Published On: Tue, Dec 6th, 2016

Nvidia Geforce and AMD Radeon Graphic Cards Memory Analysis

Memory has always been one of a some-more critical components of a graphics card. Its importance, in new times has spin even some-more distinguished – utterly interjection to a appearance of HBM technology. While a concentration of a attention has shifted to that of maximizing bandwidth while maintaining an excusable earthy footprint – there lurks in a shadows a threat that could spin out to have distant reaching consequences: appetite expenditure of memory. we suspicion it was time that we do a brief editorial on a effects of bandwidth and appetite consumption.

As many of we will know, a memory customary for striking cards of a past era is GDDR5, and while many destiny offerings will positively keep this storage type, there is a new actor in town: High Bandwidth Memory. HBM or built DRAM (which comes in many opposite names and flavors) is a new creation – with significantly augmenting bandwidth though drastically lowered appetite consumption. The whole indicate of HBM was to move high bandwidth to an affordable spin (in terms of appetite cost). And while they have flattering most succeeded in doing that for a nearby future, a problem persists when articulate about a prolonged tenure horizon. Nvidia had their GPU Technology Theater discussion during SC15 a while behind and Dr. Stephen W. Keckler, Senior Director of Architectural Research (at Nvidia) mentioned that a problem of a memory appetite expenditure stays unsolved – even with HBM.

Nvidia Looming Memory Crisis SC15

The graph shows an exponentially augmenting trend in appetite expenditure of a dual memory forms – as bandwidth increases. We can see a constructional mangle in a graph where a attention shifts from GDDR5 to HBM, though a finish outcome stays a same – despite delayed. Now since is this a problem? one competence ask. In terms of a TDP fan rigs are used to – these numbers are good within range. But we have to comprehend that a GPU that would need so most bandwidth would devour utterly a lot on a possess too – and a sum TDP will not be within excusable ranges. The appetite consumed by a GPU is a really critical number, and if usually a memory requires such a vast volume of appetite than a sum series would be flattering huge.

The appetite of striking cards is augmenting during an harsh gait that means that within a few some-more years we will need bandwidth in a region of a few thousand GB/s – that would meant that even with a appetite potency afforded by built DRAM, memory appetite expenditure will arise above a 100W symbol – that deliberation a fact that a GPU themselves will be sipping utterly alot appetite on their possess – means sum wattage above a 250W symbol high finish cards now employ.

Some of a readers reading this will substantially be wondering since we need bandwidth in a thousands of GB/s. Well, we have to remember a fact that striking cards are indeed processors that have outrageous clusters of impossibly little cores. You can call them CUDA cores, or we can call them tide processors – possibly way, it is these little processors that need a bandwidth. Now while a series of these cores has been augmenting during a really solid rate, a bandwidth afforded to them has not been means to keep pace. To elaborate a statement, we distributed the bandwidth accessible per core of new Nvidia and AMD striking cards. A really transparent trend is immediately visible.

Bandwidth per core research of Nvidia and AMD striking cards

Bandwidth Per Core AMD Radeon Graphic Cards 1Bandwidth per core of opposite Radeons, with a series of tide processors in descending order.

Bandwidth Per Core Nvidia Geforce Graphic CardsBandwidth per core of several Geforce cards, with a series of CUDA cores sorted in descending order. 

You will notice that a bandwidth accessible to any core decreases as a GPU grows some-more absolute in nature. This is since of a inherently delayed gait of stream memory technology. Interestingly, we can see that AMD has been means to keep a slope of a Bandwidth Per Core comparatively consistent while as Nvidia has usually a somewhat steeper gradient. Nvidia’s label operation in a interlude of 107MB/s to 120 MB/s per core going as high as 160 MB/s per core on a GTX 760.  This is ofcourse something we have famous all along – AMD is some-more inexhaustible with dispersal bandwidth to a striking cards than Nvidia – that uses tone optimization record to recompense for a miss of bandwidth. AMD cards operation from 110 MB/s to 140 MB/s per core, going as high as 175 MB/s per core for a R9 370. Interestingly, a usually label in a calculations that forsaken next a 100 MB/s per core symbol is a R9 285 – that has a series of 98 MB/s per core.

Lets try a appetite quandary a bit more. The appetite expenditure of flighty memory is a duty of a time speed and train breadth (aka bandwidth). According to a statement by AMD, the latest chronicle of GDDR5 (on average) uses about 1 watt per 10.66 GB/s of bandwidth whileas HBM is good for about 35 GB/s per watt. This is a phenomenal, scarcely 3.5x, boost in appetite potency – though is it enough? We computed a estimate appetite expenditure of a several striking cards and arrived during some really engaging conclusions.

Memory appetite expenditure trend of AMD and Nvidia striking cards

As approaching with Nvidia, a appetite consumed by a memory is augmenting during an approximately proportional rate to a normal appetite of a striking card. Keep in mind that cores opposite generations are not totally comparable (the normal Maxwell core is 15%-32% some-more absolute depending on a card) so a tangible series (after adjusting for opening differences) would be somewhat higher. We see that on normal an AMD label uses around 15W to 36W to appetite a GDDR5 memory customary whileas Nvidia cards use around 18W-32W depending on a accurate card. With GDDR5 – a limit volume of wattage so distant is nearby a 40W symbol (after accounting for genuine universe line losses). With HBM however, things take an altogether opposite turn. HBM memory uses usually 15Ws and provides some-more bandwidth than a top clocked GDDR5. Infact if we demeanour during a graph we will notice that in a box of a Fury lineup, a bend breaks from a approaching trail and indeed slopes downwards – indicating a constructional mangle in trend we talked about.

By adhering to a HBM customary we can discriminate what a bandwidth will be like during a appetite levels settled by Nvidia. At 120W an HBM (at today’s efficiency) will be means to outlay 4200 GB/s – that is an positively outrageous number. Unfortunately however, potency does not scale linearly and we competence as good be looking during a cause of 20-25 GB/s per watt during consummate puts that high. Which gives us a series around 2400 GB/s – a most some-more reasonable series that we can design to see in a nearby future. Currently, a appetite expenditure of a memory ranges anywhere from 8-15% of a sum TDP of a GPU though as a bandwidth increases – this series will go adult as well. Ofcourse, there are many choice technologies already in a tube – including standards being worked on by Intel, Rambus and Micron. This projection does not (and cannot) take into comment disruptive new technologies that reset a bend once some-more – a probability that is some-more than likely.

Nvidia Geforce – Memory Analysis


AMD Radeon – Memory Analysis

About the Author

Leave a comment

XHTML: You can use these html tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>