HD 5870 Discussion thread.

[Bo_Fox]

4890 --> 19% faster than 5770, 62% more bandwidth also

5770 --> slight R800 architecture optimizations, about ~3% boost

4890 --> 19% + ~3% faster than 5770 w/o R800 optimizations, TOTAL 22% increase


4890 GPU @ stock, mem overclocked to 5870's bandwidth --> ~3% extra boost

4890 GPU @ stock, mem overclocked to 5870's bandwidth w/o hidden latency increases --> ~5% boost, not 3%


5770 with same bandwidth as 4890 --> ~3% faster than 4890 thanks to R800 optimizations (22% boost overall)

5770 with bus doubled (same latency, 100% increase in bandwidth) --> 22% + 5% boost from further increase between 4890-->5870 bandwidth (= latency), TOTAL 27% boost!




5870, with each 128-bit 5770 "halves" boosted by 27% each (same 16 ROP's each) --> SAME 27% increase overall


Take that guys!!!


ATI, bring on that 512-bit!!!

Did you guys read this post? At all? It's simple math, really.

 

[Benetanegia]

Benetanegia, you cannot ignore the difference between a 5770 and a 4890, right in your face. It's not funny. What's the sole, main difference that sets the performance apart? (the bandwidth)

Please just respect the conclusion that I have arrived at. That's all I ask of you. It added up perfectly for me, so I only wanted to share it to you guys here. I showed 20+ benchmarks here to back it up. 1-2 others have posted only 1-2 benchies here.

If you cannot respect my opinion, please show some real evidence, with at least 10 real-world games at high resolutions where the CPU cannot be a bottleneck. Heartily.

We are not talking about the HD5770, we are talking about the HD5870, two different cards that belong to two different performance gaps. I have told you like a thousands times that you can't extrapolate the results. I could conceed, in fact I'm going to do it in order to preserve your sanity, that the HD5770 is limited by it's memory, but not going to say it is bottlenecked. The HD5870 is NOT, on the other hand. What you are trying to do with your comparisons is like saying that a Formula 1 would be able to reach 600 Km/h if they had 1500 HP instead of 750 HP, because a Civic with 150 HP is able to do 200 Km/h while a Civic with 75 HP is able to do 150 Km/h. It just doesn't correlate. Period.

If you cannot respect my opinion, please show some real evidence, with at least 10 real-world games at high resolutions where the CPU cannot be a bottleneck. Heartily.

You could, at least, read the links that I posted, because you have exactly what you are asking for. :shadedshu

Did you guys read this post? At all? It's simple math, really.

Sorry, but no, that's not math at all. That's just some random numbers and correlations that came out right from your ass, as I tried to say with some humor in my first reply to that post.

 

[Bo_Fox]

Look, I read into all of the linked webpages that you provided.

I understood them with sincerity.

This car analogy was cool, but the comparison is of a different kind, my cool friend.

It's extremely nice talking to you, Benetanegia.

What a cool name.

How did you come up with this?

6 vowels, wow! It's interesting!

 

[Benetanegia]

Look, I read into all of the linked webpages that you provided.

I understood them with sincerity.

Apparently not, since you are asking for what is there.

This car analogy was cool, but the comparison is of a different kind, my cool friend.

It's of the same kind, you are taking gains obtained in less performing parts and extrapolating them to greater performing ones, linearly. Specially when it comes to interfaces that is never the case. Inside the GPU with more or less fixed pipelines we can expect linearity, but outside of there in interfaces like memory and PCIe we can't, in those performance gains are rather logarithmic. Like this:

http://www.techpowerup.com/reviews/AMD/HD_5870_PCI-Express_Scaling/25.html

It's extremely nice talking to you, Benetanegia.

What a cool name.

How did you come up with this?

6 vowels, wow! It's interesting!

My nickname is in Basque and I appreciate you like it.

 

[Bo_Fox]

Basque.. hmm, I can't remember hearing of that!

Wow, NW Spain and SE France! So cool! 632,000 people had Basque as their native language! Awesome!

*edit* sorry for getting carried away from this thread!

Hey, I'm really tired right now, but I'll try to reply sincerely.. yes, i did read the websites that you linked to, with care, through what those pages and you were talking about. Also, hold on that car analogy, alrite, I'll discuss that with ya tomorrow, alrite. It's a valid "psychological" one, but I'll have to tell you the long story on why it's of a different kind.

The PCI-E 8x - 16x scaling is far, far, far, far, far, far, FARRRRRRRRRRRRRR less of a bottleneck than the 128-bit memory bottleneck on a 5770. It's a different kind of a bottleneck, in respect of the flexible semantics. Not the same kind of "logarithmic" that you just mentioned. It's a far, far flatter logarithm, sometimes actually the opposite at certain points when there's a sudden jump in the "release of bandwidth that is finally more than enough for a certain scenario." (You already know the following 2 sentences very well, duh me!, ---> Anyways, around the curve, a 5770 is half of the R870XT GPU, with half the bandwidth. Those differences are the most main and simple ones. Hey, am I being silly, lol? Ha!

What you doing tonight, dude? What's your favorite game that you would play right now? I'd play UT2004, which is where my avatar's from!

 

[Benetanegia]

Basque.. hmm, I can't remember hearing of that!

Wow, NW Spain and SE France! So cool! 632,000 people had Basque as their native language! Awesome!

*edit* sorry for getting carried away from this thread!

Hey, I'm really tired right now, but I'll try to reply sincerely.. yes, i did read the websites that you linked to, with care, through what those pages and you were talking about. Also, hold on that car analogy, alrite, I'll discuss that with ya tomorrow, alrite. It's a valid "psychological" one, but I'll have to tell you the long story on why it's of a different kind.

The PCI-E 8x - 16x scaling is far, far, far, far, far, far, FARRRRRRRRRRRRRR less of a bottleneck than the 128-bit memory bottleneck on a 5770. It's a different kind of a bottleneck, in respect of the flexible semantics. A 5770 is half of the R870XT GPU, with half the bandwidth. Those differences are the most main and simple ones.

What you doing tonight, dude? What's your favorite game that you would play right now? I'd play UT2004, which is where my avatar's from!

Dude, you are so insistent (overpost really) with the HD5770 that I was really really really starting to think that I've been reading the header wrong and this is the "HD 5770 below expectations?" thread, but no, I was right, it's the "HD 5870 below expectations?" thread, I had to take a look though.

I don't care how much the HD5770 is bottlenecked, I have not looked at it too much. I do think the HD5770 is somewhat limited by it's memory, but I have not seen tests on it.

The HD5870 is not bottlenecked.

 

[Bo_Fox]

Half of the 5870 (5770) is bottlenecked by half the bandwidth that the 5870 has.

Half of the 5870 still benefits from its bandwidth being increased to the total bandwidth that a 5870 has.

How can a 5870 still not benefit from additional bandwidth?

I do not mean to indulge you, but do you really want to provide (find) some scientifically justifiable and logical evidence, which, in my opinion, will be nearly impossible?

My friend. I'm gonna play some UT2004 now! Take care and g'nite!

 

[HalfAHertz]

Half of the 5870 (5770) is bottlenecked by half the bandwidth that the 5870 has.

Half of the 5870 still benefits from its bandwidth being increased to the total bandwidth that a 5870 has.

How can a 5870 still not benefit from additional bandwidth?

I do not mean to indulge you, but do you really want to provide (find) some scientifically justifiable and logical evidence, which, in my opinion, will be nearly impossible?

My friend. I'm gonna play some UT2004 now! Take care and g'nite!

Dang dude, it's all about bandwith with you! But hey the masses agree, size does matter, right?

Yes you will see an increase in performance if you increase the bus, but ,BUT it will not be a 100% increase, nor a 50%, nor even a 25% increase, thus rendering it inefficient, expensive and useless for the consumer.

 

[r9]

Wow.

Just wow.


mebey I could draw a picture.


a------------------------>B

a needs to get to B data location to be processed.
a doesn't give a fuck if the road is a 12 lane, 2 lane or 1 lane.
B doesn't either.

if each line represents 1ns and a is forced to drive at 1ns per second it will take X numer of seconds get there. The lines represent a unchangin timing value, not a error checking value, not a changed value. Thus the reason we can't/shouldnt mess with timings on GPU's since, DDR.


So a is driving along happily at one dah per ns, untill we give him a NOS bottle. then it becomes one dash every .70ns same number of dashes, still doesn't ive a flying fuck about the width of the highway. B is still waiting.

B runs at one cycle per ns, ever ns a is late results in a performance hit. There could be a 12 lane freeway coming to B's house and it still woudn't stop the fact that a has to drive at the same speed. not one fucking bit.


The sooner a arrives, the sooner the epic plans of their consort can continue.

So today we have learned that the faster a drives (faster memory speed) the faster a branch prediction fault can be recovered, and normal operation can resume. Again.


a------------------>B
--------------------
--------------------
--------------------
--------------------
--------------------
--------------------
--------------------
--------------------
--------------------
--------------------
--------------------


all wasted, jsut wating on little a.

The thing is that games need more than just "a" to be processed in any given moment more like chinese alphabet.
What if there are 20 packages to be delivered in every given moment.
If there is one lane and 20 packages and the time of travel is 1 second per package and there is only one lane you going to waste 20 seconds for 20 packages.
What about 20 slower lanes about 2 sec from A to B your are going to waste just 2 seconds.

So the width of the memory link is as important as the memory speed.

1------------------------------------------2sec
2------------------------------------------2sec
3------------------------------------------2sec
4------------------------------------------2sec
5------------------------------------------2sec
6------------------------------------------2sec
7------------------------------------------2sec
8------------------------------------------2sec
9------------------------------------------2sec
10------------------------------------------2sec
11------------------------------------------2sec
12------------------------------------------2sec

 

[Benetanegia]

Half of the 5870 (5770) is bottlenecked by half the bandwidth that the 5870 has.

Half of the 5870 still benefits from its bandwidth being increased to the total bandwidth that a 5870 has.

How can a 5870 still not benefit from additional bandwidth?

I do not mean to indulge you, but do you really want to provide (find) some scientifically justifiable and logical evidence, which, in my opinion, will be nearly impossible?

My friend. I'm gonna play some UT2004 now! Take care and g'nite!

A little bit of history:

9800 XT --> 23.36 GB/s ---> baseline performance --> baseline memory bandwidth
X850XT PE--> 37.76 GB/s --> 2x performance --> 1.6x memory bandwidth
X1900XTX --> 49.6 GB/s --> 4x performance --> 2.12x bandwidth
X1950XTX --> 64 GB/s ---> 4.2x card performance --> 2.75x more bandwidth
HD3870 --> 57.6/72 GB/s (GDDR3/4) --> 7x card performance --> 2.46/3x more bandwidth
HD4870----> 115.2 GB/s ---> 16x card performance --> 5x bandwidth
HD5870----> 153.6 GB/s ---> 28x card performance --> 6.6x bandwidth

If you are unable to see the trend there, the fact that double the bandwidth is NOT necessary for doubling the performance, I don't know what else to tell you to "convince" you of what everybody else knows very well.

The HD5870 is NOT 2 HD5770 put together, every half of the chip doesn't access half of the memory or has access to half the memory bandwidth. The whole chip accesses the whole memory and doesn't require double the bandwidth at all. It's about time you understand this simple concept.

* If you want to do the same with Nvidia, you can: it would start off with 6800 Ultra's 35 GB/s and end up at GTX285's 160 GB/s or 4.57x bandwidth for around a 18x increase in performance.

 

[wolf]

This thread gets more and more interesting with every page, good postings by all!

 

[Bo_Fox]

Hmm, that's cool, wolf, thanks for letting me know that it's an interesting discussion. I didnt know if it was rubbing people the wrong way or not. That's good to know.

The thing is that games need more than just "a" to be processed in any given moment more like chinese alphabet.
What if there are 20 packages to be delivered in every given moment.
If there is one lane and 20 packages and the time of travel is 1 second per package and there is only one lane you going to waste 20 seconds for 20 packages.
What about 20 slower lanes about 2 sec from A to B your are going to waste just 2 seconds.

So the width of the memory link is as important as the memory speed.

1------------------------------------------2sec
2------------------------------------------2sec
3------------------------------------------2sec
4------------------------------------------2sec
5------------------------------------------2sec
6------------------------------------------2sec
7------------------------------------------2sec
8------------------------------------------2sec
9------------------------------------------2sec
10------------------------------------------2sec
11------------------------------------------2sec
12------------------------------------------2sec

Nicely said. You're right, it's not just the bandwidth but speed also.

A little bit of history:

9800 XT --> 23.36 GB/s ---> baseline performance --> baseline memory bandwidth
X850XT PE--> 37.76 GB/s --> 2x performance --> 1.6x memory bandwidth
X1900XTX --> 49.6 GB/s --> 4x performance --> 2.12x bandwidth
X1950XTX --> 64 GB/s ---> 4.2x card performance --> 2.75x more bandwidth
HD3870 --> 57.6/72 GB/s (GDDR3/4) --> 7x card performance --> 2.46/3x more bandwidth
HD4870----> 115.2 GB/s ---> 16x card performance --> 5x bandwidth
HD5870----> 153.6 GB/s ---> 28x card performance --> 6.6x bandwidth

If you are unable to see the trend there, the fact that double the bandwidth is NOT necessary for doubling the performance, I don't know what else to tell you to "convince" you of what everybody else knows very well.

The HD5870 is NOT 2 HD5770 put together, every half of the chip doesn't access half of the memory or has access to half the memory bandwidth. The whole chip accesses the whole memory and doesn't require double the bandwidth at all. It's about time you understand this simple concept.

* If you want to do the same with Nvidia, you can: it would start off with 6800 Ultra's 35 GB/s and end up at GTX285's 160 GB/s or 4.57x bandwidth for around a 18x increase in performance.

That's a nice glimpse at history, thanks. Fun data.. I like this!

However, there continues to be benefits from increased bandwidth/speed. You're leaving out "speed", which is the latency associated with clock. It progressively got better over the years, just like CAS 3 DDR-500 and then now we have CAS 6 DDR3 2000, which is 4 times the speed with only twice the latency. That's 2x the speed that you just left out right there.

A 4890 benefited with its memory overclocked to 5870 speeds, while its GPU core was being kept @ stock. The results were consistent with different games.

How can you really believe that a 5870 will not benefit at all from increased bandwidth? Even half of a 5870's core would benefit from having the full 256-bit bandwidth of a 5870 all to its own (to each half).

It's ok if you want to disagree. We could peacefully agree to disagree, dude.

Maybe you'd be willing to buy a cut-down 128-bit version of a 5870 for $50 cheaper. I wouldnt even buy it for $100 cheaper.

I'd be willing to buy a 512-bit version for $100 more. You might think I'm dumb for doing that, but for 100% more bandwidth, hell yea! People paid $100 more for an X1950XTX over the X1900XTX, and it had only 29% greater bandwidth (which was the ONLY difference...same GPU clock), believe that or not. I thought that was retarded, but for 100% more bandwidth, I'd pay $100+.

ATI would have a super-high-end card, and a "lower" high-end card that might be cheaper to make than a 5850. Also, it'd be nice if ATI could release a 256-bit of 5770, for the same price of a 4890. It'd sell even hotter than a 6600GT from the Doom3 days.

 

[a_ump]

+1 to wolf's discussion, i check this thread every hr or so when at home to see if it has a new post. Though i must the cas latency in the memory chips is beyond me and i don't think there's anyway to know the cas settings for the memory chips is htere? some of us are just assuming.

 

[Benetanegia]

How can you really believe that a 5870 will not benefit at all from increased bandwidth? Even half of a 5870's core would benefit from having the full 256-bit bandwidth of a 5870 all to its own (to each half).

I believe it because overclocking and downclocking the memory on the actual card has demostrated that it doesn't affect too much, no more no less. It would benefit a little bit? Yes a little bit. And so would a GTX285, an HD4870, a GT240 or an Hd3650... you get the idea. All cards can benefit.

Maybe you'd be willing to buy a cut-down 128-bit version of a 5870 for $50 cheaper. I wouldnt even buy it for $100 cheaper.

No, I wouldn't buy a cut-down version of a HD5870, that would be stupid to release a card with such a chip and such a limited memory bandwidth (more so to buy it), BUT that is not the case. It's even more stupid to release a card that has a lot of bandwidth that is never going to be used. The answer: a balanced card.

I'd be willing to buy a 512-bit version for $100 more. You might think I'm dumb for doing that, but for 100% more bandwidth, hell yea! People paid $100 more for an X1950XTX over the X1900XTX, and it had only 29% greater bandwidth (which was the ONLY difference...same GPU clock), believe that or not. I thought that was retarded, but for 100% more bandwidth, I'd pay $100+.

We've been trying to tell you that despite the bandwidth being 100% more, 1000% more or 1 trillion% more you would not see a huge improvement. The HD5870 would benefit from a 100% increase as much as the X1950XTX did from that 29% increase, necause it would have moved into diminishing returns much much earlier than both +100% and 29% as the tests performaed in the HD5870 have demostrated. I don't judge people's intelligence based on how they spend their money, so I'll let that in your hands to decide if it would be retarded to buy a 512 bit HD5870 for $100 more, when you could probably buy an X2 for such an ammount, or a Nvidia card that will probably be significantly faster than a HD5870, 512bit or not. For $100 (even 50$) more and considering they would have to completely redesign to make a 512 bit HD5870, AMD could as well design a 2400 SP, 384 bit card that would crush any HD5870 512 bit, they would be designing basically an "AMD Fermi". But costs wise the 512 bit HD5870 would be no different, so the "AMD Fermi" makes much much more sense, at least you would have the performace improvement you wouldn't obtain the other way. Why not do that? Because by not doing that they've been first to market by months advantage and that is what AMD needs right now.

 

[Bo_Fox]

That is ok if you think my hypothesis is wrong. I meant that each half would benefit a lot (27% predicted). You could as well as say that it's 27% for the first half, and half of 27% (13%) for the second half of the 5870 core, if you think that it's THAT diminishing in returns (far, far worse than SLI/Crossfire scaling with two separate halves, or 4890's).

The test in which you did not see any performance difference with 5870's memory downclocked or overclocked was a theoretical test, to remind you once again. It's basically as theoretical as the TFLOPS performance rating. Also, the Unigine benchmark with the memory potentially overclocked into "error-correcting slow-down territory" is still only one test. One test alone is worth nothing in statistics when trying to prove a scientific hypothesis.

I stand firm on the ground with this stance that a 512-bit bandwidth for a 5870 would yield a ~27% improvement overall.

It's similar to what r9 has stated above. I quoted his post, because it does have logical credibility. It's like each SIMD engine (80 stream processors per engine pipeline) scaling along without diminishing returns on each pipeline if there are more. The total memory bandwidth has to feed all of the shader units and TMU's (a massively multi-core processor). It needs far, far, far more bandwidth than the multi-threaded CPU. A 5870 is "balanced" enough that it is 60% faster than a 4870 in order to be the fastest card today, but when there's competition, the engineers at ATI might need to use their 512-bit experience after doing it with an HD 2900XT (which was done out of desperation though, with only 320 shaders that could hardly handle AA all on its own).

Nvidia will most likely be feeding their Fermi GT300 with 50% greater bandwidth than that of a GTX 285. By comparison, a 5870 has only 23% more than a 4890. A gentle reminder: A 4890 still did better with its memory alone overclocked to 5870's speed. ---> A 5870 with 2x the stuff of a 4890 is likely to benefit far more from 100% increase than an X1950XTX did from that 29% increase that you talked about.

People talked up quite a storm about how 1GB was a waste for the 4870. This heated discussion lasted for months, until ATI finally released a 1GB version. There were pages and pages of threads on many forums about how 512MB was good enough. Once the 1GB version was released, the truth became clear. Yes, a few did complain about the lack of the 1GB version at first (including me), but there were so many more people arguing against those complaints, about why 1GB was unnecessary. It seems that for each person who wants an improvement, there is 3 people to argue against it.

Hey, I peacefully disagree with you, Benetanegia. Truce.

 

[HalfAHertz]

Ok what if the slow-down is not in the bus itself, but right after it? The old generation had 4TMUs and 4ROPs connected to each 64bits, for a total of 16 correct? The 5000s have twice that while still keeping the same internal connection. You can argue that doubling the bus will correct that, however if the above case is indeed true, woudn't it be cheaper to just increase the internal bandwith between those components?

 

[Benetanegia]

That is ok if you think my hypothesis is wrong. I meant that each half would benefit a lot (27% predicted). You could as well as say that it's 27% for the first half, and half of 27% (13%) for the second half of the 5870 core, if you think that it's THAT diminishing in returns (far, far worse than SLI/Crossfire scaling with two separate halves, or 4890's).

The test in which you did not see any performance difference with 5870's memory downclocked or overclocked was a theoretical test, to remind you once again. It's basically as theoretical as the TFLOPS performance rating. Also, the Unigine benchmark with the memory potentially overclocked into "error-correcting slow-down territory" is still only one test. One test alone is worth nothing in statistics when trying to prove a scientific hypothesis.

I stand firm on the ground with this stance that a 512-bit bandwidth for a 5870 would yield a ~27% improvement overall.

It's similar to what r9 has stated above. I quoted his post, because it does have logical credibility. It's like each SIMD engine (80 stream processors per engine pipeline) scaling along without diminishing returns on each pipeline if there are more. The total memory bandwidth has to feed all of the shader units and TMU's (a massively multi-core processor). It needs far, far, far more bandwidth than the multi-threaded CPU. A 5870 is "balanced" enough that it is 60% faster than a 4870 in order to be the fastest card today, but when there's competition, the engineers at ATI might need to use their 512-bit experience after doing it with an HD 2900XT (which was done out of desperation though, with only 320 shaders that could hardly handle AA all on its own).

Nvidia will most likely be feeding their Fermi GT300 with 50% greater bandwidth than that of a GTX 285. By comparison, a 5870 has only 23% more than a 4890. A gentle reminder: A 4890 still did better with its memory alone overclocked to 5870's speed. ---> A 5870 with 2x the stuff of a 4890 is likely to benefit far more from 100% increase than an X1950XTX did from that 29% increase that you talked about.

People talked up quite a storm about how 1GB was a waste for the 4870. This heated discussion lasted for months, until ATI finally released a 1GB version. There were pages and pages of threads on many forums about how 512MB was good enough. Once the 1GB version was released, the truth became clear. Yes, a few did complain about the lack of the 1GB version at first (including me), but there were so many more people arguing against those complaints, about why 1GB was unnecessary. It seems that for each person who wants an improvement, there is 3 people to argue against it.

Hey, I peacefully disagree with you, Benetanegia. Truce.

I'm tired of the conversation so I'll do it easier for me by explaining it by points:

1. You can't predict the performance increase on each half in the way you are doing. Period.
2. Increasing the memory clock a 10% only increased performance by 2% and the gains are NOT linear at all. Increasing it a 100% would increase performance by a maximum of 8% (predicted), 8% that's it. I don't know from where are you taking that 27% figure but it simply won't ever happen. Diminishing returns is that the first 10% increases perf by 2% (as we see in reviews), the next 20% increases another 2%, the next 30% another 2% and the last 40% another 2% IF AT ALL. Total: 8%.
3. In order to achieve an 8% imprvement is much better to increase the core clock and memory clock by 10%, much cheaper too, but still probably the HD5870 clocks were the best choice, due to yields, thermals, power...
4. 512 MB to 1 GB is a very different thing. It depends on if data has to be moved from and to main memory or not.

I'm peacefully disagreaing with you too man, but I have to reply, because you are simply taking numbers out from your ass constantly, numbers that defy not only logic*, but the actual data we are seeing in reviews. A performance increase greater than 10% will just never happen from doubling the bandwidth and that's it.

* If a 10% memory OC gives 2% increase, a 100% will yield an absolute maximum of 20%, which represents linear scaling. And it will never happen.

 

[wolf]

Hmm, that's cool, wolf, thanks for letting me know that it's an interesting discussion. I didnt know if it was rubbing people the wrong way or not. That's good to know

It's awesomely interesting, I'll go as far as to say easily top 5 most interesting threads on TPU that I have read and been a part of. And you certainly aren't rubbing me the wrong way

+1 to wolf's discussion, i check this thread every hr or so when at home to see if it has a new post.

It's awesome ain't it.

2. Increasing the memory clock a 10% only increased performance by 2% and the gains are NOT linear at all. Increasing it a 100% would increase performance by a maximum of 8% (predicted), 8% that's it. I don't know from where are you taking that 27% figure but it simply won't ever happen. Diminishing returns is that the first 10% increases perf by 2% (as we see in reviews), the next 20% increases another 2%, the next 30% another 2% and the last 40% another 2% IF AT ALL. Total: 8%.

Hmmm, See changing the memory clock to the value of 30% in my testing showed a delta of roughly 9% performance in multiple engines (1000mhz to 1300mhz) So I see value in predicting 100% more, for instance 2400mhz GDDR5, or 1200mhz across a 512-bit-bus, may indeed yield 27% , my guess is closer to 20%, but I think its a solid prediction based on testing performed thus far.

3. In order to achieve an 8% imprvement is much better to increase the core clock and memory clock by 10%, much cheaper too, but still probably the HD5870 clocks were the best choice, due to yields, thermals, power...

No argument there, boost memory from 1200 to 1350, and core from 850 to 925-950 and BAM there's your 8-10% I reckon.

Keep up the awesome thread guys

 

[Benetanegia]

Hmmm, See changing the memory clock to the value of 30% in my testing showed a delta of roughly 9% performance in multiple engines (1000mhz to 1300mhz) So I see value in predicting 100% more, for instance 2400mhz GDDR5, or 1200mhz across a 512-bit-bus, may indeed yield 27% , my guess is closer to 20%, but I think its a solid prediction based on testing performed thus far.

Yeah, but that is based on change from downclocking or moving below the hot spot. Remember that performance boost based on an external interface has logarithmic progression, moving to lower values will always yield a bigger change than moving to higher ones.

https://upload.wikimedia.org/wikipedia/commons/7/73/Logarithms.svg

It's important to note that you had the core at 950 Mhz 12% OC and that changes the hot spot of the memory bandwidth too. You would basically need to match the memory OC to the core OC to obtain the same balance around the hot spot.

 

[wolf]

Yeah, but that is based on change from downclocking or moving below the hot spot. Remember that performance boost based on an external interface has logarithmic progression, moving to lower values will always yield a bigger change than moving to higher ones.

https://upload.wikimedia.org/wikipedia/commons/7/73/Logarithms.svg

That's a good point and I do think they tried to hit the 'hot spot' pretty darn close, but your'e pretty certain on 8%? I'd still gamble at more than that, perhaps not 27%, maybe not even 20%, but 8% seems really low to me....

 

[Mussels]

if your design has a bottleneck, you get linear progression as you up computational power/bandwidth.

Once you get past that bottleneck, you get diminishing returns - only exceptions (EG, a program that uses more memory than most) will show benefits.


The question shouldnt be silly arguments about theoretical bandwidth limitations, but rather specific discussions on: does it limit DX9 games, does it limit DX10 games, will it limit DX11 games

If none of those hold true - then who gives a shit?

 

[wolf]

if your design has a bottleneck, you get linear progression as you up computational power/bandwidth.

Once you get past that bottleneck, you get diminishing returns - only exceptions (EG, a program that uses more memory than most) will show benefits.


The question shouldnt be silly arguments about theoretical bandwidth limitations, but rather specific discussions on: does it limit DX9 games, does it limit DX10 games, will it limit DX11 games

If none of those hold true - then who gives a shit?

See I'm really not sure about the whole bottleneck thing, after my own testing I'm still left scratching my head a little, in terms of interpreting the results.

On one hand it seems like they could-have clocked the memory even lower than 1000mhz and still not really 'bottlenecked' the card at that point, but there are consistant gains to be had moving in 50mhz increments from 1000 to 1300.

As for the games side, I am super duper keen to test core/mem speed variations on Dirt2, as soon as it's released as I got a copy on steam with the card.

 

[HalfAHertz]

See I'm really not sure about the whole bottleneck thing, after my own testing I'm still left scratching my head a little, in terms of interpreting the results.

On one hand it seems like they could-have clocked the memory even lower than 1000mhz and still not really 'bottlenecked' the card at that point, but there are consistant gains to be had moving in 50mhz increments from 1000 to 1300.

As for the games side, I am super duper keen to test core/mem speed variations on Dirt2, as soon as it's released as I got a copy on steam with the card.

There are temporary alternatives avaible like the free DX11 enabled Battleforge from EA.
Or if you feel like shelling out, you could get the russian version of the new S.T.A.L.K.E.R. from e-bay. It goes for 20 to 30$ and you can get the english language files from the mod sites...

 

[Benetanegia]

That's a good point and I do think they tried to hit the 'hot spot' pretty darn close, but your'e pretty certain on 8%? I'd still gamble at more than that, perhaps not 27%, maybe not even 20%, but 8% seems really low to me....

Well of course my number is not a hard fact at all, it's my estimation based on ALL the hints that I see in all the reviews and tests that I posted earlier, but don't think it's going to be closer to 20% than 8%. I'll try to explain my point even further with graphs and all, if I have time later. (We are working remodeling and repainting my house so I don't know if I'll have time.)

 

[Steevo]

I already did a graph, it was ignored by te bandwidth holy rollers. I and many others would like to discuss real things with these cards and find performance improvements, however the constant interjection about bandwidth ruined it for me.



Ever notice how many people after looking through review after review, and even Benetanegia (a green man) is willing to concede they are not bandwidth limited to any real degree. I did an chart, and yes if the gain were to remain the same it might net your a 20% at best boost in FPS. But not for everything, and most likely not due to the bandwidth increase.


I was goign to post the whitepaper for people to read, but the point of it is moot, as it doesn't seem it would be read or understood.


So please don't reply, I am not coming back to this thread, or any like it unless I see a solid proof that someone has gained a large amount of performance. I'm not evein going to bother sharing any of what I have found-learned as it would take a longtime and waste my time for someone who has blinded themselves.