HD 5870 Discussion thread.

[a_ump]

Not to bring back the hot debate "memory bottleneck" but i was reading some architecture documents and come to think about something. Many of you and reviewers noticed a 5-10% increase in FPS when memory was OC'ed.

5870 = 256bits (1200Mhz for 153.60GB/s)
5870 = 256bits (1350Mhz for 172.80GB/s) <--- = 11.25% more GB/s (+- 5-10% perfs)

Now what if it had a 512bit bus wide?

5870 = 512bits? (1200Mhz for 307.20GB/s) <--- = 200% more GB/s
5870 = 512bits? (1350Mhz for 345.60GB/s)

Wouldn't it make sense to expect more than 5-10% increase in FPS? i think yes
I think the issue has more to deal with "memory adressable" than the possible speed it operate.

Anyhow another good page to read.. http://www.beyond3d.com/content/reviews/53/7

um, we've had 2 members here post on this page the benefits that memory overclocking certainly aren't significant enough to make it a bottleneck, it scaled the performance up as any memory oc would. If it were bottlenecked it would increase performance more than what it has. I too used to believe that the HD 5870 had to be bottlenecked but after seeing wolf and bobzilla post their results and both cases memory didn't increase performance significantly i've come to the conclusion that memory isn't really the issue. Though if you mean something else by memory accessible then could you explain in more detail

 

[Steevo]

The real world latency drops as the cycle time becomes faster.


5 wait states at 1300MHZ is going to be less than 5 wait states at 1000MHZ, if the core is runing 1,000Mhz (or 1Ghz) then it will lose five cycles every time it has to wait for data from the vmem @ 1,000 Mhz, but if those cycles are 30% faster (1300Mhz) then it effectively gains one or two cycles per fetch back in performance, so a horribly optomized game will benefit from faster data delivery during cache misses than the supposed higher bandwidth that is being generated.


So a 30% performance gain in branch prediction failure, from faster memory results in a real world performance of X, or XX.


Again, not bandwidth driven.

 

[Bo_Fox]

It's not just bandwidth that a game needs. Many games also benefit from lower latency cache/memory.

@ a_ump, it could be the hidden "upping" of latency that prevents any performance gains being seen from overclocking the memory. Many review sites are not aware of the error-correcting algorithm that if it's overclocked too much to the point where some errors are being produced, it would actually slow down the performance. It's such a new feature for a video card.

There's no doubting that a 5870 is absolutely starved for more bandwidth.

Upping the latency does hurt performance if it's increased way too much. 500MHz memory with CAS 2 latency is practically good as 1000MHz with CAS 4 latency in most applications. The world is not one-dimensional in that everything needs sheer bandwidth or buffer size or sheer speed with as low latency as possible. Some applications benefit more from increasing bandwidth. Some from increasing core processor clock. Some from increasing buffer size (for memory hungry apps).

1000MHz with CAS6 latency is usually slower than 500MHz with CAS2 latency for most applications. What we do not know for sure is that when we overclock the memory, does the driver automatically increase the latency more than we'd like? When some overclock their system memory just to see how high their clocks can get, they try to increase the latency by a couple notches, which really hurts performance.

There's just too much logical stuff pointing to how a 5870 just needs more bandwidth without losing too much latency.

I'm not saying this to "pwn" you guys. It's human to try to explore as much as possible from every angle of thought.

 

[Bo_Fox]

http://img.techpowerup.org/091110/6thsense.jpg

lol silly! You certainly read my mind!

Just joking around with you in this thread, ya know! Keep that in mind!!!

However... 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).

 

[dir_d]

Steevo i get what you are trying to say..When i read what you typed about the latency it reminded me of a Phenom 2 NB. The NB dosent need the extra Bandwidth of 1600 its fine at 1333 but it likes tighter timing and the NB to have fast clock speeds like 2.8Ghz to 3Ghz. So the memory bandwidth of the 5870 is fine but it probably likes lower latency which we cant control but we can control the core speed which helps it out far greater than increasing the bandwidth.

 

[newfellow]

That doesn't make any sense absolutely none. Neither does original message. Which is why Steevo probably posted the nice picture.

Higher the latencies higher the speed memory can go that is win win situation lower the latencies smaller OC capability, if you look at standard DDR2 for example 800Mhz CL4 is absolutely same as CL6 1066Mhz with difference that RAM internal copying speed is actually 2000MB/s faster while actual latency of CL4 800Mhz or CL6 1066Mhz stays exactly the same as base latency also goes down on higher speed from 2.50ns to 1.66ns.

 

[Steevo]

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.

 

[Bo_Fox]

Sadly, nobody really tested for this issue concerning a 5870 yet.

Only the engineers at ATI really know how much 512-bit bandwidth would benefit a 5870 without increasing latencies.

Many thought that an X1900XTX did not need more bandwidth (since we saw useless gains with overclocking the memory and greater gains with overclocking the core). However, ATI went ahead and did an X1950XTX with the only difference being ~30% greater bandwidth that yielded about 6-8% performance increase overall. It sold at a $100 price premium over an X1900XTX for several months, which helped ATI to rake in some money.

I do not think this is dumb at all. I do not read somewhat "dumb" posts thoroughly myself, but I definitely can see how some might classify me as "dumb" if my posts are not being read.


Logic 101

Prerequisite lesson:

Check the specifications of a 4890, 5770 and 5870. See how both a 4890 and a 5770 carry half the shader units, TMU's, and ROP's of a 5870--at the same clock speed.

Lesson 1:

A 4890 benefits with memory overclocked to 5870's bandwidth alone. I posted the evidence here in this thread.

Lesson 2:

A 4770 performs overall 20% worse than a 4890 despite slight architecture improvements (which still enabled it to beat a 4890 in 2 games out of like 20 games tested).

Lesson 3:

A 4770, with 5870's bandwidth, is 99.9% guaranteed to have at least a 26% performance increase.

Lesson 4:

A 5870 (2 times a 4770 core) is predicted to perform xx% better overall with 512-bit memory at the same latency.

I could post in theoretical benchmark numbers for a 5870 with 512-bit memory here if you want!

 

[newfellow]

Wow.

Just wow.


mebey I could draw a picture.

seems we need better picture.

plus couple more roads and no lanes. jumping from 2.50ns(ddr2-800 reference) to 0.70ns and back while store to buffer. o and thanks for Vista and 7 we actually do have to "a doesn't give a fuck if the road is a 12 lane, 2 lane or 1 lane." give a f.

 

[Steevo]

Your logic goes out the door when memory latentcy, GPU core steppings, and minor revisions ilke the 4890 had over the 4870 come into play.


We have done tests that show a increased bandwidth delivery, also a factor of memory speed and timings, have done nothing noteable with the 5XXX series. Why do we all strive for faster RAM? more bandwidth? Only on a few certain select applications, like media encoding, file compression, etc..... But more so to resolve a "fault" in less time. Lookup hard faults, page fault, cache and read up for the next year or so to understand.

 

[Bo_Fox]

Your logic goes out the door when memory latentcy, GPU core steppings, and minor revisions ilke the 4890 had over the 4870 come into play.


We have done tests that show a increased bandwidth delivery, also a factor of memory speed and timings, have done nothing noteable with the 5XXX series. Why do we all strive for faster RAM? more bandwidth? Only on a few certain select applications, like media encoding, file compression, etc..... But more so to resolve a "fault" in less time. Lookup hard faults, page fault, cache and read up for the next year or so to understand.

What does a 4870 have to do with it? Honestly? I dare you to explain!

 

[newfellow]

Your logic goes out the door when memory latentcy, GPU core steppings, and minor revisions ilke the 4890 had over the 4870 come into play.


We have done tests that show a increased bandwidth delivery, also a factor of memory speed and timings, have done nothing noteable with the 5XXX series. Why do we all strive for faster RAM? more bandwidth? Only on a few certain select applications, like media encoding, file compression, etc..... But more so to resolve a "fault" in less time. Lookup hard faults, page fault, cache and read up for the next year or so to understand.

Ok, I was about to add thanks on this one, but hell I just had to read it again. This doesn't make absolutely no sense either.

We do not correct errors with higher bandwidth we do need bandwidth more than that access time and a lot of applications including games does memory caching not to even consider 'shared memory' dedicated Video memory which is these days is more RAM as there's a lot more memory in system than faster GDDRx memory.

 

[Steevo]

The 5870 is not the same chip as was the 4890, in the same manner the 4870 is not the 4890.

Just like luck of the draw CPU steppings, one GPU with a different revision code will clock differently than another. Thy make minor revisions to the process all the time and test those as they go out to determine their exact effect on performance, yield, power consumption, and ohter aspects.


the 5870 is not a 4890 doubled. Not.

The addition of thousands of extra surfaces rendered by the tessellator in test show a performance hit, but not nearly as large as the hit we would see from trying to render those surfaces the same way DX10, 9 or any other does.


We have evolved beyound the point of needing to feed the GPU that much data.


Dare away, if i don't are you going to punch your monitor, or yell at me?


If the memory was the bottleneck, then the performance increase would rise withthe memory speed directly. It doesn't.

So the only other plauseable explination is the faster data delivery helps offset soft fault delays. Something a larger on die cache might help with, but only if the branch prediction is working at 100% accuracy, and that never happens.

 

[Bo_Fox]

Ok, I was about to add thanks on this one, but hell I just had to read it again. This doesn't make absolutely no sense either.

We do not correct errors with higher bandwidth we do need bandwidth more than that access time and a lot of applications including games does memory caching not to even consider 'shared memory' dedicated Video memory which is these days is more RAM as there's a lot more memory in system than faster GDDRx memory.

Let's keep it up like this! The tune's starting to sound nice.. just kidding!

 

[Bo_Fox]

The 5870 is not the same chip as was the 4890, in the same manner the 4870 is not the 4890.

Just like luck of the draw CPU steppings, one GPU with a different revision code will clock differently than another. Thy make minor revisions to the process all the time and test those as they go out to determine their exact effect on performance, yield, power consumption, and ohter aspects.


the 5870 is not a 4890 doubled. Not.

The addition of thousands of extra surfaces rendered by the tessellator in test show a performance hit, but not nearly as large as the hit we would see from trying to render those surfaces the same way DX10, 9 or any other does.


We have evolved beyound the point of needing to feed the GPU that much data.


Dare away, if i don't are you going to punch your monitor, or yell at me?


If the memory was the bottleneck, then the performance increase would rise withthe memory speed directly. It doesn't.

So the only other plauseable explination is the faster data delivery helps offset soft fault delays. Something a larger on die cache might help with, but only if the branch prediction is working at 100% accuracy, and that never happens.

Already mentioned "slight performance optimizations" with the 5770 over a 4890 numerous times in this thread.

slight performance optimizations: stuff that you regurgitate above

Once again, the performance never scales linearly with the bandwidth. It has never been the case for the past 10 years, but we always needed bandwidth. I'd gladly trade my 128-bit 5770 for your 256-bit 4890, even if I lose DX11 features. Would you trade it with me? Yes or no?

That's the million-dollar answer. It defies all things illogical.

 

[newfellow]

I'd gladly trade my 128-bit 5770 for your 256-bit 4890, even if I lose DX11 features.

Generally this is very interesting point.

I mean of course would be a good trade in all senses, if new GPU features like DirectX 11 would be meaningless (as explained above tesselation makes no difference what so ever to memory). Also considering that 4890 comes with so called (loaning a little) '8 lanes' where as 5770 comes in '4 lanes' (more ideal explainatory of course would be to compare 'dual channel'(128-bit) to 'single channel'(64-bit) (ddr2 reference) and how much better latency you could get from single channel, lol, which ain't much, but makes a lot on access). the 4890 would "thread" twice as fast and gain on memory speed while 5770 bandwidth would be completely dependable on the lower latency where the memory is at to have also faster access time with same internal memory speed with lower read/write external memory speeds which we cannot alter with, atm, softwares like computer BIOSes to alter usual RAM latencies.

but in theory 5770 could be as fast as 4890 memory interface with lower latencies.

 

[Bo_Fox]

Yes, very true. But 1000MHz with CAS4 is still better than 500Mhz with CAS2, although both might perform very similarly in most applications. There are a few scenarios that still benefit from increased bandwidth also.

Also, I doubt that a 5770's GDDR5 memory has lower latencies than a 4890. It's actually clocked higher than 4890's, but with 128-bit bus. If anything, the latency is probably identical (CAS 11 or so).

 

[newfellow]

Yes, very true. But 1000MHz with CAS4 is still better than 500Mhz with CAS2, although both might perform very similarly in most applications. There are a few scenarios that still benefit from increased bandwidth also.

Actually if you are using this as reference to how this would act in GDDRx as in graphical memory side. It would have huge benefit from so called 'Internal Memory Copy Speed' hundreds of times more/larger meaning in GPU world than than on RAM. Write and Read speeds in this sense could basically be what ever as long it has insane internal speed as most of the applications which require graphics does everything to keep the data on GPU memory instead of moving it to store on system RAM. (Edit: with exception of 'Microsoft Shared Video Memory'-scheme but 128-bit is still as fast as Dual Channel interface.)

Also, I doubt that a 5770's GDDR5 memory has lower latencies than a 4890. It's actually clocked higher than 4890's, but with 128-bit bus. If anything, the latency is probably identical (CAS 11 or so).

No, need to doubt there's no reason to. Why would they change the latency when they can sell more cards, heh.

 

[HalfAHertz]

Yes, very true. But 1000MHz with CAS4 is still better than 500Mhz with CAS2, although both might perform very similarly in most applications. There are a few scenarios that still benefit from increased bandwidth also.

Also, I doubt that a 5770's GDDR5 memory has lower latencies than a 4890. It's actually clocked higher than 4890's, but with 128-bit bus. If anything, the latency is probably identical (CAS 11 or so).

If video ram is the same as normal ddr ram, then yes

 

[ToTTenTranz]

If the memory was the bottleneck, then the performance increase would rise withthe memory speed directly. It doesn't.

Back this up with something real or your whole point is flawed.

 

[grimeleven]

um, we've had 2 members here post on this page the benefits that memory overclocking certainly aren't significant enough to make it a bottleneck, it scaled the performance up as any memory oc would. If it were bottlenecked it would increase performance more than what it has. I too used to believe that the HD 5870 had to be bottlenecked but after seeing wolf and bobzilla post their results and both cases memory didn't increase performance significantly i've come to the conclusion that memory isn't really the issue. Though if you mean something else by memory accessible then could you explain in more detail

What i meant is the memory channels, forgot to include it.

Look at the design, 4x64bit MC and 4xL2 caches compared to 8x64bit MC on the 512bit controller. Again i agree the difference in Mhz vs performance benefits doesn't justify any overclock, my point is about the data that can be fetched in and out, (like the way Intel has built their X-25 SSD, more memory channels than any other vendors).

 

[newfellow]

What i meant is the memory channels, forgot to include it. http://www.beyond3d.com/images/reviews/cypress-arch/cypress-arch-big.png
Look at the design, 4x64bit MC and 4xL2 caches compared to 8x64bit MC on the 512bit controller. Again i agree the difference in Mhz vs performance benefits doesn't justify any overclock, my point is about the data that can be fetched in and out, (like the way Intel has built their X-25 SSD, more memory channels than any other vendors).

This has some flaws in general when we come to how the data is output/input while there's so much data that it doesn't squeeze all in there (like 3D games streaming). The internal is insanely fast, but considering all the Intructions, all the core specific speed ups and on tops on that incredible memory speed none of this will make it 'a ok' when we are considering streaming 25GB BluRay(example) of textures like 'Rage' presented in the future. That data will still be buffered on X times slower RAM and that is the point where MS Shared memory fails & where every single one of us will have the bottleneck a head. Only way to prevent this is to have graphic design equal to entire shared video memory & very low fast DDR to close match the continuous stream.

Of course level based pre-loading would kill this idea, but that's way too slow technology when we look at data size of today as well as compression techniques on HDR/Textures & general graphical data are way more important than we think and stressful to process "in action". This is where the fast acting latencies & speed is killed away to get very very fast access time to the card sadly out mainboards, RAM nor CPU ain't exactly there yet.

 

[Steevo]

For a increase of 20% you would need to increase your memory speed by 70% of stock.

Notice how the lines have a nice trend, only about a 3% increase in performance for very 10% increase in memory speed.

Now that we all have a hard factual data based of actual figures from this thread. Let discuss what it means. Obviously faster memory does increase the card's performance. It does not follow the trend established by the memory for percent of increase. So then we must look elsewhere as to WHY the trend follows WHAT rule. The next one is branch prediction fault, and the effect of latency on recovery time. We find that since the bandwidth is not directly proportional to performance we infer that the next thing related to memory speed and data delivery that can explain the results is highly probable.

 

[newfellow]

For a increase of 20% you would need to increase your memory speed by 70% of stock.

At the system not the GPU. Games streams data no matter what the 'decided' footprint on GPU memory is it still goes from low latency system memory against high latency high speed GPU Memory which we cannot match no matter what we do. Buffers does solve it sure, but what when a single scene streamed data is twice the size of your GPU memory? (sure you can dump me something like Crysis only uses 900MB on tops, but there's a lot more than 900MB behind there waiting to be written to card).

Notice how the lines have a nice trend, only about a 3% increase in performance for very 10% increase in memory speed.

Look at your own picture 900/1250 and 950/1100 that speaks against this. Raising an GPU increased same amount of speed on your test -150Mhz on memory while memory speed increase did nothing as faster command rate. This is exactly the same as lower latencies which would actually even affect more, but since tighter latencies and lower speed decreases the internal speed only place you would actually see difference would be in while streaming enough data through system.

Now that we all have a hard factual data based of actual figures from this thread. Let discuss what it means. Obviously faster memory does increase the card's performance. It does not follow the trend established by the memory for percent of increase. So then we must look elsewhere as to WHY the trend follows WHAT rule. The next one is branch prediction fault, and the effect of latency on recovery time. We find that since the bandwidth is not directly proportional to performance we infer that the next thing related to memory speed and data delivery that can explain the results is highly probable.

We do have hard 'factual' data it simply doesn't make any sense as long we do not have an bottom line benchmark to do an individual streaming through to actually compare what does the new high-end compression algorithms do against streaming data through GDDRx and system-DDRx memory buffer. Faster memory did increase the performance because of lowered latency = faster access time that is exactly the 'Directly proportional to performance'. Not because there would of been speed behind it and considering very tight latencies against low speeds access time could be tweaked to very very similar levels on lower speeds than, atm, higher clocks with very small penalty of absolute top clock speed of memory or overclock ability.

Been tweaking so damn too many computers with so damn too many manufacturers memory chips and done tech support so too damn long that I think I can make argument latency against speed any time. We're still looking exactly same bottom line no matter what DRAM we speak of and specially on an card/GPU system where the threading cannot be performed in similar speeds as on even older cards.

 

[a_ump]

ah i see grimeleven.

as for bo_fox, i have alot of respect for you and your posts but your theoretical perforamnce increases just aren't real world...as was proven by wolf and bobzilla on the previous page. did you look at those posts? they keep the same core speed and increased the memory speed from 1200-1350, each 50mhz it only improved performance by .6-.9% for a grand total of ~3% performance increase after increasing the memory by 18%...that's not a bottleneck.

grimeleven's statement does make better sense though and i think there is some merit to it.