Why is memory so overclockable on certain GPU's?

[Recon-UK]

Let me go a bit back in history to give you an idea of my experience.

Radeon 4870 / 9800GTX+ neither could be pushed very far in regards to memory... +50 at max.

GTX 480 was able to hit 2050mhz from it's base rate of 1848 but those chips were rated for 2000 so... Nvidia purposely dropped those clocks.

GTX 580 could hit 2200mhz!

GTX 670 is ridiculous... base mem clock of 3000 i have not found the memory maximum clock but i am already +512mhz on it at over 7ghz whch is above the GTX 770 mem speeds.

I don't know if it is me but i noticed a trend, the newer the tech, the further it overclocks?

I can somewhat apply that to GPU clocks too.

GTX 480 max stable clock 920mhz 700 base
GTX 580 960mhz 850mhz base though this wa sa EVGA FTW HC2 - real stock is 775mhz
GTX 670 1300mhz stock 980mhz.




WTF?

 

[P4-630]

You're an unstoppable train man with your threads!!!

 

[Recon-UK]

You're an unstoppable train man with your threads!!!

 

[Toothless]

You're an unstoppable train man with your threads!!!

Google does not exist anymore, for TPU is the google and sanctum of tech wiki. W1z is our god that we trust in reviews and our mods are the angels with spears of ram sticks.

Silicon lottery, get it right.

 

[Recon-UK]

Google does not exist anymore, for TPU is the google and sanctum of tech wiki. W1z is our god that we trust in reviews and our mods are the angels with spears of ram sticks.

Silicon lottery, get it right.

Most cards will hit those clocks if you know what you are doing. GTX 480 typically maxed at 850-900 on air but as it's me i can crank more as i custom cool some stuff my own way.
People with 670's have hit higher than 1300 and well a lot have gotten over 7ghz mem too.
4870 and 9800GTX+ were not that well known for overclock prowess lol.

 

[peche]

You're an unstoppable train man with your threads!!!

+1

 

[dorsetknob]

Google does not exist anymore, for TPU is the google and sanctum of tech wiki.

Boy Scouts modified campfire Song
Bing Bang Google goolie what u can Baidu
set lyrics to this soundtrack

 

[Recon-UK]

Well since this thread is now a spam post thread i shall just leave.

 

[Toothless]

Most cards will hit those clocks if you know what you are doing. GTX 480 typically maxed at 850-900 on air but as it's me i can crank more as i custom cool some stuff my own way.
People with 670's have hit higher than 1300 and well a lot have gotten over 7ghz mem too.
4870 and 9800GTX+ were not that well known for overclock prowess lol.

Again, silicon lottery. Not every card will clock the same and you're comparing cards across families like fermi against kepler. It doesn't work that way.

 

[dorsetknob]

Guess if you leaving we will all say

Baidu

 

[Recon-UK]

Just so you know, memory chips have nothing to do with architecture.
GG

 

[peche]

Just so you know, memory chips have nothing to do with architecture.

actually they have to ... architecture is what controls / domanis or what ever you want to say to every process and memory management, clocks and transfers are linked to architecture ... also instrucionts are linked to memory, thats why you se some changes on frecs, clocks, volatage and of course memory .... senseless coment...

Regards,

 

[Toothless]

Just so you know, memory chips have nothing to do with architecture.
GG

Just so you know, you're wrong although Peche beat me to the reasoning.

GG m8

 

[peche]

Just so you know, you're wrong although Peche beat me to the reasoning.

GG m8

 

[Bill_Bright]

Why is memory so overclockable on certain GPU's?

I don't see where this was answered so I will.

Setting aside the fact the memory is on the card, not the GPU, the answer has everything to do with the RAM chosen by the card makers, not the technologies. The limits of technologies are based on the restrictions imposed by the capabilities of the raw materials, manufacturing techniques, and of course, the Laws of Physics. These limits imposed by divine intervention and human knowledge determines the maximum speed RAM (or CPUs, or motherboards, or car engines, or fan motors, etc.) can achieve.

The "engineers" in the development departments design these devices to the limits based on those restrictions and the "designers" of the cards (as influenced greatly by marketing weenies ) restrict (dummy down?) those speeds to meet the needs of that particular marketing demographic/target.

This dummy down process is done for several reasons. One has to do with the last line in my sig. Another main reason is to extend the marketability of the product so the company can recoup more of their R&D costs without spending more on R&D.

So Brand A can market their brand new Graphamatix 1000 today, milk consumers for a few months, then OC the card and come out with a "new and improved" and 10MHz faster Graphamatix 2000 and milk consumers for a few more months.

The other main reason has to deal with the manufacturers' intense desire to avoid zero profit, or worse, losses. This might be easier to illustrate (or at least me to explain) using CPUs.

Using AMD to illustrate (but Intel, NVIDIA, Motorola and the others do it too), they set out to manufacturer 4GHz quad core processors and the production run is for 1000 units. But testing shows only 500 pass testing at 4GHz so they label, package and ship those off to stores at the highest prices.

They test the remaining 500 to 3.6GHz. But testing shows only 400 pass 3.6GHz on all 4 cores. They label, package and ship those off to stores as 3.6GHz processors, but at a lower cost.

So now they still have 100 left and they don't want to toss them in the bin. So they disable the bad core, label them as "Tri-core" processors and sell them at an even lower cost so they can at least break even, if not turn a little profit.

 

[Recon-UK]

I don't see where this was answered so I will.

Setting aside the fact the memory is on the card, not the GPU, the answer has everything to do with the RAM chosen by the card makers, not the technologies. The limits of technologies are based on the restrictions imposed by the capabilities of the raw materials, manufacturing techniques, and of course, the Laws of Physics. These limits imposed by divine intervention and human knowledge determines the maximum speed RAM (or CPUs, or motherboards, or car engines, or fan motors, etc.) can achieve.

The "engineers" in the development departments design these devices to the limits based on those restrictions and the "designers" of the cards (as influenced greatly by marketing weenies ) restrict (dummy down?) those speeds to meet the needs of that particular marketing demographic/target.

This dummy down process is done for several reasons. One has to do with the last line in my sig. Another main reason is to extend the marketability of the product so the company can recoup more of their R&D costs without spending more on R&D.

So Brand A can market their brand new Graphamatix 1000 today, milk consumers for a few months, then OC the card and come out with a "new and improved" and 10MHz faster Graphamatix 2000 and milk consumers for a few more months.

The other main reason has to deal with the manufacturers' intense desire to avoid zero profit, or worse, losses. This might be easier to illustrate (or at least me to explain) using CPUs.

Using AMD to illustrate (but Intel, NVIDIA, Motorola and the others do it too), they set out to manufacturer 4GHz quad core processors and the production run is for 1000 units. But testing shows only 500 pass testing at 4GHz so they label, package and ship those off to stores at the highest prices.

They test the remaining 500 to 3.6GHz. But testing shows only 400 pass 3.6GHz on all 4 cores. They label, package and ship those off to stores as 3.6GHz processors, but at a lower cost.

So now they still have 100 left and they don't want to toss them in the bin. So they disable the bad core, label them as "Tri-core" processors and sell them at an even lower cost so they can at least break even, if not turn a little profit.

Thank you and top answer, why this was so difficult for others was beyond me, or better yet? don't really know? don't post

 

[Toothless]

I now proclaim @Bill_Bright as TPU's head of tech wiki manager since the amount of raw data output always amazes me.

 

[Frick]

I now proclaim @Bill_Bright as TPU's head of tech wiki manager since the amount of raw data output always amazes me.

I kinda miss lilhasselhoff. He could be an ass but man he wrote essays.

 

[Toothless]

I kinda miss lilhasselhoff. He could be an ass but man he wrote essays.

I wonder where he's been lately. I miss reading books on here on my lunch breaks.

 

[xorbe]

Do benchmark before you claim that +500 on the vram is actually effective. It has been noted in the past that performance actually decreases past a point -- theorized to be some sort of in-band texture crc (but not vram ecc). It's historically easier to kill a card with vram oc versus gpu oc.

 

[Recon-UK]

Do benchmark before you claim that +500 on the vram is actually effective. It has been noted in the past that performance actually decreases past a point -- theorized to be some sort of in-band texture crc (but not vram ecc). It's historically easier to kill a card with vram oc versus gpu oc.

I shall conduct tests for you sir.

Stock vs 512+ memory clock, also will use a standard core OC with it on both tests.

 

[Recon-UK]

+512mhz memory.

Stock memory clock.

Resident Evil 6 was barely using over 850MB of Vram, clearly this game is barely affected by memory clock though it still showed a performance improvement, it is negligible.

OC settings.

GPU will boost to 1248mhz if it stays below 70c which is impossible on the reference cooler this card uses so as you can see i am using a stupidly aggressive profile to force the max clock out of it, it hit's 1150mhz average, 1138mhz minimum and 1248mhz maximum.

I will be putting it under a custom cooler in time.

GPU-Z never reports correct boost clocks, even at stock the 670 will blow passed 980 reported to around 1050mhz.

 

[Grings]

Generally overclocking vram simply involves upping the mhz number, on newer cards as you up the ram speed the timings get relaxed, allowing for higher speeds (just as ram modules have higher timings as the speed increases)

I would assume on older cards they did not adjust the timings (or at least as much)as you increase the speed

 

[Recon-UK]

Dying Light on 60+FPS settings, first image shows the GPU hitting 1248mhz in the menu.

Both screen shots are done at the settled GPU core clock and my saved games was loaded up in the same spot on each run to produce a reliable comparison, this game is more demanding yet we settled at a higher boost clock which is kinda weird.

+512mhz

Stock.

Clearly a decent gain, at higher settings this is the difference between 50FPS and closer to 60FPS, makes a difference

 

[Iceni]

Gddr5 is based on Ddr3 chips that have been modified.

The chips themselves get upgrades. In 2007 they were made at 60nm. In 2015 that had been updated several time to a 20nm process.

In addition to that, controllers are faster and need less power due to GPU die shrinkage.

It's not unfair to compare it to CPU evolution where at the start of the DDR3 evolution we saw speeds of only 800Mhz. DDR3 800 isn't common because everyone wanted 1066Mhz or 1333Mhz.

Now DDR3 and DDR4 are now in that crossover period. DDR4's starting speed is 1600Mhz, but you would have to search to find that, And DDR3's top finishing speed was 2133Mhz but I know 2400Mhz is perfectly available (it's just not part of the official specification). That extra speed is because the whole manufacturing process of DDR3 is at it's peak. Flaws in the original spec have been corrected. And the factories are running systems that have been tweaked to perfection. In return as a consumer you get better chips that need less voltage, and will clock up far higher than was possible a generation or 2 ago.

GDDR5 is at the same level. Currently a sister spec GDDR5x is available, but GDDR6 will not be far away. And AMD have already moved some cards onto HBM and HBM2. Once the cards move over onto the new specification you may find that the insane overclocks are no longer possible for a few generations.