Slashdot Mirror
DDR3 RAM Explained
Das Capitolin sends us to Benchmark Reviews for an in-depth feature on DDR3 memory that begins: "These are uncertain financial times we live in today, and the rise and fall of our economy has had [a] direct [effect] on consumer spending. It has already been one full year now that DDR3 has been patiently waiting for the enthusiast community to give it proper consideration, yet [its] success is still undermined by misconceptions and high price. Benchmark Reviews has been testing DDR3 more actively than anyone. ... Sadly, it might take an article like this to open the eyes of my fellow hardware enthusiast[s] and overclocker[s], because it seems like DDR3 is the technology nobody wants [badly] enough to learn about. Pity, because overclocking is what it's all about."
Great, now that I know how well certain DDR3 chips overclock, can I see how DDR3 compares to DDR2 in terms of raw performance and overclocking ability? I'm not an expert on how DDR2 works, so DDR3 could be better explained to me by showing me how it is better than DDR2.
DDR3 is still 5-10 times the cost of DDR2, and the performance gain is not big (maybe 10% at best) on overall system performance.
Lawyers, MBA's, RIAA? A jedi fears not these things!
Memory bandwidth is not problem right now. I/O is too slow. We need faster disks and LANs first.
Actually, with the current technology, such as 900 MHZ CPUs, 1 gig of RAM and Linux, things are selling. I am referring to the eeePC along with the other UMPCs that are growing in use. The slowing US economy has made computer makers adapt to lower spending incomes or face lower sales. Other computers such as the gPC reflect this. Along with the growing use of Linux (which is free) helps make the computers cheaper so they sell better. MS has been failing lately due to Vista's high-priced hardware need, compare that to Linux which can run on a Pentium II or III. New technology doesn't sell well in this economy, cheap, older technology does.
Taxation is legalized theft, no more, no less.
... really careless[ly] written submission[s] that the editors [have] to fix [up] before they can be [read] by us[?]
Generally, bash is superior to python in those environments where python is not installed.
What is the one and only thing that has ever helped companies servive a depression?
Burning down the factory and collecting the insurance?
What?
It's always nice to see a tech writer full of teenage enthusiasm, but this article goes a little over the top.
It's supposed to be an article about a performance enhancement, and there's barely any performance values at all (except for the theoretical peak throughputs on page 3, and we know how much that means).
I think what the guy really wanted was to write about planes, not about computer hardware.
Because some value stability over speed.
I was asked to purchase 4 of the fastest desktop PC's I could find (just less than a year ago) for a contract placement; I ended up going with the QX6850 and for 3 machines DDR2-800, and with the extra money on the 4th machine, going with a capable motherboard and DDR3-1333; hoping for at least some speed difference.
First of all, as the technology was brand new at the time, the motherboard, although capable of 1333mhz ram, it only detected it as 1066 (we double checked they sold us the right stuff), so we manually set the RAM in the BIOS to run at 1333.
After all the setup, on otherwise almost identically configured machines, we found absolutely 0 performance gain on the DDR3 machine over the other 3 DDR2-800 machines. Although one might argue that our applications we were using to test were not so memory intensive, the fact is it was a computationally intensive task that regularly accessed about 200-300mb of data from ram. I would think that even if everything would be pre-fetched into the 8MB CPU cache before it was used, we should at least see some small difference.
In the end, it seems that we spent an extra $800 for no noticeable performance gain.
Hard disk speed, until SSDs become very common this is one of the causes of decreased speed because a HD can only run so fast
And the actual reason memory manufacturers have such a hard time selling memory performance is an extension of this; most purchases of memory have an implicit tradeoff: amount versus speed.
In most cases the real memory pain will come when you hit swap. That means you will almost always notice having less memory more than you'll notice having slower memory. So basically the only ones who end up having 'fast memory' on their system purchase checklist is those for whom money isn't an issue at all (ie, there is nothing else you could buy for that money that would improve your system (or life) more than faster memory). Or those who have very small applications they need to execute blazingly fast. Such as benchmarks.
Until those problems get fixed, faster RAM won't make a bit of difference to the end-user.
More accurately, faster RAM will make less of a difference than more RAM.
Hardware hack, smart guy.
I think I'm being trolled.
"You can either have software quality or you can have pointer arithmetic, but you cannot have both at the same time."
The problem with sales of DDR3 is probably down to the fact that most computers are being sold with versions of Vista 32. To really make a difference with faster SRAM would probably also mean having a larger amount of SRAM. But the problem is Vista 32 only addresses just under 4GB of SRAM. As many computers for the home market max out by being supplied with 4GB of SRAM as new, the computers are effectively un-upgradable unless a different operating system is used. This leaves an interesting situation:- 1. Run Vista 64 - but then a large portion of hardware and software will not work as you cannot sign unapproved drivers. 2. Run Linux 3. Buy a MAC And there is the problem, the alternatives to using DDR3 are in fact called using a different operating system and adding standard DDR2 memory, which is more cost effective that paying the extra for the DDR3 SRAM. The only area I could possibly see DDR3 technology in the commercial hardware is in cutting edge graphics cards, where performance is everything. Unfortunately, as Vista 32 is already at is technology limit and that many potential users of this technology will baulk at the price, I can see games console manufacturers adopting this technology before the general PC market.
You're still lacking the source for
* BIOS
* BMC
* WLAN Card
* Disk Controllers
etc. pp.
While there's some of that, there's also the fact that the silicon itself wasn't fabricated at a certain speed. Once cut up, it's evaluated for the speed. If the silicon was capable of performing at higher speeds, but the market is only paying for lower ones, guess what do the cpu makers do?
Yes, mark it at a lower speed, and sell it as such.
Why do you think they "lock" things?
I wouldn't call it a success, either. I'd wager that figure is 90%+ copies that came with new PCs. The large majority of which probably end up in a corporate setting where it was re-imaged with XP Pro (happens all the time where I work and for our clients).
Uh the Opteron 165 in this machine is supposed to be running at 1.8Ghz. It's been doing 2.4Ghz for several years now with no problems at all. It will do 2.6 no problem but I like to back off a bit from what is possible.
Every machine I have had since my 120 Pentium Doom special o/c'd to 133 has been overclocked with no problems. Once you have everything stable there is almost always at least 10% with room to spare.
He's like a jet engine supporter immediately after World War II:
"Jet engines are inherrently capable of much greater speeds than propeller engines!"
"OK, so show me one that goes much faster than the prop driven spitfire?"
"Well, uh, there's the Gloster Meteor!"
"It does about 500mph*, right? That's not bad compared to the Spitfire XIX's 460mph. But there are tons of Spitfires out there, available cheaply vs. paying several times the cost for the Meteor for about a 10% speed improvement." *note: F-3 variant, not the "overclocked" tweaked versions that set speed records.
"Well, but the point isn't that it's better today. It's a better technology! It'll be better in the future! Props will never go supersonic. Jets can potentially go several times supersonic."
"That is cool. Doesn't really help me today though, does it? I'm still paying several times the cost for a small improvement, today."
"Yeah, but if you don't buy jets now, how will their prices ever come down? How will we ever reach the heady perfection they're capable of?"
"Again, not helping me today, is it?"
"But! But! It's really cool!"
Yes, the technology shows promise. But its future promise with only small increases today doesn't justify its current high cost.
If more people bought it, the cost would come down over time and more investment would mean unlocking more of that promise. Which is great in the future but gives you very little today in exchange for that high cost.
The argument he seems to be making is that everyone should adopt it right now, not because it actually gets them much for their money but because their investment will enable him to buy even faster stuff for a lower price later.
Not really compelling.
I really like ECC memory. I've built a lot of machines here and there for various purposes and people and can safely say that ECC should be mandated. People oft talk of blue screens, panics, etc. A lot of these are due to bad memory. And these types of errors can often be hard to replicate. I believe EETimes had an article about how bad bad memory is.
Anyhow, I'm on a 975X chipset with 4GB of DDR2 800 MHZ unbuffered ECC memory machine now. Not a single unforced error since I bought this machine and assembled it December, 2006. Nothing. This unit is primarily a gaming rig, the 3DMark 2006 score is 11500 with an 8800GTX, all with ECC memory.
The most irritating thing for me is, looking at the great new CPUs available, is the utter lack of any unbuffered ECC memory in the DDR3 range. This to me is unacceptable. I refuse to compromise so I will wait. Intel has a motherboard featuring DDR2 800 fully buffered memory for the 'high end workstation' , D5400XS, this is $600+. Supermicro offers something similar.
The X38 does DDR2 ECC, and for whatever reason the X48 took that away. I don't get why Intel wants to deny us DDR3 unbuffered ECC? Its really a selling point, a very good thing. If you overclock, its nice to have because it can tell you the limits minus the guesswork, not that I would bother with OC personally.
Fundamentally, without ECC, do you even know if the memory works at all? My experience leads me to believe that without ECC present, the answer is no at all.
Legalize the constitution. Think for yourself question authority.
This article seems to be paid advertising, with some bending of the facts.
And the fact is that the double prefetch buffer is the sole reason for the double bandwith and the double latency. As a matter of fact the speed of the individual memory chips on the ram module are the same as on ddr2 (see that table in the article, just divide the ddr3 clock by 2 to get the corresponding ddr2 speed) - but instead of reading 1 bit from 4 chips at once into the prefetch buffer (that is four bit prefetch buffer), they are reading from eight at once (so that's the 8 bit prefetch buffer), so double the amount of data can be read in the same time (hence the double bandwidth). However because the memory chips are the same speed as in ddr2, the time needed to program the individual chips stays the same - so because the clock is double the speed, it takes twice as many clocks to tell the individual chips which bit we want to fetch. And that bullshit about lower latencies is also not quite right: ddr3-1600 cl6 is the same latency as ddr2-800 cl3 - and such modules have been sold for years.
Of course ddr3 is better, because it has higher bandwidth, and absolute latency is not worse than ddr2's. Also there are in deed technological advances (e.g. the lower voltage). But this article is still not exactly honest.
I run memtest for a minimum of 48 hours on any new system I build and have never had any problem with RAM that has passed that. This is the best I can do without the premium of paying for ECC capable motherboards and RAM.
And conversely; a reasonable number fail. Working as a system builder and memtesting every machine I build, I've seen around 15% failure rate in all brands of RAM.
[clever sig]
I have yet to have a customer offer a contract for a full ECC system, but I expect that type of customer to be ordering in the dozens or hundreds. For one machine or 100 I would still give 48 hours of memtest. For ECC I would still give 48 hours of memtest.
If you can recommend a better test, I am open to suggestions.