. I think (or thought) those hiccups are something unavoidable regardless of FPS.

Originally posted by: Kuzi
Great review lopri. I have a question, at the resolution you play at, does it make a difference in image quality if you turn on AA or not?

I would have thought it's not really needed at such a high res.

n7 kindly answered your question. Even x2AA, if nothing else, is appreciated.

Originally posted by: n7
I am concerned about the Gears of War AA thing though...well more, the UE3 AA thing.
You have any other UE3 games you can test?

If not, what about GoW AA enabled @ say 1600x1200 or a lower resolution?
Does it still choke to death?

I went back to GoW, and tried DX10/AA @1920x1200. Good news - FPS is practically stuck @62 while moving around. It looks like only when there are firearm explosions the FPS drops to 40's or 50's.

That encouraged me to try it again @2560x1600, but the result was the same (~10FPS). However, I have found something interesting. I am looking into it now, and will post it ASAP.

Since I started real gaming on this rig, the lack of profiles in CCC is becoming cumbersome. I've added this to the 'Negatives' list in the original post.

 

[natty1]

8800 gt sli was a terrible choice for 2560 res, unless of course you bought your monitor later on

 

[thilanliyan]

Great review. I think I've found my next card...unless the X2 is a significant step forwards.

 

[SlowSpyder]

It's nice to get some subjective opinions vs. just average frame rates. I've kind of thought that things have gotten too complicated to just say rig x gets 60 FPS and rig y gets 45 FPS... at least when comparing a system that uses multiple GPU's. I think there's more to it then just frame rates as you pointed out. Good info. I'm definitely going to go the single GPU route yet, I'm sure multiple GPU's will improve, but I'm not sure they're ready for mainstream for a lot of us yet. Based on what you said I can certainly say that for me a single 4870 will make me much happer then any multiple GPU set up. Luckily, or unluckily depending on how you look at it, my smallish 22" monitor really doesn't require more then one GPU. With AMD's new cards I can easily push 1680x1050 res at under $200 now.

 

[n7]

Hmm, so it would seem that it's running out of vRAM when running AA @ 2560x1600, but not at 1920x1200?

That make sense, or no?

 

[lopri]

Originally posted by: n7
Hmm, so it would seem that it's running out of vRAM when running AA @ 2560x1600, but not at 1920x1200?

That make sense, or no?

Yes, but I think there is more to the story. Wait a few.

 

[AzN]

Originally posted by: n7
Hmm, so it would seem that it's running out of vRAM when running AA @ 2560x1600, but not at 1920x1200?

That make sense, or no?

ATI does better with handling vram. It seems more efficient compared to Nvidia. It might have that same effect on SLI considering it only has 512mb of vram.

 

[lopri]

HD 4870 definitely runs out of memory in Gears of War @2560x1600 if both DX10 and in-game AA are enabled. Can't explain otherwise because the game is real smooth @1920x1200. So n7's observation is correct, IMO. But then it occurred to me - why isn't it happening to the same severity with G92?

That's when I started looking at the screenshots more carefully. I'll describe the procedure one by one.

First, the following in-game setting was used for both configurations.

Video Settings @2FPS :laugh: (And control panel = High Quality, 16AF for both systems)

Now here is a screenshot without AA enabled as a reference. (But still under DX10 path)

HD 4870 @56FPS

Now, in-game AA enabled under DX10 path.

8800 GT SLI @19FPS
HD 4870 @9FPS

G92's superior filtering quality is evident if you look at the pillars in close proximity. But I am not sure whether each cards' are equally taxed when it comes to AA. I picked a couple spots showing the difference of AA/AF.

EDIT: I've just realized that the screenshot was taken while 'Mipmap Detail Level' was at 'Quality, not 'High Quality' in CCC. Duh. Not sure how much details are lost in the shot above. Gah. That makes the whole comparison (Not performance, but IQ) murky. *$(*&@#$^ I am not going to delete all the screenshots but please disregard both AA and AF quality comparisons. Performance issue of DX10 and AA is still valid, however.

Green box and Red box

If you go back and compare the shots, green-boxed area clearly shows when it comes to AF G92 > RV770. For AA comparison, I picked a far-away wooden stick from each shot (red-boxed area) and zoomed them in.

G92 DX10 AA
RV770 DX10 AA

Remember that this AA is not forced by control panels, but selected through in-game menu. There are a few more spots I could see difference (usually in the distance) between the two AA's. I am no expert of AA so I cannot say which one is superior, but they're definitely different. And to my eyes, it looks as if RV770 is using more samples. I could be wrong or I might need better screenshots. (I tried to shoot the skull's eye) Experts' opinion is needed here. (BFG where are thou?)

Anyway, DX10 with AA in this game is a no-go for both setup. Apparently HD 4870 doesn't have enough memory for AA samples @2560x1600 in this game. Great looking game, nonetheless. I will update the first post accordingly.

 

[Kuzi]

Originally posted by: n7

Originally posted by: Kuzi
Great review lopri. I have a question, at the resolution you play at, does it make a difference in image quality if you turn on AA or not?

I would have thought it's not really needed at such a high res.

You thought wrong

Pixel pitch isn't that much lower than what it is on a 17" 1280x1024 or 20" 1680x1050 or even 24" 1920x1200, so yes, just like you see jaggies on those displays, you see jaggies @ 2560x1600.

Lower pixel pitch essentially hides aliasing, so if you had 2560x1600 on a 24", yeah, it'd be hard to see the aliasing.
But since pixel pitch isn't really that much lower with a 30", it's still nice to run AA.

Here Anandtech lists pixel pitch for common display sizes: http://www.anandtech.com/displays/showdoc.aspx?i=2961

Thank you for the explanation

 

[thilanliyan]

Originally posted by: lopri
G92 DX10 AA
RV770 DX10 AA

Remember that this AA is not forced by control panels, but selected through in-game menu. There are a few more spots I could see difference (usually in the distance) between the two AA's. I am no expert of AA so I cannot say which one is superior, but they're definitely different. And to my eyes, it looks as if RV770 is using more samples. I could be wrong or I might need better screenshots. (I tried to shoot the skull's eye) Experts' opinion is needed here. (BFG where are thou?)

The ATI sample looks more detailed but maybe the NVidia sample was was taken when that part of the screen was too bright maybe?

 

[NullSubroutine]

Originally posted by: lopri
HD 4870 definitely runs out of memory in Gears of War @2560x1600 if both DX10 and in-game AA are enabled. Can't explain otherwise because the game is real smooth @1920x1200. So n7's observation is correct, IMO. But then it occurred to me - why isn't it happening to the same severity with G92?

That's when I started looking at the screenshots more carefully. I'll describe the procedure one by one.

First, the following in-game setting was used for both configurations.

Video Settings @2FPS :laugh: (And control panel = High Quality, 16AF for both systems)

Now here is a screenshot without AA enabled as a reference. (But still under DX10 path)

HD 4870 @56FPS

Now, in-game AA enabled under DX10 path.

8800 GT SLI @19FPS
HD 4870 @9FPS

G92's superior filtering quality is evident if you look at the pillars in close proximity. But I am not sure whether each cards' are equally taxed when it comes to AA. I picked a couple spots showing the difference of AA/AF.

EDIT: I've just realized that the screenshot was taken while 'Mipmap Detail Level' was at 'Quality, not 'High Quality' in CCC. Duh. Not sure how much details are lost in the shot above. Gah. That makes the whole comparison (Not performance, but IQ) murky. *$(*&@#$^ I am not going to delete all the screenshots but please disregard both AA and AF quality comparisons. Performance issue of DX10 and AA is still valid, however.

Green box and Red box

If you go back and compare the shots, green-boxed area clearly shows when it comes to AF G92 > RV770. For AA comparison, I picked a far-away wooden stick from each shot (red-boxed area) and zoomed them in.

G92 DX10 AA
RV770 DX10 AA

Remember that this AA is not forced by control panels, but selected through in-game menu. There are a few more spots I could see difference (usually in the distance) between the two AA's. I am no expert of AA so I cannot say which one is superior, but they're definitely different. And to my eyes, it looks as if RV770 is using more samples. I could be wrong or I might need better screenshots. (I tried to shoot the skull's eye) Experts' opinion is needed here. (BFG where are thou?)

Anyway, DX10 with AA in this game is a no-go for both setup. Apparently HD 4870 doesn't have enough memory for AA samples @2560x1600 in this game. Great looking game, nonetheless. I will update the first post accordingly.

Are you using Temporal AA or Multisampling or Supersampling? Both take a good hit in performance but make AA much better. Some games MS is faster others SS is.

 

[BenSkywalker]

Remember that this AA is not forced by control panels, but selected through in-game menu. There are a few more spots I could see difference (usually in the distance) between the two AA's. I am no expert of AA so I cannot say which one is superior, but they're definitely different. And to my eyes, it looks as if RV770 is using more samples.

Neither of them are applying proper AA in your provided shots by the looks of it. You notice it differently depending on the particular board, but both of them are providing different odd errors. The G92 is exhibiting some flat out wrong colored pixels and you have some improper gamma compensation by the looks of it(may be something else, but that is what it looks like). The RV770 seems to completely miss pixels every once in a while and have the blended value misplaced by one pixel in a particular direction. Hmmm, the more I look at it, the more that could also be a bit of hosed gamma compensation. Either way, we need more then this screenshot to figure out what is going on, we need near vertical, near horizontal and near 45 degrees to give a proper assesment

 

[zod96]

Would the 4870 benefit me at 1680x1050 with games like Stalker and Oblivion? I'd like to play those with max AA and AF I currently have a 8800GT myself.