HD 5870 Discussion thread.

[phanbuey]

look at the ultimate product... 1 card is going toe to toe with the current top gen dual cards - and will beat them in DX11 w/ tesselation nd all the API improvements. This card is not perfect - but its out and it is the fastest single gpu on the planet since 1900XTX for ati.

NV, on the other hand is nowhere to be seen. They have doubled performance ONCE since the G80, and their cycle is obviously longer than ATI's. From the ultimate product standpoint, I would take a 5870/5850 over a gtx285 or 295 any day - its a better product, irrespective of the point that the specs are double but the perf is only 85%greater in dx10 titles with one month old drivers.

the 5870 is practical, cheap to make, has a 256bit bus which makes it even cheaper, and delivers great performance. The gt300 is nowhere to be seen - most likely due to the issues that are inherent in the development of such an incredibly complex design.

 

[ToTTenTranz]

The HD5870 is very bandwidth-limited.. when using DX9 and early DX10 engines.


DX11 is a lot more about computing and smart usage of resources (take tesselation and multitasking i.e.) than the previous generations.

So for old engines, the HD5870 isn't twice as fast as the HD4890. But for future fully fledged DX11 engines wich use tesselation and compute shaders, the HD5870 should become faster than the HD4870X2 and even two HD4890 crossfired.


History has told that generally, ATI's graphics cards tend to be very future-proof (except for that R4xx generation fiasco). However, that future-proofness can only be noted when no one cares for the graphics card anymore.
Compare a X1900XT to a 7900GTX in a recent game and the first doubles the performance of the second.

 

[phanbuey]

Yes DX11 is all about efficiency and using compute shaders to perform stuff like AA and tesselation, dx10.1 shows how this is done. - Ati cards simply spank nv cards in 10.1 titles, and deliver a higher quality image with less cost.

When compute shaders will be the limiting factor (and they will be in dx11), then the 5870 will shine.

 

[Halos]

History has told that generally, ATI's graphics cards tend to be very future-proof (except for that R4xx generation fiasco). However, that future-proofness can only be noted when no one cares for the graphics card anymore.
Compare a X1900XT to a 7900GTX in a recent game and the first doubles the performance of the second.

can you proof that statement?That x1900xtx doubles the 7900gtx performance? I am very interested in that as i have x1900xtx...

Some rewiev or benches?

 

[Woody112]

can you proof that statement?That x1900xtx doubles the 7900gtx performance? I am very interested in that as i have x1900xtx...

Some rewiev or benches?

I use to have 1900xtx and a 1900xt crossfire card. you know the old school getto adapter/bridge. Yes that 1900xtx spanked the 7900gtx something offal, not sure about double performance but the 1900 defiantly treated the 7900 like a red headed steep child.

 

[Halos]

Well, dn. I trust you guys, but as the time is passing i really see that in my own spot. I had (a few months ago) Foxxcon 8800GTX, now i have 3870x2...3870x2 was 8800GTX/Ultra side by side card, and they were in match. NOw it seems that 3870x2 is stronger than GTX, all new games i have played on GTX are better on 3870x2...i dont want to generaly speak about future proof ATI, but in 3-4 years it seems like it is future proof

 

[ToTTenTranz]

can you proof that statement?That x1900xtx doubles the 7900gtx performance? I am very interested in that as i have x1900xtx...

Some rewiev or benches?

There aren't many.. but the most recent the game is, the bigger the difference in performance:

http://www.tomshardware.co.uk/charts/gaming-graphics-charts-q3-2008/Assassins-Creed-v1.02,736.html

http://www.tomshardware.co.uk/charts/gaming-graphics-charts-q3-2008/Call-of-Duty-4-v1.6,742.html (COD4 is the perfect example, with 60fps for the X1900XT against 30fps for the 7900GTX)

http://www.tomshardware.co.uk/charts/gaming-graphics-charts-q3-2008/Crysis-v1.21,747.html

http://www.tomshardware.co.uk/charts/gaming-graphics-charts-q3-2008/Mass-Effect,771.html


These results are over a year old, from Q3 2008. I remember seeing a Bioshock benchmark where the X1900XT showed more than double the performance of the 7900GTX. I just don't remember where.
But compare them to Q1 2008:

http://www.tomshardware.co.uk/charts/gaming-graphics-charts-q1-2008/Battlefield-2142,552.html

http://www.tomshardware.co.uk/charts/gaming-graphics-charts-q1-2008/Doom-3,561.html

and you'll see that in the beggining of 2008, the 7900GTX 512MB and the X1900XT 512MB were trading blows in BF2142 and Doom3, whereas in newer games the X1900XT seems to be in a whole other level.
The biggest problem is that the X1900XT was only benched widely when it came out 2006. So the general opinion that stuck was that the X1900XT was just on the same level as the 7900GTX.

And the most worrying example:

http://www.tomshardware.co.uk/chart...2008/3DMark06-v1.0.2-HDR-SM3.0-Score,540.html
Here, the X1900XT is actually 33% slower than the 7900GTX.
It just shows how 3dmark is not representative of the card's performance for future engines, contrary to what Futuremark says all the time.

 

[Bo_Fox]

Yep, 7900GTX had 24 TMU's which enabled it to do well in older games like Doom3 etc.. while X1900XTX had only 16TMU's (same for HD2900XT and HD3870). The thing is that newer games use far, far more shaders. DX9.0c is a lot about using much more shaders. X1900XTX's 48 shader units tripled the shader performance of an X1800XT, which was already beating a 7800GTX 256MB by a small margin. Ya know how Unreal3 engine games like Bioshock are so dependent on shaders? For the last 3-4 years, nearly all of the games use shaders like never before. However, the X1900XTX was a big, hot, and power-hungry chip. It ate nearly as much power as an 8800GTX, believe it or not. The amazing thing is how the 8800GTX was made on the same 90nm process, and yet hardly ate any more power despite its ~50% bigger chip size, and was oh, so much faster. I had two X1900XTX's since the first day it was released ($600 a pop), and it was quite a challenge for my power supplies and an array of heatsinks from back then (when a Zalman VF900 failed to cool it for stable operation at stock settings). My PSU booster came to the rescue after 3 500-600W PSU's failed!

Back to the point..

Just one more time:

A 5870 chip, which is like 2x 4890 chips infused into one, should never lose to two 4890's in crossfire. However, it frequently loses to 2x 4890's in more games than not, when it never should especially due to the removal of inefficient crossfire scaling.

Why does it lose out to 2x 4890's overall (and also 4870X2), and by quite a margin in some games? What is the reason for this?

The only thing that is not doubled over 2x 4890's is the memory bandwidth (and possibly cache also). The memory bandwidth of a 5870 vs a single 4890 is only increased by 23%, rather than double (100%).

 

[wolf]

Well.... Below expectations....? yes.

I had hoped everything would be doubled or more, which they did in some areas.

What I really WANTED the 5870 to have (on top of what it already has) is 2gb 512-bit GDDR5 @ 4+ghz, or whatever combination they need to get up over about 230gb/s

but maybe they tested silicon like this, and that STILL wouldst double a 4890's performance, so they cut it back so its NOT double, so everyone can find some justification for it not performing twice as fast.

maybe...

 

[Bo_Fox]

Well.... Below expectations....? yes.

I had hoped everything would be doubled or more, which they did in some areas.

What I really WANTED the 5870 to have (on top of what it already has) is 2gb 512-bit GDDR5 @ 4+ghz, or whatever combination they need to get up over about 230gb/s

but maybe they tested silicon like this, and that STILL wouldst double a 4890's performance, so they cut it back so its NOT double, so everyone can find some justification for it not performing twice as fast.

maybe...

Maybe, just maybe, but I think it got more to do with the 5870X2 reasons I stated in post 75.

 

[a_ump]

yea that was a good thought you had in that post Bo_Fox . i just would like to have tech magically jump forward about 2 months so we could see what HD 5870's potential is with mature drivers as well as Nvidia's response to it if there will be any new news on their side lol

 

[Bo_Fox]

Proposal for a benchmark test (for W1zzard and Btarunr)

Given that a 5870 chip should perform 100% better than a 4890 chip plus some DX11 features, in theory, especially since both are of the same architecture and a 5870 actually looks like two 4890's infused into one chip, a suite of benchmark tests is needed to confirm exactly how much the memory bottleneck is hampering a 5870. Despite the theoretical 100% performance increase of the chip, a 5870 card has only 23% greater memory bandwidth than a 4890.

Whenever a 5870 loses out to a dual 4890 in Crossfire (in identical settings), the memory bottleneck (and possibly L1/L2 cache bottleneck also) is to be blamed rather than the drivers. Crossfire scaling is never 100% efficient due to additional CPU overhead and some other factors. If the drivers could scale two 4890's so that it beats a 5870 by a wide margin in some games, the drivers should be able to scale all of the shaders/cores in a 5870 chip just fine since it is a very similar chip architecture. ATI is not known for introducing a new driver set that brings such incredible performance boosts like Nvidia has done a couple times in the past several years, so this is not to be expected.

An excellent test would be to reduce the memory bandwidth of each 4890 to exactly half of a 5870, and then crossfire the 4890's together to see how it compares against a 5870. If the 5870 is now at least equal to 2x 4890's in all of the games in which the 5870 was previously beaten by 2x 4890's (let alone a 4870X2), it would confirm this issue for once and for all. The first hardware-review site to do this test could gain quite a bit of popularity and respect.

 

[phanbuey]

Given that a 5870 chip should perform 100% better than a 4890 chip plus some DX11 features, in theory, especially since both are of the same architecture and a 5870 actually looks like two 4890's infused into one chip, a suite of benchmark tests is needed to confirm exactly how much the memory bottleneck is hampering a 5870. Despite the theoretical 100% performance increase of the chip, a 5870 card has only 23% greater memory bandwidth than a 4890.

Whenever a 5870 loses out to a dual 4890 in Crossfire (in identical settings), the memory bottleneck (and possibly L1/L2 cache bottleneck also) is to be blamed rather than the drivers. Crossfire scaling is never 100% efficient due to additional CPU overhead and some other factors. If the drivers could scale two 4890's so that it beats a 5870 by a wide margin in some games, the drivers should be able to scale all of the shaders/cores in a 5870 chip just fine since it is a very similar chip architecture. ATI is not known for introducing a new driver set that brings such incredible performance boosts like Nvidia has done a couple times in the past several years, so this is not to be expected.

An excellent test would be to reduce the memory bandwidth of each 4890 to exactly half of a 5870, and then crossfire the 4890's together to see how it compares against a 5870. If the 5870 is now at least equal to 2x 4890's in all of the games in which the 5870 was previously beaten by 2x 4890's (let alone a 4870X2), it would confirm this issue for once and for all. The first hardware-review site to do this test could gain quite a bit of popularity and respect.

aye aye... also for one with new drivers. I feel like the drivers bottleneck the cards as well.

 

[a_ump]

well there was something mentioned in one of these threads about 9.12 supposedly going to give a big boost to HD 5XXX performance and that the series is currently not operating at full capacity according to the article. Now that makes me wonder, cause 9.11's aren't even out so why can't they implement these rumored improvements into 9.11? and how would they know bout how to improve performance when 9.11's haven't been released for ATI to get feedback. So if they already know how to improve performance but aren't doing it for current drivers then it makes me think that they are purposely holding back HD 5XXX performance. Why? i can only guess, if the rumors are true, that ATI's been holding back performance on purpose to surprise nvidia if they release in 09 by boosting HD 5XXX performance with 9.12's. It seems logical and possible to me since from a technical stand point the HD 5870/50 should be performing better.

 

[Mussels]

9.11 leaked a while back actually, its main focus was on openCL


its easy enough to see that 9.11 focuses on openCL for all compatible cards, which gives them time to work on 5xx0 performance in 9.12

 

[a_ump]

well i'd be jumping for joy if 9.12's boost performance by 10%+ in everything. more pressure on Nvidia

 

[[I.R.A]_FBi]

kick out teet

 

[Steevo]

I love all the NV fanbois in tehis thread. It really shows how there is NOTHING NEW at the NV camp besides some left over vapor they all are breathing.


Drink the green kool aid in your own thread.

I know you all love your precious NV and they can do no wrong, and you must come spread you revealing light upon us poor souls and show us the way and the light, but really we are happy with our new generation of hardware, and the time to refine drivers, and the use of DX11. We understand and forgive you for your holy nvidia jihad.


Now go play some DX9 games and spank off to that.

The HD5870 is very bandwidth-limited.. when using DX9 and early DX10 engines.

No.

It's not odd at all. One suspects that doubling specs is the way of trying to achieve twice the performance. If not, if Ati doubled the specs knowingly expecting a mere 50% performance increase, then Ati's architecture and their long term strategy is a BIG BIG fail.

Thie plan I believe is to use the shader cores to perform the work NV makes you think only they can do, you know, a poorly spelled and implemented proprietary hardware and software design that they drop support for faster than a prom dress?

RV870 is not a redesing, far from it. And I say it's a fail, if they knew it from the drawing board, because if they have to double up the silicon to obtain a 50% increase in performance they won't get too far in the future.

Besides GT300 will not be expensive necessarily, it won't be expensive to make, not like GT200 or even GT200b. Price will depend entirely on how fast Nvidia wants to recover the R&D money and if they decide to come back with a vengance to RV770... I mean Ati cards have doubled specs and have not scaled accordingly, Nvidia has more than doubled specs, if they scale like they did in every previous generation...

NV 56% vs ATI 66% increase?

Yeah. 16 bit video anyone? Green Kool Aid?

And lastly, does you notice massa the SLI beating the poor ol 9800 X2, oh loawdy fo sho who eveah made dis heah chart mussa been from da devil!!!!!

 

[[I.R.A]_FBi]

I am amused by previous post

 

[Bo_Fox]

well there was something mentioned in one of these threads about 9.12 supposedly going to give a big boost to HD 5XXX performance and that the series is currently not operating at full capacity according to the article. Now that makes me wonder, cause 9.11's aren't even out so why can't they implement these rumored improvements into 9.11? and how would they know bout how to improve performance when 9.11's haven't been released for ATI to get feedback. So if they already know how to improve performance but aren't doing it for current drivers then it makes me think that they are purposely holding back HD 5XXX performance. Why? i can only guess, if the rumors are true, that ATI's been holding back performance on purpose to surprise nvidia if they release in 09 by boosting HD 5XXX performance with 9.12's. It seems logical and possible to me since from a technical stand point the HD 5870/50 should be performing better.

That's because a few of the people wanted to blame it on the drivers. That's called "wishful thinking" because so many, many things support my theory that it should be related to the "slightly" increased memory bandwidth over a single 4890, which already showed consistent improvements when the memory alone was overclocked, and related to the "loss" of the 5870 to a 4870X2, which is worse than 2x 4890's in crossfire.

Let the logic do the talking instead of always believing the first few random people by coincidence.

Also, Steevo is correct in that the GTX 280 did not perform twice as good as a 9800GTX, but people already expected this right before the release of GTX 280's. The GTX 280 had rather slower core clock speeds and hardly any more TMU's than a 9800GTX. However, the 5870 has exactly twice the cardinal specification figures of a 4890 in that it is predicted to be exactly twice as fast as a 4890, and perhaps even more so if the memory was more than twice as amperous as that of a 4890.

I'm just pissed off at ATI because I want more speed now!!!

 

[Zubasa]

Given that a 5870 chip should perform 100% better than a 4890 chip plus some DX11 features, in theory, especially since both are of the same architecture and a 5870 actually looks like two 4890's infused into one chip, a suite of benchmark tests is needed to confirm exactly how much the memory bottleneck is hampering a 5870. Despite the theoretical 100% performance increase of the chip, a 5870 card has only 23% greater memory bandwidth than a 4890.

Whenever a 5870 loses out to a dual 4890 in Crossfire (in identical settings), the memory bottleneck (and possibly L1/L2 cache bottleneck also) is to be blamed rather than the drivers. Crossfire scaling is never 100% efficient due to additional CPU overhead and some other factors. If the drivers could scale two 4890's so that it beats a 5870 by a wide margin in some games, the drivers should be able to scale all of the shaders/cores in a 5870 chip just fine since it is a very similar chip architecture. ATI is not known for introducing a new driver set that brings such incredible performance boosts like Nvidia has done a couple times in the past several years, so this is not to be expected.

An excellent test would be to reduce the memory bandwidth of each 4890 to exactly half of a 5870, and then crossfire the 4890's together to see how it compares against a 5870. If the 5870 is now at least equal to 2x 4890's in all of the games in which the 5870 was previously beaten by 2x 4890's (let alone a 4870X2), it would confirm this issue for once and for all. The first hardware-review site to do this test could gain quite a bit of popularity and respect.

What you are doing is basically compareing a 4Ghz single core against a 2Ghz dual core.
A chip that has double spec is never double as fast, its simple law of diminishing returns dude.

It all depends on how the game is programed.
The 4870X2 is actually slightly slow than the 4870 1GB CF, and the GTX 295 is in many cases beaten by GTX 275 SLI.

Another point of interest is CF and SLI in some cases scales over 100%.
I believe Wolf did a quick testing on that.

 

[Bo_Fox]

What you are doing is basically compareing a 4Ghz single core against a 2Ghz dual core.
A chip that has double spec is never double as fast, its simple law of diminishing returns dude.
Another point of interest is CF and SLI in some cases scales over 100%.
I believe Wolf did a quick testing on that.

A 5870 is like a "perfectly" efficient CF of 2x 4890's plus some DX11 effects. Mind you, if the memory bandwidth was also 100% greater than that of a 4890 (with equal latencies), using 512-bit memory, it would actually "never" lose out to the less efficient crossfired pair of 4890's in any of the games, even if it was beaten so badly (to the point where it's only like 60% of a 4870X2's performance).

 

[Zubasa]

A 5870 is like a "perfectly" efficient CF of 2x 4890's plus some DX11 effects.

It is not perfectly efficient, for nothing really is.
The funny thing is CF and SLI can sometimes scales more than "perfect".

 

[wolf]

What you are doing is basically compareing a 4Ghz single core against a 2Ghz dual core.
A chip that has double spec is never double as fast, its simple law of diminishing returns dude.
Another point of interest is CF and SLI in some cases scales over 100%.
I believe Wolf did a quick testing on that.

I sure did, but I can't find the link, trawling right now.

It is a crazy anomaly, but yes I and other users have indeed seen 2 cards scale in excess of 100% over 1, and in a well optimized engine, scaling consistently at 97-99% is not uncommon either.

 

[Bo_Fox]

It is not perfectly efficient, for nothing really is.
The funny thing is CF and SLI can sometimes scales more than "perfect".

Yes, in terms of theoretical core performance and bandwidth (both memory and cache) and latency (more will be explained about how each can become a bottleneck, and how there's a threshold to each of those things that when exceeded, give much bigger increases).

Two certain factors come into mind. If the cache is not also linearly increased in the terms of bandwidth or latency, or both in combination (those are two different but similar factors), simply multiplying two "half" cores could actually give faster output even if the memory bandwidth is not doubled using the same latency -- as long as the cache does not become bottlenecked.

But I bet with 50:1 odds that using a 512-bit bandwidth with the same speed GDDR5 memory will bring dramatic results to a 5870.


The reason why CF and SLI sometime scale more than 100% is because the memory "bottleneck" threshold is finally exceeded, or because the ROP back-end is no longer suffocated like the back of a thin exhaust on a cheap car.

So should a 5870 be even more likely to scale more than 100% since it is not "subservient to Crossfire demands or programming inefficiency".

Another bottlenecking factor here is the CPU. Overclocking Core i7's to 4+ GHz and benching at resolutions of 1920x1200 or more will keep nearly all of the CPU-bottlenecking factor out of most of the benchmarks, and have been done in quite a number of reviews.