Slashdot Mirror
Memory Manufacturers Could be Cheating
Mark Brown writes "Tom's Hardware is live-testing DDR2 memory products in order to determine whether memory manufacturers submit cherry-picked products for reviews. 'GeIL DDR2-667 that was claimed to be purchased performed worse than the review samples they got: 471 MHz for the review samples vs. 421 MHz for the retail memory.'"
Retired Washington Post food critic Phyllis Richman used to work around trumped-up meals like this by dressing down and not telling restaurants what days she'd be coming in to try the food. She even went so far as to hiding most of her face in photos so no one could publicly identify her. Really, one of the best critics to ever be published, I really miss her reviews.
Product reviews in general are a bit more difficult. Although the aforementioned Consumer Reports has a great thing going for them in purchasing products from stores, the thing is that they can AFFORD to do that. Most publications and websites can't, forcing them to rely on review samples. Car companies in particular are notorious for fine-tuning their review vehicles, which why Consumer Reports is highly respected for their year-end car accolades.
Taking a sample size of 1, not really. Their test leaves something to be desired. They really ought to be testing both memories in both systems, several times before jumping to conclusions. Slight variations in PCBs and silicon can build up to cause appreciable differences. Ultimately overclocking is taking entire designs well outside their specified operating limits. To do this reliably you need to test thoroughly on many samples.
The part of it that convinced me that they're right anyhow is the memory supply voltages. "Normal" on the cherry picked Gigabyte board was ~2.2V, normal on the storebought was ~1.83V (FVI 1.8V is the DDR2 spec supply voltage). You'll have to take my word for it, but THAT variation is huge. People who build computers do not tolerate voltage discrepancies like that, it's out of spec for the devices which usually allow 5% variation (1.71V-1.89V). You can verify this by going to Hynix/Micron/Infineon and pulling down a DDR2 component datasheet.
The headline is beyond wrong though, it's probably actually criminal. GeIL does not control the memory supply voltage (they make the DIMM), Gigabyte does (they make the mobo). GIGABYTE is cheating.
It's very easy to figure out if memory makers are cheating: take the heatsink off, look at the device part numbers and look them up. There's not a whole lot to tweak that doesn't involve a complete redesign of the DIMM. If they cheat it's almost always because they used a DDR2-400 device but branded their DIMM as DDR2-something_higher.
http://www.hardocp.com/reviews.html?cat=MjUsRGVza3 RvcCBDb21wdXRlcnMsaGNvbnN1bWVyLCws
What they are doing is having other people buying systems and then reviewing those systems. They will only review systems where they have an agreement with the manufacturer that the computer can be returned at the end of the review. The key is that the manufacturer never knows who is getting a system which may be subject to review.
It actually works well for both parties. Some manufacturers are proactive in the forums and even acted on complaints received, strengthing their processes.
* Winners compare their achievements to their goals, losers compare theirs to that of others.
Let's see - the GeIL memory is rated at DDR2-533. The module from the vendor ran at DDR2-942. The module from the store ran at DDR2-842. Now, Tom makes this out to be some big controversy, but it seems to me that a module running 36% faster than specified is no small thing, particularly at that high of a data rate.
I'm an engineer who designs memory modules. In most cases, our modules are overclockable, at least to some degree - some go faster than others. At the sort of speed that Tom's Hardware is running, I'm not really surprised that there's more than a 2 or 3% variation in performance, espeically if the chips on those modules came from different manufacturing lots. At the outer limits of memory speed performance, the tiniest changes in parasitic capacitance can be death to performance - and those values change from lot to lot, even from wafer to wafer.
When manufacturers specify that 2% to 3% tolerance, they're referring to the module's performance at its rated speed, and that makes sense. Plug two modules into a system and they will run in virtual lockstep - at their rated speed. There are a million analogies that I could use, but the bottom line is that there are assumptions and statements in Tom's article that just aren't right.
Maybe the module was cherry-picked and maybe it wasn't, but, if nothing else, a sample of two doesn't make for much of a study. After all, if the retail module had been DOA, a pedantic person could say that GeIL cherry-picked the evaluation samples and sends all the defective modules to retail.
-h-