5870 Bottleneck Investigation (CPU and/or Memory Bandwidth)

[Raptor-X]

dont forget to check for memory error correction (more details on overclocking page of my review)

Link Please

[W1zzard]

http://www.techpowerup.com/reviews/A...D_5870/33.html

[Raptor-X]

http://www.techpowerup.com/reviews/A...D_5870/33.html

Thank you.

[jaredpace]

Thanks lightman, wish elmore would let us know how he got his vmem up and clocked the mem at 1400mhz...
however, from your tests on the canyon flight FPS @ 850mhz core

memory mhz / FPS / fps Increase
0900 / 091.7 / 0 0%
1000 / 095.2 / +3.5 +3.8%
1100 / 097.9 / +2.7 +2.9%
1200 / 100.3 / +2.5 +2.5%
1300 / 102.4 / +2.1 +2.1%

According to this test, as the memory clock increases, gains are evident at 1200 -> 1300mhz. Even though they are diminishing as it goes higher, you are still getting benefits above the stock mem clock, which backs peoples assumptions that the card is bottlenecked @ 153gbps. Would like to see memory at 1400mhz, if it yields a ~2% increase over 1300mhz, it's even further proof.

Lightman's overall FPS is still increasing so his memory is not producing errors yet

"Overclocking the memory on these cards is quite different from any other card so far. Normally you'd expect rendering errors or crashes, but not with these cards. Thanks to the new error correction algorithm in the memory controller, every memory error is just retransmitted until everything is fine. So once you exceed the "stable" clock frequency, memory errors will appear more often, get retransmitted, but the rendered output will still look perfectly fine. The only difference is that performance drops, the further you increase the clocks, the lower the performance gets. As a result a normal "artifact scanning" approach to memory overclocking on the HD 5800 Series will not work. You have to manually increase the clocks and observe the framerate until you find the point where performance drops."
http://www.techpowerup.com/reviews/A...D_5870/33.html

[Lightman]

Yes, I'm aware of ECC on Cypress. So far my memory won't cooperate @1400, but it is working @1350 (more tests needed to check for ECC)
This lets me believe 1300MHz should be stable.
BTW my core scales to 950MHz on default volts.

From my tests what caught my attention was Fillrate tests and how nice they respond to memory (for pixel) and engine (for texel) clocks! Also perlin noise seems to be purely GPU limited.

I think for what it's worth 1200MHz is a good compromise at this point of time.
This leaves AMD with doors opened for HD5890 with higher clocks and at the same time higher performance! It looks like 1GHz/1400 should be doable on current tech. using 6Gbps IC's (interconnect limited?).

[saaya]

BTW first and foremost bottleneck is triangle setup. Same speed per clock as in RV770!
Let's look for more

huh? how come?

but it cant be a big limitation, otherwise a 4870x2 would pull ahead notably in most tests...

[saaya]

thx lightman for the infos!
you dont happen to have a 4890 or 4870 or better yet 4870x2 so you can do exactly the same thing on that card in the same rig?

i threw the numbers in a graph:
Pixel shader, multi fillrate, perlin noise, vertex shader simple are not bandwidth limited... yes some of these increase slightly, but the boost they get from increasing the memory clocks and hence bandwidth by almost 50% can be ignored...

complex vertex shader, shader particles and fillrate single are clearly bandwidth limited... but the gains are linear and proportional, so the gpu is definately not bandwidth crippled or starved like some people suggested... i dont think that memory bandwidth is limiting the 5870... i think its internal bandwidth or some internal architecture bottleneck...

[CrimInalA]

very interesting information

[demonkevy666]

thx lightman for the infos!
you dont happen to have a 4890 or 4870 or better yet 4870x2 so you can do exactly the same thing on that card in the same rig?

i threw the numbers in a graph:
Pixel shader, multi fillrate, perlin noise, vertex shader simple are not bandwidth limited... yes some of these increase slightly, but the boost they get from increasing the memory clocks and hence bandwidth by almost 50% can be ignored...

complex vertex shader, shader particles and fillrate single are clearly bandwidth limited... but the gains are linear and proportional, so the gpu is definately not bandwidth crippled or starved like some people suggested... i dont think that memory bandwidth is limiting the 5870... i think its internal bandwidth or some internal architecture bottleneck...

http://www.xbitlabs.com/articles/vid...on-hd5870.html

I'm still looking for this "bottleneck" on the architecture.
page three...

[chispy]

http://www.techpowerup.com/reviews/A...D_5870/33.html

Than you W1zzard .

Regards. Chispy

[hurleybird]

Well, I'm getting my card on monday and should be able to do some more detailed tests, but at first glance it doesen't seem like bandwidth is that big of bottleneck. Either there's some obscure and unforseen bottleneck in the architecture, in which case we can probably expect some kind of very slightly modified but much faster fixed "RV890" chip to compete against GT300, perhaps in time to coincide with Nvidia's launch. More likely though, it's just immature drivers and broken optimizations that need to be updated for RV870. The fact that the card is launching relatively early, while eyefininty isn't working over CF, and SSAA is a broken blurry mess in the current drivers would seem to suggest this.

Again, there's no reason an enhanced and doubled HD4890 shouldn't outperform the HD4870X2, 4890CF, and the GTX 295. At the moment performance is good, but isn't anywhere close to where it should be.

[wiak]

sounds like a good combo with GDDR5 memory that is rated 5Gbps plus 1Ghz Core and 2GB memory
if you remember 512MB > 1GB on 4870 cards did boost preformace

[STEvil]

Just so you guys know, clocking the card to 1/2 speed will not result in the performance "scaling" to 1/2 performance.

[Crankybugga]

Just so you guys know, clocking the card to 1/2 speed will not result in the performance "scaling" to 1/2 performance.

Dont let any facts get in the way of good ol 5870 bashing.

I can see some experts here trying to track down some mysterious bottleneck crippling the HD5870, even though it out performs nvidias biggest and best dualcore monstrosity in some game benchmarks and poses a serious threat in most others.
Only nvidiots would portray a 40% performance improvement between models as too little, then claim to know about a mysterious bottleneck to back this opinion.
Its the same every model release when the opposing side feels threatened.

[dekruyter]

Dont let any facts get in the way of good ol 5870 bashing.

I can see some experts here trying to track down some mysterious bottleneck crippling the HD5870, even though it out performs nvidias biggest and best dualcore monstrosity in some game benchmarks and poses a serious threat in most others.
Only nvidiots would portray a 40% performance improvement between models as too little, then claim to know about a mysterious bottleneck to back this opinion.
Its the same every model release when the opposing side feels threatened.

Gotta say, I love this response, and agree. The 5870 is what it is....a "product", build on design, manufacturing, cost, and marketing/sales limitations. If the components are not equally matched, and thus there is hidden performance, so what. It is what it is. No one could ever claim that the 5870 components were perfectly tuned....to say within a few % of each other. It is "technically" interested what these test show, but in the real world what do they mean......they certainly have nothing to do w/ the decision to buy, or not buy.

And for disclosure's sake, I get my XFX 5870 on Monday.




.

[okorop]

Gotta say, I love this response, and agree. The 5870 is what it is....a "product", build on design, manufacturing, cost, and marketing/sales limitations. If the components are not equally matched, and thus there is hidden performance, so what. It is what it is. No one could ever claim that the 5870 components were perfectly tuned....to say within a few % of each other. It is "technically" interested what these test show, but in the real world what do they mean......they certainly have nothing to do w/ the decision to buy, or not buy.

And for disclosure's sake, I get my XFX 5870 on Monday.




.

I agree with you, but that 2% more performance only overclocking the memory is a sign that the card is a little bit bandwidth limited at least with same game, I will buy 2 of 5870 for sure but I would like to see some test with the memory at 1400mhz to see the difference......

[CrimInalA]

Gotta say, I love this response, and agree. The 5870 is what it is....a "product", build on design, manufacturing, cost, and marketing/sales limitations. If the components are not equally matched, and thus there is hidden performance, so what. It is what it is. No one could ever claim that the 5870 components were perfectly tuned....to say within a few % of each other. It is "technically" interested what these test show, but in the real world what do they mean......they certainly have nothing to do w/ the decision to buy, or not buy.

And for disclosure's sake, I get my XFX 5870 on Monday.
.

This got everything to do with buying the best product and overclocking it to the max . We are on xtremesystems here ..
Most people here want to get the most of of the product which they have bought . And not only that , but we want to learn more about this product and squeeze every last bit of perfomance out of it .

You seem to be one of those few people on here that is not interested too much in this material . But that's ok too

[saaya]

I'm still looking for this "bottleneck" on the architecture.
page three...

my guess is having the units in 2 800sp blocks might cause problems with the hw scheduler... that should be possible to at least be improved if not fixed with driver updates tho i think...

Well, I'm getting my card on monday and should be able to do some more detailed tests, but at first glance it doesen't seem like bandwidth is that big of bottleneck. Either there's some obscure and unforseen bottleneck in the architecture, in which case we can probably expect some kind of very slightly modified but much faster fixed "RV890" chip to compete against GT300, perhaps in time to coincide with Nvidia's launch. More likely though, it's just immature drivers and broken optimizations that need to be updated for RV870. The fact that the card is launching relatively early, while eyefininty isn't working over CF, and SSAA is a broken blurry mess in the current drivers would seem to suggest this.



Again, there's no reason an enhanced and doubled HD4890 shouldn't outperform the HD4870X2, 4890CF, and the GTX 295. At the moment performance is good, but isn't anywhere close to where it should be.

ill correct myself, bandwidth IS limiting the 5870, but not by a lot...

yupp, i find it hard to believe that a 4870x2 is using the two gpus on it 100% efficiently and hence performs the same as a chip that has the same resources in one gpu... that makes no sense whatsoever...

not to mention 5870 is even slower on average than a 4870x2... theres def something limiting...

Just so you guys know, clocking the card to 1/2 speed will not result in the performance "scaling" to 1/2 performance.

yeah, like with any chip there is a base performance... 5870s responce to memory bandwidth is about the same as most vgas though, it doesnt seem very memory bandwidth limited at all...

Dont let any facts get in the way of good ol 5870 bashing.

I can see some experts here trying to track down some mysterious bottleneck crippling the HD5870, even though it out performs nvidias biggest and best dualcore monstrosity in some game benchmarks and poses a serious threat in most others.
Only nvidiots would portray a 40% performance improvement between models as too little, then claim to know about a mysterious bottleneck to back this opinion.
Its the same every model release when the opposing side feels threatened.

sigh, thanks for... contributing to this thread

if only there would be more people like you who just ignorantly take things for what they are and mock people who try to understand and possibly improve things around them, this world would be a much better place... NOT

you might wanna check out afghanistan next vacation, youd find lots of friends over there for sure
fertilizers? god made plants grow fast enough as it is, there is no limitation slowing it down, impossible! only a blasphemer would doubt gods perfect creations! ati is great! err god i mean!

This got everything to do with buying the best product and overclocking it to the max . We are on xtremesystems here ..
Most people here want to get the most of of the product which they have bought . And not only that , but we want to learn more about this product and squeeze every last bit of perfomance out of it .

You seem to be one of those few people on here that is not interested too much in this material . But that's ok too

well said bro!

and yes, if your not interested in this thats fine, but why would you bash people and call them fanboys and idi0ts if they are curious how the tech they bought works and how to make it faster? this is THE essence of overclocking and xtremesystems!!! posting here, on THE overclocking, tweaking and pc technology forum worldwide mocking this is quite bizarre...

[saaya]

this might be interesting to benchmark! id love to test a 4870 vs a 5870 with half its blocks disabled to see how big of a diference this makes!

The computing cores can communicate on both local and global levels. ATI claims a considerable increase in cache bandwidth. Particularly, the speed of fetching data from the L1 cache is now as high as 1 terabyte per second while the bandwidth of the link between the L1 and L2 caches is increased to 435GBps. The L2 caches have become larger from 64 to 128KB.

xbitlabs keeps mentioning the reduced bandwidth "probably" causing some impacts at high resolutions compared to the 4870x2... they claim to have written this passage of the article before doing any tests, but i dont buy it
im pretty sure they KNEW what perf was like when they wrote this part...

anyways, if bandwidth is really the limiting factor, then why doesnt the 5870 drop behind the 4870x2 notably at high resolutions and with high anti aliasing modes, and why doesnt the 5870 perform notably better at low resolutions where the memory bandwidth doesnt matter?

yes, you CAN see this trend, its definately there, but its veeery subtle... we are talking about a few percent points here, nothing mayor you would expect from the 5870 having more than 30% less bandwidth than a 4870x2...

i think i found something...
rv870 is a REALLY massive chip with a huge load of raw processing power... keeping all those sps busy is not easy... and i think thats whats holding the 5870 back...



thread limit was only increased from 768 to 1024...
could this explain the limitation? 768 to 1024 is only a 30% boost... and a 5870 is roughly 40% faster than a 4890...

[hurleybird]

First of all, that's just the number of compute threads possible, and has nothing to do with limiting the shader power in games. Secondly, the 5870 can only run a maximum of 1600 / 5 = 320 threads, because you can only run one thread per group of 5 "stream processors", so in terms of DirectCompute even TriFire 5870's will not be held back by the thread limit, although QuadFire might be.

[Chumbucket843]

thx lightman for the infos!
you dont happen to have a 4890 or 4870 or better yet 4870x2 so you can do exactly the same thing on that card in the same rig?

i threw the numbers in a graph:
Pixel shader, multi fillrate, perlin noise, vertex shader simple are not bandwidth limited... yes some of these increase slightly, but the boost they get from increasing the memory clocks and hence bandwidth by almost 50% can be ignored...

complex vertex shader, shader particles and fillrate single are clearly bandwidth limited... but the gains are linear and proportional, so the gpu is definately not bandwidth crippled or starved like some people suggested... i dont think that memory bandwidth is limiting the 5870... i think its internal bandwidth or some internal architecture bottleneck...

they doubled registers and cache so internally it is fine.

[Chumbucket843]

i think i found something...
rv870 is a REALLY massive chip with a huge load of raw processing power... keeping all those sps busy is not easy... and i think thats whats holding the 5870 back...



thread limit was only increased from 768 to 1024...
could this explain the limitation? 768 to 1024 is only a 30% boost... and a 5870 is roughly 40% faster than a 4890...

it is held back by the compiler. i dont see how this would be any different from rv770 though so obviously scaling linearly with shaders isnt goint to happen for ATi unless they change something in their memory system.

[iandh]

I'm benching my unlocked 720BE against a guy with that has a 5870 and i7 over at OCN... results are less than exciting (for me at least)

I just ran Crysis Warhead bench 0.33 (avalanche flythrough)


1920 0xAA

i7 @ 4.2Ghz Min: 29.55 Max: 46.81 Avg: 37.04

Phenom II x4 @ 3.6Ghz Min: 25.12 Max: 36.73 Avg: 30.99


1920 8xAA

i7 @ 4.2Ghz Min: 21.29 Max: 37.74 Avg: 28.65

Phenom II x4 @ 3.6Ghz Min: 18.24 Max: 29.33 Avg: 23.83



note: Unlock from X3 to X4 gave me a slight avg fps increase, and almost 2fps increase in min fps.

It appears since the min FPS numbers are all very close, they are revealing spots where the bench is GPU limited, and the Max FPS numbers are quite different, showing spots where the bench is CPU limited

Anyone who owns a 5870 and wants to add benches into the mix is welcome... I'll run the same bench if possible on my system and we can see what can be worked out about this. OBVIOUSLY a 3.6Ghz Phenom II is no match for a 4.2Ghz i7, but I would have though that the results would have been much more GPU limited in this bench.

[jaredpace]

crysis likes cpu as much as gpu

[Gurr]

Not sure why you're less than excited. You're 600 mhz below him and only 5-10 FPS behind him in possibly the most intense game(graphically) there is. Still.

Nice results, are you 2 going to be doing any other tests together?