Overclocked HBM? It's true, and it's fast

[mirakul]

Well, at least you remembered that this is a topic about HBM. Two big paragraphs talking about nVidia really tired me out. And you seems to forget that I mentioned "memory intensive scenario".

You are right about HBM. The benefit of OC mostly lies in latency improvement due to low clock. The performance gain is real, and will become significant when it comes near to 600MHz zone.

More example of HBM overclocking
Stock Memory: http://i.imgur.com/mVQU5AL.png

+10MHz Memory: http://i.imgur.com/cKL2J4i.png

+50MHz Memory: http://i.imgur.com/vhxRF3z.jpg

 

[RejZoR]

So, as it seems, massive bus width at very slow clock (R9 Fury X) isn't as beneficial as moderate bus width with ultra fast clock (GTX 980Ti)...

 

[Aquinus]

Well, at least you remembered that this is a topic about HBM. Two big paragraphs talking about nVidia really tired me out. And you seems to forget that I mentioned "memory intensive scenario".

You are right about HBM. The benefit of OC mostly lies in latency improvement due to low clock. The performance gain is real, and will become significant when it comes near to 600MHz zone.

More example of HBM overclocking
Stock Memory: http://i.imgur.com/mVQU5AL.png

+10MHz Memory: http://i.imgur.com/cKL2J4i.png

+50MHz Memory: http://i.imgur.com/vhxRF3z.jpg

So lets think about this. At stock you have a score of 14250 and overclocked (50Mhz) you get a score of 14423. So you have a 1% overclock that yields a 0.012% increase in score. Assuming that 0.012% isn't simply error (which most scientists would say it is,) a 10% overclock on memory would get you a theoretically (and error prone,) result of 0.1% improvement. The problem is that latency probably isn't the problem. Much like modern CPUs, caching algorithms have got pretty damn good and hit rates are higher than they've ever been. So I suspect that memory bandwidth and memory starvation are not the problems. In fact I think the IMC can feed the cores too quickly and the GPU is simply memory biased. Considering the difference in peak power and average power, I suspect most of the shaders on Fiji have been going vastly underutilized because the ROPs aren't pumping frames out fast enough. There is more than enough memory bandwidth to be had. The problem, to me, seems to be that the GPU can't render frames fast enough, which isn't a shortcoming of the shaders, but probably the bad ratio of shaders to ROPs. I suspect Fiji will shine when most of the memory is actively getting used when shaders have a little more work to do and rasterizing the frame becomes less of a burden to the GPU.

Now, most of that is speculation but, what isn't, is that Fiji is clearly not memory starved and the links you provided kind of point to that as being the case.

 

[MrGenius]

Higher bandwidth doesn't necessarily provide higher performance(depends on GDDR vs. HBM mainly). Whereas higher memory clock speed(and/or lower latency) always does(regardless of GDDR vs. HBM). Whether or not the bandwidth is consumed, or used to increase performance, depends on various factors(most are not important to this conversation, with the exceptions of its relevance to the bus width of HBM and which resolution is being displayed). However, whether or not the memory speed(and/or latency) is directly correlative to performance is a really stupid question to ask. Of course it is(just like the core speed is). HBM changes the game, but not that much. Running HBM at higher speed will always yield higher performance(just as it would with GDDR). And increasing the memory bus width is essentially the same as increasing the memory speed(regardless of GDDR vs. HBM). Which is why HBM at relatively low clock speeds utilized across a MUCH wider bus performs as well as it actually does. A 4096 bit bus is essentially 10.66 times faster than a 384 bit bus. Lower speed times a wider bus divided by .66 the amount of GB available still gives you relatively more bandwidth too(if we're still on the Fury X vs. 980 Ti debate). Which does not always, as previously stated, translate into higher performance. Hence the 980 Ti performing better than the Fury X at less bandwidth consuming resolutions(albeit with significantly lower bandwidth available). Simple stuff if you comprehend the specs and know the math involved(basic addition, subtraction, multiplication and division).

Amazing to me the amount of disinformation in this thread(or in threads on this site in general). Wrong, wrong, wrong. Just plain wrong. Way off base. Complete misunderstanding for the most part.

 

[15th Warlock]

With HBM's bandwidth being so high I don't think you'll benefit much in real games and apps from OCing the memory for this card, memory bandwidth doesn't seem to be an issue for the Fury at the moment.

Chances are it'll be a while before you see any games or apps top 512GBs of memory bandwidth and perceive a tangible benefit from OCing the memory for this card.

 

[EarthDog]

So lets think about this. At stock you have a score of 14250 and overclocked (50Mhz) you get a score of 14423. So you have a 1% overclock that yields a 0.012% increase in score. Assuming that 0.012% isn't simply error (which most scientists would say it is,) a 10% overclock on memory would get you a theoretically (and error prone,) result of 0.1% improvement. The problem is that latency probably isn't the problem. Much like modern CPUs, caching algorithms have got pretty damn good and hit rates are higher than they've ever been. So I suspect that memory bandwidth and memory starvation are not the problems. In fact I think the IMC can feed the cores too quickly and the GPU is simply memory biased. Considering the difference in peak power and average power, I suspect most of the shaders on Fiji have been going vastly underutilized because the ROPs aren't pumping frames out fast enough. There is more than enough memory bandwidth to be had. The problem, to me, seems to be that the GPU can't render frames fast enough, which isn't a shortcoming of the shaders, but probably the bad ratio of shaders to ROPs. I suspect Fiji will shine when most of the memory is actively getting used when shaders have a little more work to do and rasterizing the frame becomes less of a burden to the GPU.

Now, most of that is speculation but, what isn't, is that Fiji is clearly not memory starved and the links you provided kind of point to that as being the case.

That is a 1.2% gain in score...

...and 50Mhz is a 10% overclock...

.... your point still remains though!!!


Outside of some math, my question would be, what would the scores of a 980ti do when adding 10% to its memory clocks?

 

[mirakul]

As I said, HBM overclock is more about latency. I don't know why some of you keep talking about bandwidth.

And as I said (again), latency improvement must be coordinated with memory timing profile in BIOS. Too far from a timing profile yields more errors and mitigate performance gain, such as in 550MHz case. At 600MHz however, it was a 20% jump in fs graphic score, from 16k stock to 19k oc.

This is another example of 600MHz with significant performance gain.

[quote name="Neon Lights" url="[URL]http://www.overclock.net/t/1547314/official-amd-r9-radeon-fury-nano-x-x2-fiji-owners-club/1720#post_24106947[/URL]"]
I also had the memory bug.

In the "Furry and Tessy" Test (1920x1080, 4xMSAA) in MSI Kombustor 2.5.0 600MHz memory clock (and standard core clock) gives me 57FPS instead of 49FPS[/QUOTE]

 

[LightningJR]

At 600MHz however, it was a 20% jump in fs graphic score, from 16k stock to 19k oc.

[/QUOTE]

And again show us where you got these numbers. I looked and didn't find them, you need to stop making stuff up.

I have never saw such a hardcore fanboy before, the manipulation of facts is unbelievable with you. I really hope this is not how you conduct yourself in the real world.


"In the "Furry and Tessy" Test (1920x1080, 4xMSAA) in MSI Kombustor 2.5.0 600MHz memory clock (and standard core clock) gives me 57FPS instead of 49FPS"

You're basing the FPS gains from VRAM overclocking on a stress test? You know why there's no tests from games using the memory overclock? Because the FPS increase is so low it's not even worth reporting.

 

[dorsetknob]

I have never saw such a hardcore fanboy before, the manipulation of facts is unbelievable with you. I really hope this is not how you conduct yourself in the real world

You have a very short memory Like beetlejuice you dont say his name 3 times in a row but there was

Spoiler: Beetlejuice Beetlejuice Beetlejuice

@xfia

]

 

[mastrdrver]

Reading that thread over there (oc.net), I think it's possible that AMD might be completely rewriting their driver for W10. It would explain why there has been no official driver for so many months. Evidence to support this idea can be seen from Project CARS where people using it on W10 are seeing something on the order of at least a 20% increase in performance.

As always time will tell, but there may be a lot of performance locked up in the driver as the little increases in performance from overclocks seem odd for such a card (even on the core).

 

[Folterknecht]

Stop using Firestrike as an argument when it comes to OC or performance - its fucking useless. Who plays that shit?

OC to the max stable values, run some games and post min + avg FPS numbers.

 

[Mussels]

Reading that thread over there (oc.net), I think it's possible that AMD might be completely rewriting their driver for W10. It would explain why there has been no official driver for so many months. Evidence to support this idea can be seen from Project CARS where people using it on W10 are seeing something on the order of at least a 20% increase in performance.

As always time will tell, but there may be a lot of performance locked up in the driver as the little increases in performance from overclocks seem odd for such a card (even on the core).

i'm seeing better performance in win 10 on the insider edition with the forced windows update drivers than i did in 8.1, even in DX9 games. Feels like some CPU limitations have been improved (better multithreading?)

that said, its going a little off topic.

 

[Pill Monster]

Reading that thread over there (oc.net), I think it's possible that AMD might be completely rewriting their driver for W10. It would explain why there has been no official driver for so many months. Evidence to support this idea can be seen from Project CARS where people using it on W10 are seeing something on the order of at least a 20% increase in performance.

As always time will tell, but there may be a lot of performance locked up in the driver as the little increases in performance from overclocks seem odd for such a card (even on the core).

Not this again.....
The AMD drivers are not the reason for performance boost in W10, the increase is due to WDDM 2.0. The vendors have to rewrite the drivers to be compliant.

 

[Aquinus]

That is a 1.2% gain in score...

...and 50Mhz is a 10% overclock...

.... your point still remains though!!!


Outside of some math, my question would be, what would the scores of a 980ti do when adding 10% to its memory clocks?

My appologizes, I put a decimal place in the wrong place. Point still stands though.

 

[Mussels]

My appologizes, I put a decimal place in the wrong place. Point still stands though.

to be fair, being out by a single decimal point really changes the results a lot. this thread seems really confused by basic math.

OC = faster, but not always because not everything is VRAM bandwidth limited.

 

[LightningJR]

Here are some benchmarks I did on my machine.

http://www.3dmark.com/3dm/7576650?

http://www.3dmark.com/3dm/7576735?


This is a stock 670 and a VRAM overclocked 670. A 275Mhz overclock to be exact. An 18.3% overclock on the memory with a 3.15% increase in performance when using 3DMark Firestrike. Real world gaming performance increases will usually be even less. A timing "sweet spot" is unrealistic at best and an increase of 3k points in 3DMark Firestrike's graphics score is impossible if the tests are done correctly.

 

[Mussels]

its got nothing to do with timing sweet spots and everything to do with most engines/benchmarks favouring GPU performance over VRAM performance.

 

[mirakul]

Here are some benchmarks I did on my machine.
http://www.3dmark.com/3dm/7576735?


This is a stock 670 and a VRAM overclocked 670. A 275Mhz overclock to be exact. An 18.3% overclock on the memory with a 3.15% increase in performance when using 3DMark Firestrike. Real world gaming performance increases will usually be even less. A timing "sweet spot" is unrealistic at best and an increase of 3k points in 3DMark Firestrike's graphics score is impossible if the tests are done correctly.

I guess you have never tried to OC a Tahiti/Hawaii cards. Memory timing is a real issue with their BIOS. Also I can see you have also never touched a Maxwell, in which the latency improvement has a great impact on its delta color compression and the overall performance. Memory timing in nVidia BIOS normally is well set by the way.

About the 19321 graphics fs score, I'm seriously disappointed with your eyesight. It's right on the first post of this thread.

Here is another FuryX at same core clock and normal memory clock, 16735 graphics score
http://www.3dmark.com/fs/5223480

 

[LightningJR]

I did see the first post, I didn't know where you got your second number, the 16k. I was expecting to see both benchmarks done by the same person in order to keep variables to a minimum.. I don't know how many science experiments you have done but you can't compare those two results, it's bad science.

No I haven't overclocked Tahiti/Hawaii or a Maxwell. From reading benchmarks and overclock results of what is probably hundreds of video cards from multiple sources and my own tests that go way back to the S3 Virge days I can tell you that what me and many others are saying has the highest probability.

 

[EarthDog]

Basically, this is what needs to happen... Someone with this card needs to do a before and after WITH THE SAME EXACT SYSTEM and post in this thread. This mention of 'it jumped 3K points on the GPU score' but only showing the overclocked result is a bit sketchy.

I'm curious though, Mirakul, I am actually someone that 'plays' Fire Strike/3d11/Heaven etc. I have overclocked more Tahiti/Hawaii based cards than you have seen on store shelves... with that said, there was ALWAYS a relatively linear rise of my graphics/overall score in FS with those kind of cards. My 295x2 is a BEAST on overclocking its ram. I can go from 1250 MHz to the MSI AB limit of 1624 MHz and bench all day long there... what I noticed in finding the limit of that card in particular (and again, same with the others), was not what you are saying happens. There were no large increases in score it was fairly linear most of the way up. I am unable to speak to the Fury X at this time. I should have a Fury in my hands for review soon enough.

Also, I have heard of what you speak, however when you increase clocks, the timing table will LOOSEN thus negating any 'peaky' type of gains as you may likely know increasing the timings raises latency thus minimizing the effect on speed increases. So while timing tables may change its changes typically loosen things up for stability. When you lower the clock the timing table should tighten things up.

 

[Mussels]

Thread has been cleaned up a little, disagreeing and debating is fine but please no more insulting people and random memes.

 

[mirakul]

Basically, this is what needs to happen... Someone with this card needs to do a before and after WITH THE SAME EXACT SYSTEM and post in this thread. This mention of 'it jumped 3K points on the GPU score' but only showing the overclocked result is a bit sketchy.

I'm curious though, Mirakul, I am actually someone that 'plays' Fire Strike/3d11/Heaven etc. I have overclocked more Tahiti/Hawaii based cards than you have seen on store shelves... with that said, there was ALWAYS a relatively linear rise of my graphics/overall score in FS with those kind of cards. My 295x2 is a BEAST on overclocking its ram. I can go from 1250 MHz to the MSI AB limit of 1624 MHz and bench all day long there... what I noticed in finding the limit of that card in particular (and again, same with the others), was not what you are saying happens. There were no large increases in score it was fairly linear most of the way up. I am unable to speak to the Fury X at this time. I should have a Fury in my hands for review soon enough.

Also, I have heard of what you speak, however when you increase clocks, the timing table will LOOSEN thus negating any 'peaky' type of gains as you may likely know increasing the timings raises latency thus minimizing the effect on speed increases. So while timing tables may change its changes typically loosen things up for stability. When you lower the clock the timing table should tighten things up.

The first one with 19k graphics score was a bench of Hardware.info. They did mention the stock score (in the first post, you can follow the link to check), but it is overall, not the graphics score. Graphics score is more GPU dependent, hence I used a score from a different bench with the same core clock for comparison. The different in score is BIG enough to get to the conclusion though.

About memory timing, in addition to synthetics I also "played" Litecoin mining, one of the most memory intensive application. The problem with memory timing of Tahiti/Hawaii is more severe there than in gaming though. Problem lies in the inability to write a proper BIOS of some branch. You can search for The Stilt bios to get a clear picture of this. Interestingly, it is somehow identical to what HBM is showing. It is just like AMD planed to release HBM at other clock (maybe 625MHz as the early rumours), but had to change and settled at 500MHz.

 

[EarthDog]

The graphics component will go up (or down) with different CPUs friend... They have to be the same hardware at the same clock speeds in order to make a valid comparison... was that done in the two sources you site (honus is on you to look since it is your assertion).

Not sure you understood my point on Tahiti/Hawaii, it is NOT showing the same behavior you are claiming happens. As I said, I have overclocked dozens of these and never saw anything but a linear increase.

 

[mirakul]

I got your point on Tahiti/Hawaii, but it's the fact that I had that exact problem with my cards. Many guys on mining forum had it too.

And do you think the difference between 5960x and 5930k is enough to boost the graphics core that much? Note that this is the graphics score, which is mostly GPU bounded.

 

[15th Warlock]

Can't wait to see if HBM OCing translates to real game performance gains as opposed to synthetic benchmarks.

Hope Fury owners can contribute to this thread, AMD put all its chips on HBM, and it would be awesome to see the potential unlocked, as the whole industry is moving in that direction starting next year.