ATI Radeon HD 4000 Series discussion

[Baron]

How does the 4870 look at comparing to one 4850 or a couple in crossfire?

[biohead]

CF dont gives you 1024mb frame buffer, it only gives you 2x512mb. You will expire the same slowdowns when the 512mb are filled.

If CF, only with 2x1024mb pls.

I was about to say that. We need 1GB cards.

[Stuen4y]

How does the 4870 look at comparing to one 4850 or a couple in crossfire?

There were some tests showing 3d mark differance between 26 and 28% but in the last two hours I have viewed more than 20 reviews, so I can't find the link right now . Found it Click

[Apoc@]

RV770 AIB Info Slides leak :P


http://www.ati-forum.de/general/news...-the-partners/

[perkam]

source

At stock on air

Perkam

[AliG]

Well, the memory hasn't had a significant affect on the benchmarks thus far (20% increase is from 20% core clock increase). I'm assuming the memory speed is bottlenecked somehow, but I have 0 experience in GPU memory bottlenecks and what instigates them

Eeither way, these will fly on LN2, run on air, and make breakfast at stock (Yes these will be hot).

@Cadaveca, there should be 1GB GDDR5 versions out later, and frankly, the GTX 260 not having 1GB does not escape it from the same criticism. I do, however, find your comment a tad juvenile. The increased speed on the GDDR5 modules does help in its ability to process textures faster than GDDR3 at the same memory size. That is why we have memory bandwith, and it does account for something

Perkam

Yeah that's odd, especially since the performance difference is still only around 25% even at crysis with 8xaa, the memory seems to really only make a 5% difference (performance gap goes from 20% to 25% once you hit the max details) and only with max details.

[Luka_Aveiro]

becuase the cpu goes idle waiting for vgas..this "lapse" while cpu is idle, then loaded, leaves spurts of traffic over the pci-e bus, causing stutter as data from the cpu slows the data flowing in the card out to monitor.


Luka, just check out the slide pics in this very thread about microstutter...it's shown there too.


this is not the only cause of microstutter, but it's definately an issue.



Phenom isn't too bad...it could use 3.2-3.4ghz and a slightly faster bus. Core2 @ 3.0-3.2 400mhz FSB or higher is ideal.



Remember boys, I'ver got dual 3870x2's...I'm dealing with the issue on a daily basis. I've almost completely eliminated the problem...some apps are more prone to it than others though.


It's a problem with memory control...within the vga. vga gets data from cpu, it must buffer the data...no data, it works only on data available, and will use as much of the ringbus as possible.

Cpu sends some data down....it goes on the ringbus too...but now it interferes with rendering as the dispatch controller must deal with pixels in flight, and pixels data incoming.

So we need more efficient and wider bus-interfaces, so data can flow and not stutter.

Nehalem should bring that along with triple channel, shouldn't it?

But it also brings a problem, if my line of thinking is correct: internal ring-buses should also be wider, in order do deal with so much data, correct?

I mean, what advantage would be having so much bandwidth from CPU to GPU, if the GPU ring-bus isn't able to deal with so much info? Coming to think of that, it surely seems a cause to multi-gpu solutions failure, small ring buses on each gpu, do not allow them to create a fluid data path, does it?

[emoners]

so CF benefits more on higher FSB?

[ORBR]

[Luka_Aveiro]

So, if i understood this correctly, 2 controllers on different GPUs will be able to access the same 32bit memory module, through a 16-bit path each, allowing them to share the same buffered info.

Sweet

[shiznit93]

That slide has nothing to do with our problem, clamshell is for using double the chips not multiple controllers. Notice there is only one controller in both pictures.

[cadaveca]

So, if i understood this correctly, 2 controllers on different GPUs will be able to access the same 32bit memory module, through a 16-bit path each, allowing them to share the same buffered info.

Sweet

no, one controller can address two moduels at once in clamshell. It means that one gpu could potentially write to another's framebuffer. AS noted in the pic above, clamshell only allows for doubling the buffer size. You could run into page conflicts when two memory controllers each address the same pcs of GDDR5.

Got E-peen, Cadaveca ?

When the prices for GDDR5 go down a bit, you will see 1GB cards. You can also crossfire two 4850s if you need 1GB that badly, and most people that have a $1000 to spend on a monitor usually do go xfire and sli, so again, your concern is a tad unfounded.

Perkam

Yes, got my epeen alright.

Um, Perkam, I've got 3870x2's...two of 'em...512mb framebuffer for each gpu, 1gb for each card, and it's not enough for 2560x1600.

Unfounded? lol. You look really out of place now, don't you? Remember, I'm the guy buying high-end parts on release...I'm the one that has those concerns.

You, on the other hand, buy entry-level parts, so I understand your perspective, but you should also understand mine, rather than trying to undermine it, and you should also face reality a bit here.

I'm the ATI fanboi...I'm not knocking the products, merely highlighting my needs as an extreme enthusiast. Surely you can understand that?

[Xello]

Indeed they bloody well better have 1gb frame buffer 4870X2's out, at launch. Don't care about price, don't care about the heat, or the power usage. Having to switch to crossfire because of the wreck that was the 790i is a pain enough but if i'm going to be stuck with 512mb buffers... Not worth thinking about.

[Luka_Aveiro]

no, one controller can address two moduels at once in clamshell. It means that one gpu could potentially write to another's framebuffer. AS noted in the pic above, clamshell only allows for doubling the buffer size. You could run into page conflicts when two memory controllers each address the same pcs of GDDR5.

Doesn't this happen already with 1GB 8800GT (or cards that have doubled vram over reference)? If so, nothing new, here...

[Jowy Atreides]

I'm the ATI fanboi...I'm not knocking the products, merely highlighting my needs as an extreme enthusiast. Surely you can understand that?

[Luka_Aveiro]

http://news.softpedia.com/images/news2/NVIDIA-to-Release-GTX-280-GX2-3.jpeg

[zerazax]

Ok that pic is too hilarious lol!

[cadaveca]

So we need more efficient and wider bus-interfaces, so data can flow and not stutter.

Nehalem should bring that along with triple channel, shouldn't it?

But it also brings a problem, if my line of thinking is correct: internal ring-buses should also be wider, in order do deal with so much data, correct?

I mean, what advantage would be having so much bandwidth from CPU to GPU, if the GPU ring-bus isn't able to deal with so much info? Coming to think of that, it surely seems a cause to multi-gpu solutions failure, small ring buses on each gpu, do not allow them to create a fluid data path, does it?

NOw you may understand why R600 had 512-bit memory control. ATI foresaw the problem with so much data in flight, as well as issues in re-syncing the dispatch processor with the new data efficiently.

AMD stepped in and lopped off half of that bus, and stutter is much worse on RV670 than R620 because of it. There are other issues plaguing RV670(powerplay) though, so anything that exhibited behavior like stutter has been singled out now.

Of course, a faster bus may help with stutter, but more importantly, it will get data to the cards much quicker, and when dealing with 4 gpus...this is very important. It's(beating stutter) more about a proper balance to the workload, and with todays hardware, it's easy to end up having a bottleneck fixed, only to have another appear elsewhere!

[Rammsteiner]

Cheapest HD 4850 I see now is 150 Euro's
HD 4870 is on pre-order, cheapest I saw 269 Euro

[ghost101]

I want to clear up some review sites showing awful results for the hd4850.

Lets take cod4 as an example.

The following all show the 4850 outperform the 9800gtx

computerbase.de - http://www.computerbase.de/artikel/h...call_of_duty_4

guru3d - http://www.guru3d.com/article/amd-at...-powercolor/11

techpowerup - http://www.techpowerup.com/reviews/P...HD_4850/6.html

Then these benchmarks show awful performance

pcgh - http://www.pcgameshardware.de/aid,64...k-Test/&page=5

pcper - http://www.pcper.com/article.php?aid...e=expert&pid=3 <--- using heavily oced nvidia cards

The PCGH review clealrly has something wrong with their test system and should be ignored. There are systematic differences throughout their benchmarks. Bioshock being the other game with terrible performance.

[zerazax]

And even in the PCGH review, they use Qarls Texture Pack and at 1920 x 1200, 8xFSAA, 16xAF, they beat the GTX280 with a single lowly 4850:

http://www.pcgameshardware.de/aid,64...icle_id=648091

Somehow the 512MB card is beating the 1GB card even with the texture pack at those settings. Either something is really wrong w/ the GTX280 or the 4850 has some ridiculous power

And yeah, the PCPER article is showing the 8800GT OCX (700 core, 1728 shader) and the 9800GTX (BFG base 9800GTX is at 700 core) so those are all OC'd G92's and the 4850 still hangs with and even beats it, especially w/ AA on.

[Hornet331]

I want to clear up some review sites showing awful results for the hd4850.

Lets take cod4 as an example.

The following all show the 4850 outperform the 9800gtx

computerbase.de - http://www.computerbase.de/artikel/h...call_of_duty_4

guru3d - http://www.guru3d.com/article/amd-at...-powercolor/11

techpowerup - http://www.techpowerup.com/reviews/P...HD_4850/6.html

Then these benchmarks show awful performance

pcgh - http://www.pcgameshardware.de/aid,64...k-Test/&page=5

pcper - http://www.pcper.com/article.php?aid...e=expert&pid=3 <--- using heavily oced nvidia cards

The PCGH review clealrly has something wrong with their test system and should be ignored. There are systematic differences throughout their benchmarks. Bioshock being the other game with terrible performance.

lulz, why does it sound like fanboy talk... review is not like it was expected, so ignore it?

maybe you would consider that they use other bench settings (custom time demo etc.) befor saying they are rubish....

[zerazax]

I don't think the PCGH article setup is bad, I mean it clearly shows that at times even the GTX280 somehow loses to the 4850, such as my Oblivion example.

The PCPER article should've clearly stated the 8800GT and 9800GTX were OC'd versions to eliminate confusion, but I guess they did a rush job to put the article up. So either way you look at it though, the 4850 hangs and even beats OC'd versions of the GTX.

[AliG]

yeah that PCPER article was a horrible review, if anything they should have oc'd the 4850 to 700mhz core if they were going to compare the card to gt and gtx oc cards. Samething that was said about the gtx 280, its a new card, give the drivers some time, look at what happened with the r600's performance, those last 4 days it went up by almost 20&#37;


I think we'll be happy with the results in the end

[hurleybird]

AS noted in the pic above, clamshell only allows for doubling the buffer size. You could run into page conflicts when two memory controllers each address the same pcs of GDDR5.

Perhaps, but you have no way of knowing if ATI has found away around that problem. This info certainly adds weight to the rumors of the two RV770s of R700 communicating to each other through memory though.