HD 5870 Discussion thread.

[Bo_Fox]

i agree with Bo_Fox. I mean the GTX 260 is hell of alot bigger than the GTS 250 with much better specs, yet the GTS 250 is only 6% on avg slower and even less at higher resolutions. I personally agree that had nvidia gave the GTS 250/9800GTX a 384-bit bus with 24 ROP's it probly would have closed that gap with less shaders. But eh we all can only think of what would have made sense.

Thank you, you're the man!

Thanks for understanding.. if Nvidia read this, Nvidia would definitely feel this! Especially Jen Hsung (I forgot how to spell his last name, but was it right?)


Nvidia could've at least had a G92, G90, and a GT200. That is, a G90 on 65nm and then a G90 and GT200 on 55nm. That way, Nvidia would've maintained a lead on a 4870 with GTS 250 with 384 bits and 24 ROP's on 65nm process, then continued the lead over a 4890 with GTX 285 on 55nm. HalfAHertz was right about there being additional manufacting costs for more variations and less volume per design, yet Nvidia would have yielded much better profits (yes, the research/development costs would've been a bit higher with less "returns" from mass-production at TSMC for a specific design, but it still would have paid off). But Nvidia just wanted to go with a monster chip from the start, at whatever cost, after the "supreme" success of monster G80 chip.

 

[HalfAHertz]

Thank you, you're the man!

Thanks for understanding.. if Nvidia read this, Nvidia would definitely feel this! Especially Jen Hsung (I forgot how to spell his last name, but was it right?)


Nvidia could've at least had a G92, G90, and a GT200. That is, a G90 on 65nm and then a GT200 on 55nm. That way, Nvidia would've maintained a lead on a 4870 with GTS 250 with 384 bits and 24 ROP's on 65nm process, then continued the lead over a 4890 with GTX 285 on 55nm. HalfAHertz was right about there being additional manufacting costs for more variations and less volume per design, yet Nvidia would have yielded much better profits (yes, the research/development costs would've been a bit higher with less "returns" from mass-production at TSMC for a specific design, but it would still have paid off). But Nvidia just wanted to go with a monster chip from the start, at whatever cost.

You have to look at the big picture here. Nvidia want to be something more than just a graphics vendor. They ant to become one of the main players in High Performance Computing. The G80/90 started mainly as a graphics core, having just the basic components for computing. It was the equivalent of Nvidias baby steps. The 200 series was a 50/50 thing, it was strating to more and more look like the real thing.
The 300 series will be the whole enchelada: Functionaliries like ECC corection, programable L1 cache, huge shared L2 cache, C compatibility, just name it - they have it... The huge size of the die is just the negative side-effect from those efforts.

 

[Bo_Fox]

Look, I was supportive of you, when I mentioned your name in my last post.

Oh man, oh yes, the GT300 will be the ultimate enchilada!!! Are you Mexican, huh? If so, oh yeah, the best Enchilada ever!!!

Nvidia should have also made a 384-bit/24 ROP/768MB version of G92 on both 65nm and then 55nm, along with a GTX 285/275 on 55nm (a bit less "plentiful" supplies, of course).

The 65nm version would've at least been head-on with a 512MB version of 4870. The 55nm could've as well been head-on with a 1GB version of 4870, if not better (assuming that the core clock could be pushed upwards of 800MHz and shader clock approaching 2000MHz within the 210W power envelope).

Ahhhhhhh, remember the X800XT / 6800 Ultra days? During the 1-year period between summer 2004 and summer 2005, both of those cards were selling for an average of $500 a pop. Both companies made a killing on those cards, especially ATI with their low-power X800XT chips. Both companies had equal competition, so both were able to compete "well" financially-wise. If Nvidia did a 384-bit/24 ROP/768MB version of G92, Nvidia could've sold it for $300 along with ATI selling their 4870's for $300 (and made a killing on it while ATI also made a killing on theirs), for at least 6 months until Nvidia releases an improved 55nm version. And Nvidia could have continued to sell the 55nm version for $~250 for another say, 8 months (while ATI starts to sell the 1GB version of 4870 for $250 for 8 months also. Also, Nvidia could've sold their GTX 285 for $400 for basically a year, and a GTX 275 for $330 also. (Note, amazingly, those are the selling prices right now, after they've been boosted a lot in the past month!)

Damn, do you guys hate me for sounding so supportive of the profits of those companies?!???!? Ahhh, maybe I do hate myself for saying that, being supportive of the profits of those companies at the expense of us who got those GTX 280 monsters quite cheaply?

 

[HalfAHertz]

The prices were so high back then due to price fixing. That's a lot of could haves and would haves in your last post. Things may have happened like this, or they may not but as you said it yourself, if things were different, we may not have had great cards like the GTX 285 and the 4890
Competition is trully a marvelous thing

Still I think we should stop taling about Nvidia in an Ati tread

 

[wolf]

...Nvidia should have...

Look I enjoy reading all your posts dude, but this is entirely speculative, and yeah doesn't really belong in this thread, I'll gladly participate in a thread about possible alternate G90/G92 configurations if you make one.

Hey do any applications FULLY support 5870/50 bios modifications, like clock speeds, clock states, fan ramping, voltages at given states etc? I know RBE has support, but how far does that extend?

we may be able to make a "5890" bios of our own or at least 5870 OC bios that anyone can use by taking values that work on say 95%+ of cards.

 

[Bo_Fox]

The prices were so high back then due to price fixing. That's a lot of could haves and would haves in your last post. Things may have happened like this, or they may not but as you said it yourself, if things were different, we may not have had great cards like the GTX 285 and the 4890
Competition is trully a marvelous thing

Still I think we should stop taling about Nvidia in an Ati tread

All right, cool.

However, I think both Nvidia and ATI would still have done a 4890 and GTX 285 (maybe a couple months later) anyways, and everything would've been slightly more expensive.

You're right that price fixing ain't good, baby! And I'm a hypocrite for being supportive of corporate profit at the expense of the consumer!

Look I enjoy reading all your posts dude, but this is entirely speculative, and yeah doesn't really belong in this thread, I'll gladly participate in a thread about possible alternate G90/G92 configurations if you make one.

Hey do any applications FULLY support 5870/50 bios modifications, like clock speeds, clock states, fan ramping, voltages at given states etc? I know RBE has support, but how far does that extend?

we may be able to make a "5890" bios of our own or at least 5870 OC bios that anyone can use by taking values that work on say 95%+ of cards.

Thanks, I appreciate your compliment.

Somebody answer Wolf's question above? I dunno..

 

[Bo_Fox]

Anandtech just did a rather specific article on 5970 overclocking. It appears that some of the VRM chips are severely overheated when overclocked, and that the VRM temperatures are not connected to the fan speed.

http://www.anandtech.com/weblog/showpost.aspx?i=657

Hope this helps to answer your question, Rabid Koala... that's all I know for now (if I had a 5870, maybe I'd know a bit more about this BIOS tweaking stuff that I did a while ago with 3870/4850's).

At least there's a new version of ATI Tray Tools that supports the R800 series!!! It also allows for tweaking of LOD which is useful to sharpen the blurry SSAA modes. It's a lot like Rivatuner, but with a few things that RT does not have.

Latest version of ATT here: http://www.softpedia.com/get/Tweak/Video-Tweak/ATI-Tray-Tools.shtml

 

[yogurt_21]

I need a "I read through this whole thread and all I got was this lousy t-shirt" shirt

seriously 17 pages of bandwidth talk when theres a simple law of design double specs does not = double performance. not to mention the fact that we're looking at paper specs upon release and have not had a chance to take a look at an in depth design and the actual coding of the card. Then we have to disect the the drivers and ackowledge they're release drivers (which most seem to be ignoring).

and maybe we'll be somewhere.

but based on what I've seen out of both ati and nvidia drivers improve performance over time it may not always be huge but it does happen. Of course there's the argument that the 5XXX series is only a redesign + dx11 and so the drivers should already be mature for dx9 and dx10 but I still hold back on that.

So comparing mature drivers to immature, stating double paper specs (at least those we are given) should = double performance is not practical. I see the 5870 as a strong card that is overpriced due to an earlier release than nvidia. The card performs very well and is the fastest single gpu card out there.

And no it's not memory bottlenecked and stating that something is architecture bottlenecked is erronious. I mean hell we've had 16 rop cards coming out of ati for many years and each generation improved. Obviously each prior generation wasn't as good as the latter at taking advantage of each rop. But that's the whole nature of the game. If the 5870 didn't have any bottlenecks we wouldn't need a new card until dx12 came out and even then only after dx12 mainstream games came out. The name of the game is push out a card, improve on it for several generations so that you buy a new one each time, then change architecture and do the same thing.

is the performance double? no. should it be? not really. new cards only need to be fast enough to entice without being too fast to be overcome with future gens.

 

[Bo_Fox]

^^ yogurt21, yeah, after 17 pages of whether we could use more bandwidth or not, (and you still cannot figure it out yet) then it's pretty pointless by now! All you're getting is a sweat-stained stinking t-shirt off my back! Hehe!

As more demanding games come out, the performance difference will eventually become more than 100% greater simply due to the fact that a 4890 can do Crysis 3 Extreme at only 4 fps and a 5870 can do it at 16 fps.

Happy now?

60% greater performance over a 4890 is still great! Oh yeah, almost exactly like the GTX 280 over 8800GTX. You say it's overpriced right now, but that we should be happy. Also, you say that we should not expect double the performance even if the specs are double.

Well, the 4870 had 2x the performance of a 3870 (yeah, it had over 2x the shaders and TMU's, but it still had the same ROP's and ~60% greater bandwidth). The amazing thing is that the chip had only like 50% more transistors with the same ROP's and lower clock!

A 7800GTX had nearly 100% the performance of a 6800 Ultra (with 2x the processing power), yet it still had the same ROP's.

16 ROP's does not always mean just 16 ROP's.

As the core clock increases, the ROP's become faster. Sometimes, it's slightly bottlenecking (like as if the term is a 4-letter word), sometimes it's not a bottlenecking factor like with a 3870/2900XT that was already limited by the shaders.

I just find the statistics (specifications) interesting. It's like following baseball, which is a sport all about the stats. Some of the baseball fans are completely obsessed about the stats like as if nothing else exists.

Hey dude, the T-shirt motto of this thread is that the memory bandwidth is not the primary indicator of performance, but rather that it complements the main muscle (GPU) that does the actual work. If the penalty of reduced memory bandwidth is almost as bad as the penalty of reduced core clock, then we know that it would definitely help to have some more bandwidth.

Here's an article that tries to focus on a potential bottleneck of 5870/5770. Please do not forget to look at the numbers, and see how the penalty of 20% reduced bandwidth is actually almost equal to 20% reduced core clock in 10+ games tested:
http://alienbabeltech.com/main/?p=12474 BFG10K is an excellent hardware/gaming enthusiast who has been participating on the Nvidia forums for years, and is now doing a great job writing articles.

By the way, here's a groundbreaking article that he did on the new circular AF image quality: http://alienbabeltech.com/main/?p=12648&page=2

 

[Bo_Fox]

HalfaHertz, hope you didn't take anything personally. When I asked you if you were a Mexican, I was only shooting out goofy questions.. if you are, then that's cool! I love enchiladas and think Mexicans are cool.

 

[Benetanegia]

By the way, here's a groundbreaking article that he did on the new circular AF image quality: http://alienbabeltech.com/main/?p=12648&page=2

Yeah I remember I read that the last week. TBH that's far more below expectations than the performance. The Image Quality is horrible!!

 

[Steevo]

http://alienbabeltech.com/main/?p=12474&page=3

Ohh no.

Someone did the same research and concluded that the 5XXX series isn't bandwidth limited too.

 

[Bo_Fox]

Ohh yes, he did, after reading my comment at the bottom of that page you just linked to. Read it and enjoy!

I'll go ahead and quote it:

I’d concur that raising memory bandwidth could boost performance almost as much as raising the core could.

 

[Bo_Fox]

Yeah I remember I read that the last week. TBH that's far more below expectations than the performance. The Image Quality is horrible!!

Hopefully ATI fixes that ASAP. BFG10K did an excellent job at uncovering the details. To me, that's "brilinear AF" instead of trilinear AF. There are not any visible grids under the transparent red area.


ATI R800:

Now, look at Nvidia's and see how nearly half of the red area is transparent in showing the moire grid.
Nvidia:

And then look at the 4890:

(source: http://alienbabeltech.com/main/?p=12648&page=2 )

I'd rather have Nvidia's trilinear AF method any day over ATI's brilinear method (these sharp lines of transition are unacceptable)! The article shows how it affects true image quality in games!!!!!!!!!!!

Let's spread the word around.. the quicker we do it, the more ATI is pressured to improve it ASAP.


(UPDATE: I'm wondering if this brilinear AF "optimization" was quietly introduced to just 5770 cards, because this one from bit-tech.net:

shows transparent details under the red area for 5870's new AF method.

Perhaps it's just with the 5770 that AF "optimizations" are secretly being forced on, as revealed by BFG10K?

_OR_ perhaps the checkerboard image from bit-tech.net is not necessarily hidden anyways, so using the program that BFG10K had would still show the same problems with both a 5770 and a 5870?

 

[Steevo]

Ohh yes, he did, after reading my comment at the bottom of that page you just linked to. Read it and enjoy!

I'll go ahead and quote it:



to you, Steevo! Heh!

Conclusion

The results generated today seem to disprove the commonly accepted idea that the 5770 is primarily held back by memory bandwidth. In actual fact ATi seems to have equipped the card with enough bandwidth to make it a reasonably balanced part overall.

Also based on the results generated today, if you’re trying to get more performance from your 5770, you should clock the core as high possible.

 

[Bo_Fox]

AND you should clock the memory as high as possible.

Yes, it's as Benetanegia said, in that the R800 cards are still "balanced" (with the 4890 performing "only" 20% better than a 5770). I already said that the GPU core is the main muscle that does the work, not the memory. You do not have to keep on repeating things after me, dude.

I'm just spoiled by cheap memory bandwidth. I want MOAR!!!

Not just the core! The memory too! 128-bit is just too cheap for my tastes. I havent had anything less than 256 bit in the last many years. My 9800 Pro in early 2004 was 256-bit. Then I had an X800XT All-in-Wonder, then a 6800 Ultra AGP, then 6800 Ultra PCIE, then X1900XTX, then 7900GT, then 7900GTX SLI, then 8800GTX, then HD 3870, then HD 4850 (DOA), then another 8800GTX, then HD 4870. Even my Geforce 4 Ti4200 in 2002 was 128-bit (the same year that 9700 Pro made good use of 256-bits). I've been such an enthusiast for more than 10 years with lots more cards before that, and I really understood just how much each card could have done better with more bandwidth (some by only a very little bit, like the HD 3870 or 7900GTX). The card with by far the most "redundant" bandwidth ever was HD 2900XT. All of the above cards were overclocked (and a few over-volted). The only card that I had that was nearly as "short on bandwidth" as the 5770 was the 4850, but it was DOA so I returned it for another 8800GTX which did just as well (at least until I got that 4870 1GB).

Just to save the company only like $10 in manufacturing costs by making 128-bit instead of 256-bit (definitely worth $20 if the performance increase is 20% or perhaps upwards of 35% in some cases.. I thought that the G92 series were rather bottlenecked by the 256-bit bandwidth, and that's 2 generations ago! The R800 series could also use a whole heck lots more bandwidth at a lower cost than having to resort to a bigger chip with more shaders/TMU's. I want a good quality 300HP sports car with a 5 or 6-speed tranny that is not so expensive to build nowadays, not a 300HP car with a cheap 4-speed tranny! If that 6-speed tranny improves its 0-60 acceleration time by 20% (4.8 seconds instead of 6 seconds), then hell yeah you're getting a lot more out of that 300HP engine.

Dude, I'm just not so against having more memory bandwidth.

 

[HalfAHertz]

HalfaHertz, hope you didn't take anything personally. When I asked you if you were a Mexican, I was only shooting out goofy questions.. if you are, then that's cool! I love enchiladas and think Mexicans are cool.

Lol haha don't worry I didn't.
Those are some sweet articles on alienbabletech. Thanks for the links.

 

[Bo_Fox]

Lol haha don't worry I didn't.
Those are some sweet articles on alienbabletech. Thanks for the links.

LOL...

Conclusion

The results generated today seem to disprove the commonly accepted idea that the 5770 is primarily held back by memory bandwidth. In actual fact ATi seems to have equipped the card with enough bandwidth to make it a reasonably balanced part overall.

Also based on the results generated today, if you’re trying to get more performance from your 5770, you should clock the core as high possible.

Even though you quoted my post, I do not think you understood it or even read it. I quoted BFG10K's new conclusion-- see post # 439 (scroll up just a little bit).



By the way, after you admitting this two times already in this thread:

I also am a bit buzzed.

This is me with a few shots, just ogt back from ......uhh.......yeah.

Please try to refrain from P.U.I. (Posting Under the Influence)!

 

[Bo_Fox]

Hey guys, perhaps it would make sense if I explained a bit more about the bandwidth increase being less linear than the core clock increase, but more linear than the buffer size increase, using the following quote (by the way, if he's a Nvidia guy talking about Fermi/future cards, it does not matter because both ATI and NV cards still need a similar amount of bandwidth):

With 3-D interconnects, it can vertically connect two much smaller die. Graphics performance depends in part on the bandwidth for uploading from a buffer to a DRAM. "If we could put the DRAM on top of the GPU, that would be wonderful," Chen said. "Instead of by-32 or by-64 bandwidth, we could increase the bandwidth to more than a thousand and load the buffer in one shot."

Based on any defect density model, yield is a strong function of die size for a complicated manufacturing process, Chen said. A larger die normally yields much worse than the combined yield of two die with each at one-half of the large die size. "Assuming a 3-D die stacking process can yield reasonably well, the net yield and the associated cost can be a significant advantage," he said. "This is particularly true in the case of hybrid integration of different chips such as DRAM and logic, which are manufactured by very different processes."

http://www.semiconductor.net/article/print/438968-Nvidia_s_Chen_Calls_for_Zero_Via_Defects.php



See, when John Chen said that the vastly increased bus helps to load the buffer in one shot, he was talking about the memory bandwidth and not the buffer size, like having 2GB or 4GB of video RAM. Sometimes, having more bandwidth actually reduces the need for more RAM (similar to the super-fast embedded RAM found on PS3, etc).


I mean, wow!!! It's not just me, but John Chen who would like some more serious bandwidth, upwards of (1024x8 = 8192-bit bus!!!) Even though the GTX 285 already has more bandwidth than a 5870 (and the GT300 is going to have ~50GB/s more than that), he's still dissatisfied enough to talk about needing MOAR.. especially moar than having 512-bit GDDR5!

ATI or Nvidia, bring on those vertically stacked dies!

(Now, I can go to bed knowing that it's not only me who wants MOAR!!! )

 

[Abir053]

hey all, i bought a sapphire 5870 1gb yesterday. i replaced my xfx gtx260 core 216. but i'm not very happy right now, coz i was expecting a huge change in gameplay. anyws the performance is not very consistent with games like crysis, crysis warhead, nfs shift. the strange thing is last night i was getting only 22fps avg. at crysis(very high) but there is a sound choppyness, but then i changed my x-fi control panel entertainment mode to game mode and wow!! the fps went 22 to 32-34. so... i should be happy right?? not really, coz after restarting pc same problem occured with game mode!!!


overall very confusing gameplay...sometimes it's working properly....but soetimes not. why is that??

anyone have any suggestion what might be causing this.

and sorry for my bad english!!!

 

[Bo_Fox]

hey all, i bought a sapphire 5870 1gb yesterday. i replaced my xfx gtx260 core 216. but i'm not very happy right now, coz i was expecting a huge change in gameplay. anyws the performance is not very consistent with games like crysis, crysis warhead, nfs shift. the strange thing is last night i was getting only 22fps avg. at crysis(very high) but there is a sound choppyness, but then i changed my x-fi control panel entertainment mode to game mode and wow!! the fps went 22 to 32-34. so... i should be happy right?? not really, coz after restarting pc same problem occured with game mode!!!


overall very confusing gameplay...sometimes it's working properly....but soetimes not. why is that??

anyone have any suggestion what might be causing this.

and sorry for my bad english!!!

I think it has more to do with the audio drivers than the 5870. Try to duplicate the "tweak" again by changing from entertainment mode to game mode again, and if it works, then great. If that's the only thing that you changed, then that must be it, I guess.

Also, run drivercleaner or driversweeper (forgot which one is the newer one..) in safe mode to properly remove all the leftovers of nvidia drivers, then reinstall catalyst.

 

[Abir053]

I think it has more to do with the audio drivers than the 5870. Try to duplicate the "tweak" again by changing from entertainment mode to game mode again, and if it works, then great. If that's the only thing that you changed, then that must be it, I guess.

Also, run drivercleaner or driversweeper (forgot which one is the newer one..) in safe mode to properly remove all the leftovers of nvidia drivers, then reinstall catalyst.

yeah, my x-fi driver is conflicting with ati catalyst for sure. i updated x-fi driver through windows update and it's better now....but not perfect. i'm still getting some sound choppyness mostly in crysis and when there is choppyness the frame rate drops like hell...

i guess i need to remove both my ati and x-fi driver and reinstall them like u said..

actually i'm a die hard fan of nvidia and this is my first ati card. so pls help me out.....thnx.

 

[Zubasa]

i agree with Bo_Fox. I mean the GTX 260 is hell of alot bigger than the GTS 250 with much better specs, yet the GTS 250 is only 6% on avg slower and even less at higher resolutions. I personally agree that had nvidia gave the GTS 250/9800GTX a 384-bit bus with 24 ROP's it probly would have closed that gap with less shaders. But eh we all can only think of what would have made sense.

Well...
A G92 with 384-bit bus and 24-ROPs is exactly a G80 with a die shrink
The whole point of the G92 is to create a more affordable GPU....

 

[Mussels]

Well...
A G92 with 384-bit bus and 24-ROPs is exactly a G80 with a die shrink
The whole point of the G92 is to create a more affordable GPU....

exactly.

large memory buses crank up the cost of the cards - a smaller bus with more bandwidth (higher clocks, GDDR5) is cheaper to produce, meaning cheaper to sell

 

[Lionheart]

They should change this thread to HD5970 below expectation!