Hardware question on HD 5770

[Wogdog]

I just got a XFX 5770 a week ago, so far haven't had any problems in Win 7 Home 64bit. Got it for the lower power consumption and smaller form factor than the GTX260 or 4870. I have seen a lot of people are really disappointed about the 128bit memory bandwidth and immature drivers. Also see a lot of posts that will change once Fermi comes out and the ATI/AMD will increase performance up with drivers, hope it comes before then. My question is, could the HD 5770 performance be held back by the firmware(BIOS) and once ATI needs/wants to, can they simply release a new firmware flash and new drivers to increase the memory bandwidth of the card? Is the card intentionally handicapped by ATI, as some people think, so they can still sell the 4890's? The 4870 is already EOL'd. Just an asking cause I am not sure if the PCB has the traces for better memory bandwidth addressing or not, but have seen better firmware(BIOS) improve other products.

 

[nemesismk2]

My question is, could the HD 5770 performance be held back by the firmware(BIOS) and once ATI needs/wants to, can they simply release a new firmware flash and new drivers to increase the memory bandwidth of the card? Is the card intentionally handicapped by ATI, as some people think, so they can still sell the 4890's? The 4870 is already EOL'd. Just an asking cause I am not sure if the PCB has the traces for better memory bandwidth addressing or not, but have seen better firmware(BIOS) improve other products.

I think the 5770 is just a budget video card and I don't think it's possible to turn it into a 5870 just with a firmware flash. Could firmware turn 128Bit Memory Interface into 256Bit? I don't think so but I could be wrong. lol

I have seen Firmware improve video cards as well but that it just enabling some disabled pixel shaders.

 

[jvroig]

BFG10K seems to think ATi is intentionally holding back performance pending Fermi's arrival , let's see what he has to say about this. I also asked about the viability of that hypothesis on another thread, and he's still due to reply (or perhaps he's already explained that in yet another thread). He'll be sure to see this one and probably set us all straight.

 

[biostud]

the 128-bit bus is still DDR5 so that equals the 4850 256-bit DDR3 bus, so for a midrange card I would say it's to be expected. I seriously doubt that they hold performance back, but newer drivers helps fix bugs and sometimes also improve speed for some games.

 

[BFG10K]

Yes, I personally believe the drivers are currently holding back performance with the entire 5xxx series, whether it’s intentional or not. The drivers could simply be immature, or it could be ATi not wanting to play their hand too early, or because they want to get rid of old stock.

A firmware flash can’t enable features that aren’t there (e.g. widening the memory), but a driver update that addresses units not being utilized properly is certainly possible. For example, even though all of the shaders are reported by utilities, it’s possible the driver isn’t actually using all of them.

Also based on my testing, it appears the 5770 is currently primarily limited by its core, not by its memory.

BFG10K seems to think ATi is intentionally holding back performance pending Fermi's arrival , let's see what he has to say about this. I also asked about the viability of that hypothesis on another thread, and he's still due to reply (or perhaps he's already explained that in yet another thread). He'll be sure to see this one and probably set us all straight.

Hmm, I don’t recall seeing that question, so I must’ve missed it. Can you please link to it for me? Thanks.

 

[jvroig]

Hmm, I don’t recall seeing that question, so I must’ve missed it. Can you please link to it for me? Thanks.

First of all, thanks for answering it here.

Second, I know this will make me look crazy (but I'm not), but even before you dropped by here, almost as soon as I made my post here that you answered, I searched for that post I said I made that asked you about it. After an hour of searching, I came up empty. I am bewildered myself Perhaps I meant to post it last night, but my connection got interrupted and I forgot to try again and now merely assumed I did. I am embarrassed.

I distinctly remember that in that supposed thread where I posted that question to you, you posted something where you mentioned you are excited for Fermi to come out and get an updated driver for your 5770 so you can re-test it. In reply, I posted that I have seen you mention that before (I mean re-testing the 5770 after AMD "uncripples it") to ask you how you can believe that despite an AMD engineer already answered the "why only 33% improvement?" in a blog by saying that given the "fact" that games rely on so many factors, with graphics cards being only less than 30% of the equation, it's already amazing they were able to get a 33% performance increase. Not that I believe his line of reasoning, but that seemed to point that as far as AMD is concerned, they aren't doing any holding back.

There, I'm sorry about "losing" that post, wherever it may be or not be (I only have a few posts, and I practically scoured each one of them to try to find that post I referenced). Anyway, I practically reproduced my questions in the last paragraph, if you could, then I'd appreciate a response.

Thanks.

 

[BFG10K]

jvroig, I found your post. It's in the Fermi thread:

http://forums.anandtech.com/showpost.php?p=28881850&postcount=13



Do you happen to have a link to that blog? That sounds more like PR than engineering to me.

There’s absolutely no way I can believe the graphics card amounts to less than 30% of the equation, especially not when we see the 4870X2 often pulling higher scores than the 5870.

Also my GTX285 bottlenecks me 100% in the vast majority of situations I game on it, even with a “slow” E6850 at stock. In fact, I bet I could underlock my CPU to 2.4 GHz (20%) and not see a significant performance drop at my gaming settings.

Here are some CPU scaling tests I did with a E6600 and E6850 on a 8800 Ultra:

http://episteme.arstechnica.com/eve...65/m/492007078831?r=492007078831#492007078831

The scores didn’t budge with the settings I used at the time (see the right-most results), except in Painkiller which wasn’t really indicate of actual performance (I now use a much better demo in that game). Now I use higher settings on my GTX285, with more demanding games.

If you configure your games to always run at the highest playable settings, the GPU will be your primary bottleneck in the majority of situations, by far. With the specs of the 5xxx series, it simply isn’t performing where it should be relative to previous cards.

 

[jvroig]

BFG10K: Thank you. I am not crazy after all For the record, I was sure I'm perfectly sane. Now I am embarassed that you were able to find my own post better than myself.

As for the blog, I didn't remember the link, but googling "CarrelK responds to only 33%" immediately brought it back, first hit:

http://www.siliconmadness.com/2009/09/amd-engineer-comments-on-radeon-hd-5870.html

Carrel Kilbrew, of course, is the wizard behind RV770. Definitely not just marketing/PR guy, a real engineer. I am bewildered myself why he would say such a thing, except perhaps to defend his current baby?

Please understand that I am no GPU expert. My only possible similarity to you would most likely be the hope that it is immature drivers (intentional or not) that cause the not-so-spectacular performance of the 5xxx series.

Now that you've read the link, what are your thoughts about it, and in particular, of Kilbrew's comment, especially since it is coming from one such as him, a lead engineer and not a PR guy? Does your opinion regarding drivers being held back on purpose not change?

Thanks.

 

[jvroig]

I just read the link again myself and saw that it links to AT

Of course, his comment has since moved to a different page, and is now found here: http://anandtech.com/video/showdoc.aspx?i=3643&p=11&cp=22#comments


EDIT: I'm surprised myself that CarrelK's comments in the AT article itself was not responded to by anyone. I blame SiliconDoc for making a complete mess of the comments; it was during those days when he practically swamped every article about video cards.

 

[Wogdog]

Thanks for the answer BFG. . I'm satisfied with the card for what it is right now, the lower power reqs and features. Have to hold on and wait for the driver updates. If in the future I need better performance, I will get another and crossfire. I just think it's odd that there isn't a 5830 or something between the 5770 and 5850.

 

[v8envy]

Plenty of people have pointed out strange performance anomalies with the 5 series.

The 5770 is a shrunk 4890 with much less memory bandwidth. Predictably, it performs a bit under a 4870. I have no idea how it's possible to draw the conclusion that it's GPU not memory limited from that. It's not reasonable for less hardware than is present on the 4890 to miraculously perform better unless the 4890's drivers are suboptimal. If there is some performance to be unlocked from the 5770 it's unlikely to be awe-inspiring across the board.

The 5870 is twice the hardware of a 4890. And yet, is only about 35% faster on average. Increasing memory bandwidth or CPU power independently seems to indicate neither alone is the bottleneck. Software issues are a possible explanation.

I expect future driver revisions to make better use of the 58xx hardware, but the 57xx is already there. Don't be surprised if in a year's time the 58xx outperforms the 295 by at least 10% while the 5770 is pretty much where it is today, on average.

 

[cbn]

deleted.

 

[RussianSensation]

Also based on my testing, it appears the 5770 is currently primarily limited by its core, not by its memory.

It's interesting that Xbitlabs concludes that it's the memory which is holding back 5770 series (since its raw compute power is more or less equal 4890s):

Pixel Fillrate = same for both cards at 13,600 MPixels/sec
Texture Fillrate = same for both cards at 34,000 MTexels/sec
GPU Core clock = same for both cards at 850 mhz
ROPs = same at 16 for both cards
Texture units = same at 40 for both cards

The only logical conclusion is that it's the memory bandwidth which is holding back 5770. This makes sense since 4870 is about 20% faster (750 vs. 625mhz GPU) in computation compared to 4850, but is often 30% or more faster in benchmarks at higher resolutions with AA, which are memory bandwidth starved:
http://www.techreport.com/articles.x/16681/9

Fallout 3 - 2560x1600 4AA/16AF
4870 1GB = 61.1 avg / 46.0 min (+35% / +35%)
4850 1GB = 45.2 avg / 34.0 min

"The Radeon HD 5700 models both boast good overclockability without any special tools or methods. In some applications overclocking makes the senior Radeon HD 5700 series card as fast as the Radeon HD 4890 whereas the junior model, when overclocked, easily overtakes its senior cousin. The average performance growth of the Radeon HD 5770 overclocked to 940MHz GPU and 1445 (5870) MHz memory frequencies is about 12%, reaching 14% in some applications.

For the Radeon HD 5750 overclocked to 870MHz GPU and 1430 (5720) MHz memory the average and maximum performance growth are 14% and 16%, respectively. This is a real gift for overclockers, especially as PC enthusiasts will surely reach better results using more advanced coolers. On the other hand, we have to note that the main bottleneck of the new series is the 128-bit memory bus which is going to limit the effect from overclocking at high resolutions."

http://www.xbitlabs.com/articles/video/display/radeon-hd5770-hd5750_17.html#sect0

 

[AzN]

BFG10K seems to think ATi is intentionally holding back performance pending Fermi's arrival , let's see what he has to say about this. I also asked about the viability of that hypothesis on another thread, and he's still due to reply (or perhaps he's already explained that in yet another thread). He'll be sure to see this one and probably set us all straight.

I doubt it. It's performing by the specs. If 5870 is "intentionally" being held back so is RV770 and the variety.

 

[AzN]

Also based on my testing, it appears the 5770 is currently primarily limited by its core, not by its memory.

You can't just point to average frame rates and say it's being limited by core. Bandwidth up's minimum frame rates and core peak maximum frame rates. When core peaks maximum frame rates average frame rates improve much more than bandwidth.

 

[BFG10K]

Now that you've read the link, what are your thoughts about it, and in particular, of Kilbrew's comment, especially since it is coming from one such as him, a lead engineer and not a PR guy? Does your opinion regarding drivers being held back on purpose not change?

My opinion doesn’t change at all. He may be an engineer, but I still think it’s a PR response he’s given. It could be as you say that he’s defending his baby, or he’s been instructed not to blow the cover of what ATi has planned when Fermi hits. If he keeps quiet now he can always state later that ATi found room for driver improvements.

Let’s look at the facts: if the platform/game is the limitation, how is it that four other graphics configurations are scoring higher than a single 5870 in that graph? In particular, how is it that adding a second 5870 increases performance by 40% over a single 5870? That’s a pretty big increase.

If the platform/game was the limitation, then all performance past a single 5870 would be flat-lining on that graph, yet it isn’t.

We also know it isn’t memory bandwidth, because it’s been proven that the core affects performance more than memory by a factor of more than 2 to 1: http://www.firingsquad.com/hardware/ati_radeon_5870_overclocking/

So with double the core processing over the 4890, that leaves the drivers as the most likely candidate. Unless there’s some kind of deep-rooted bottleneck in the pipeline somewhere.

To add more evidence to the drivers, I can find instances where my 5770 absolutely whips my GTX285, yet has less than half the memory bandwidth.

 

[BFG10K]

You can't just point to average frame rates and say it's being limited by core.

Why not? It was good enough when I put up the 8800 Ultra results. Have you changed your mind since then?

Bandwidth up's minimum frame rates and core peak maximum frame rates. When core peaks maximum frame rates average frame rates improve much more than bandwidth.

You absolutely cannot make such a sweeping generalization. That fact is, anything that’s a significant bottleneck has the potential to affect minimum/average/maximum in an equal fashion. That and higher minimums can affect the average just as much as higher maximums can.

I tested nine games and in only one did memory exhibit significantly more movement from memory (Bioshock). The rest were either close, or the core had a massively larger influence. In particular, the core had more than twice the impact of memory in CoD 5, Wolfenstein, and CoD 4.

You can find the link to my 5770 bottlenecking tests in this thread: http://forums.anandtech.com/showthread.php?t=334977

Overall the 5770 was 11.20% (core) vs 8.46% (mem), so it’s quite a balanced part with the core having more influence overall. I’m not saying it’s not limited by bandwidth because it clearly is, given it’s a balanced part.

What I am saying is that memory isn’t the primary limitation as common belief suggests. This is unlike the 8800 Ultra where the core affected the results by more than twice the amount compared to the other clocks.

This coupled with the points I made about the 5870 above, and the fact that my 5770 can whip my GTX285 in certain bandwidth heavy situations (e.g. 2560x1600 with 8xAA) despite having less than half of it, leads me to suspect drivers are holding back the performance of the core.

P.S. I’ve asked the administration to correct your post count.

 

[RussianSensation]

BFG, valid points. I think 5870 is not memory bandwidth limited since it has plenty of it. I am still not convinced that you can infer the same about 5770 from the firingsquad review you linked.

Although I am not sure what modern game can be played on 5770 at 2560x1600 8AA? Sure there are some improvements in RV870 core over 770 but this is not an architectural change. Also, 4890 sometimes gives 5770 a beating by such a large, that it's difficult to suggest a 30%-40% performance deficit is driver related since it's more or less the same architecture:
http://www.tomshardware.com/reviews/radeon-hd-5770,2446-11.html

If AMD can get another 20%+ from 5770 across the board, that would be amazing, and they would do so ASAP since they surely would want to sell a heck of lot more 40nm 128-bit 5770s at $150-$160 than 55nm 256-bit 4890s!

Further, since 5870 hardly doubles 4890's performance, despite double the theoretical performance increase (other than memory bandwidth), my instinct tells me there is a bottleneck somewhere else in the architecture itself. After all, this is more or less the same thing since R600. I hope R970 is a brand new design for their sake.

 

[AzN]

Why not? It was good enough when I put up the 8800 Ultra results. Have you changed your mind since then?


You absolutely cannot make such a sweeping generalization. That fact is, anything that’s a significant bottleneck has the potential to affect minimum/average/maximum in an equal fashion. That and higher minimums can affect the average just as much as higher maximums can.

I tested nine games and in only one did memory exhibit significantly more movement from memory (Bioshock). The rest were either close, or the core had a massively larger influence. In particular, the core had more than twice the impact of memory in CoD 5, Wolfenstein, and CoD 4.

You can find the link to my 5770 bottlenecking tests in this thread: http://forums.anandtech.com/showthread.php?t=334977

Overall the 5770 was 11.20% (core) vs 8.46% (mem), so it’s quite a balanced part with the core having more influence overall. I’m not saying it’s not limited by bandwidth because it clearly is, given it’s a balanced part.

What I am saying is that memory isn’t the primary limitation as common belief suggests. This is unlike the 8800 Ultra where the core affected the results by more than twice the amount compared to the other clocks.

This coupled with the points I made about the 5870 above, and the fact that my 5770 can whip my GTX285 in certain bandwidth heavy situations (e.g. 2560x1600 with 8xAA) despite having less than half of it, leads me to suspect drivers are holding back the performance of the core.

P.S. I’ve asked the administration to correct your post count.

Thanks for correcting my post count.

Your 8800ultra benchmarks were years ago and still is core starved for the amount of bandwidth it has.

Your Benchmarks were expected. Little more core and a little more bandwidth. Nothing much changed since 4850.

I have a slight different perspective as to say a card is bandwidth or core starved. Instead of looking at "only" average frame rates which usually sway by core because core peaks higher max frame rates to get more average frame rates. The minimum frame rates matter much as average frame rates if not more to a gamer which is swayed by bandwidth. Factor that in to the equation there you'll see the bottleneck.

 

[solofly]

I'm very satisfied with the performance of my XFX HD5770... (it replaced a 9800gt)

 

[davidb72]

This thread is interesting to me because I chose to purchase two 5770's to put in my P55 build. I would have preferred to buy one 5850 now and one later, but they are a little hard to find right now.

Since the P55 is limited to 8x on the two PCI-E slots when using Crossfire, I somehow justified the 5770's to myself because of their limited bandwith. I figured they would be the perfect match for the architecture of the platform.

Am I thinking correctly or am I grasping at straws trying to justify purchasing two slightly less powerful cards than what I would have liked to have purchased?

 

[toyota]

This thread is interesting to me because I chose to purchase two 5770's to put in my P55 build. I would have preferred to buy one 5850 now and one later, but they are a little hard to find right now.

Since the P55 is limited to 8x on the two PCI-E slots when using Crossfire, I somehow justified the 5770's to myself because of their limited bandwith. I figured they would be the perfect match for the architecture of the platform.

Am I thinking correctly or am I grasping at straws trying to justify purchasing two slightly less powerful cards than what I would have liked to have purchased?

well overall its faster than a 5850 plus a 5850 is hard to find so dont worry about it.

BTW you really need to redo your sig because taking up 22 lines is pretty silly.

 

[hans030390]

well overall its faster than a 5850 plus a 5850 is hard to find so dont worry about it.

BTW you really need to redo your sig because taking up 22 lines is pretty silly.

Counting that his sig takes up 22 lines is silly.

 

[toyota]

Counting that his sig takes up 22 lines is silly.

and replying to a post just to tell me that is stupid. the only reason I said anything to him is because there is no reason to take up that much space for a damn sig. plus if I am not mistaken most forums have a 10 line limit for sig in the first place.

 

[davidb72]

well overall its faster than a 5850 plus a 5850 is hard to find so dont worry about it.

BTW you really need to redo your sig because taking up 22 lines is pretty silly.

That's cool, when I set that sig up in the first place I thought it was a little obnoxious...

Sorry for sidetracking the thread...