Dan's Data letters #35
Publication date: 6-Mar-2003.Last modified 03-Dec-2011.
Constructive criticism
Just read your article on Thermaltake's RAM sinks.
I'm sorry, but it is just complete, unproven, nonsense.
Firstly, nowadays, the CPU FSB and memory speed is one of the biggest bottlenecks in a system. The primary goal in overclocking is now to get as high a FSB as possible, raw CPU MHz barely matters at all. As an example of a top overclocked system, 10MHz of FSB increase will give you far more gain than 100MHz of CPU speed increase. I'd go as far to say that 5MHz of FSB increase will be more beneficial.
This certainly holds true on an AMD platform, where memory and bus bandwidth is an even bigger bottleneck compared to the quad pumped buses of current P4 platforms. Although modern P4 platforms gain huge performance increases from running memory async to the bus and trying to get as high a memory speed as possible - due to the fact that a P4's FSB bandwidth is greater than the bandwidth on their memory bus.
Can you show me your benchmarks showing FSB not to have much gain? Because I can show you dozens of benchmarks (REAL WORLD benchmarks, not synthetics) that show the complete opposite to what you have said.
Here's one I did a few months back to point out how much FSB matters over CPU MHz.
There are dozens more where that came from. You only have to look at places like hardocp.com and hexus.net to see that high Front Side Bus (AMD platforms) or high RAM speed (Intel platforms) result in the biggest performance boosts - far more than a CPU MHz overclock alone.
Finally, to the comments about the Thermaltake RAM sinks being decorative only. You have done some real world testing then? These heat sinks are aimed at overclockers. So I assume you have taken some high quality RAM that doesn't come with heat spreaders, say TwinMOS (Winbond) PC3200 modules, run them at 210-230MHz, 3 volts or higher, INSIDE a closed case system (not on a nice bench open to outside air), or inside a water cooled system that has even less air flow around the RAM area. Then attached accurate thermal sensors to the RAM chips.
If you did, you will find them just a little bit more than "warm". More like too hot to touch, for fear of burning your fingers! You wonder why the top PC3200/3500 modules come with heat spreaders? It's because if they didn't they'd be likely to fry within 6 months of use, especially in an overclocker's hands, who will shove 3 volts or more into them.
RAM sinks wont get you hardly any increase in overclockability alone. What they will allow you to do is put a higher voltage to your RAM without resulting in fried modules. This will then give you your extra overclock. So the net result is an increase in overclockability, and reliability.
Unfortunately most "small time" reviewers like yourself don't really test products properly for the benefit of an overclocker. Since this product is aimed at overclockers then tests should be done in that way - i.e., any reviews on these heat spreaders should involve getting some very high speed RAM and overclocking it past its already high specification with voltages above specification. You will then see just exactly what the product will or will not give you.
It's plain obvious that you have done very little proper testing or research on this matter. Please don't publicise articles that are unproven, non factual, nonsense.
Gavin
Answer:
Memory speed isn't just a bottleneck now; it's always been the biggest
bottleneck in PC systems, ever since CPU core speeds started being multiples
of the system RAM speed. Arguably, since before then, actually; see my reply
about the "von Neumann bottleneck" in this column.
Since multipliers have been increasing faster than the RAM bus speeds, though, the problem can be expected to be getting worse and worse. But it's mitigated by larger and larger caches, and by the fact that most PC tasks don't lean very hard on main memory.
As you say, though, in current systems with un-linked FSB, RAM and PCI/AGP bus speeds, you can indeed realise considerable gains by boosting the FSB greatly, with little or no core overclock.
That wasn't the case when I wrote the Atomic magazine column you object to, back in 2001, or when I put the slightly updated version of it up on the Web at the start of 2002. Back then, super-high (for the time) RAM bus speeds inescapably meant super-high PCI/AGP/ATA bus speeds as well, which was no way to run a railroad. But no matter; you weren't to know when the column went up.
The question is - do these system RAM sinks do anything useful?
The results chart you mention demonstrates that, on recent hardware, 1.5 times the RAM clock speed delivered 1.3 times the performance, for a couple of pretty system-RAM-intensive 3D game benchmarks. That's more or less on par. Some tasks, as you know, lean lightly enough on RAM that overclocking it does nothing in particular; usually, for what passes for a really RAM-intensive task on a PC, you get 50 to 60% of the RAM overclock in performance gain.
It's instructive to note that this gain hasn't actually changed a great deal despite the more rapid advance in CPU speeds, compared with system bus speeds. Back when DDR was available but stupidly expensive, the gain-per-RAM-bandwidth wasn't much smaller than your recent tests show, which is why the more-than-twice-as-expensive DDR wasn't a good buy.
And yes, in answer to your question, I did real world testing of Thermaltake RAM cooling gear, when I first reviewed it here. That was back in the SDR memory days, mind you, but memory chips have only gotten cooler since.
With regard to the test you suggested, a 210-230MHz clock speed for PC3200 DDR RAM is only a 5 to 15% overclock. All this effort for 15% more RAM speed than you can get at stock settings - which'll give, going by your own numbers, less than 10% better performance - seems a bit excessive to me. Never mind, though.
What you said about closed cases and the heat inside is more significant. If your case has crappy ventilation, and you haven't gone the whole hog and put water cooling blocks on everything in sight, then you shouldn't have ripped out or slowed down all of your fans. The solution to a high inside-case temperature is not better heat sinks on everything inside; that'll just run you straight into a diminishing-returns situation. If your computer is badly ventilated, you need to give it better ventilation. If that spoils the silence of J. Random Overclocker's zooty water cooled creation, then that's J's problem; he or she should have come up with a better engineered solution. Filling that box of hot air with anodised aluminium won't help much.
I couldn't find any spec sheets for the TwinMOS/Winbond RAM you mention, but Micron have specs online for their stuff, which isn't very different, so let's work from those.
Micron's 256 megabit DDR400 (PC3200) chips have an absolute maximum specified current, for continuous four-bank interleaved reads, of 470mA at the nominal 2.6V. That's 1.222 watts, but it only applies if you're doing nothing but four-bank interleaved reads without any idle time (which is unlikely) or writes (which is extremely unlikely) or refreshes (which is, as I understand it, impossible). Real world working-hard numbers aren't likely to be much more than half of the peak numbers, but let's be pessimistic and say three quarters.
Now we've got 0.9165 watts as an improbably high average power consumption, and therefore heat output, for one chip. It helps to remember, at this juncture, that whatever the actual real world heat output is, this RAM is specced to handle it with no special cooling in an ordinary case, and demonstrably does. Memory overheating is not a problem for DDR-RAM PCs running at stock speed - well, unless the case is plugged up solid or the fans have failed, so everything else is overheating as well.
So. Start with that DDR400 memory, and let's say you're a star overclocker who's managed to increase the clock speed by a big 20%, to 240MHz, and the supply voltage to 3.1 volts - almost 20%. You'll get 20% more heat from the higher clock and 42% more heat from the higher supply voltage (because heat output for ICs, in general, increases linearly with clock speed and also with the square of the input voltage), for a total heat increase of about 71%.
Now you're up to 1.56 watts per chip. A 512 megabyte module, with 16 chips, thus has a pessimistic maximum heat output of just about exactly 25 watts.
That's quite respectable - if it were a CPU, it'd need a heat sink with fan - but it's spread over the whole module, which I remind you was A-OK to run at about 60 per cent of this heat output with no special cooling at all.
Now, since you can buy RAM off the shelf (like this imposingly heat-spreadered stuff...) that's rated for 233MHz at a mere 2.7 volts (and which therefore, using the above pessimistic estimates, is probably only producing 1.1 to 1.2 watts per chip), I'm unsure why you'd want to crank the RAM voltage well above what most BIOSes will let you just to get a few more per cent. But, again, no matter.
On the subject of thermal sensors on RAM - if you've figured out a way to get accurate readings of the RAM temperature after you've attached a heat sink to the module, I'm all ears. As far as I can see, it's not possible, at least without some serious surface mount rework. Sensors on top of the chip will disrupt the thermal contact and give you an average of the cool heat sink temperature and the warm RAM temperature, and there's no room for sensors underneath.
Measuring the temperature of a RAM heat sink/spreader from the outside is, of course, not informative; if the heat sink/spreader is cooler than the RAM was before, that may mean the RAM underneath is being well cooled, or it may just mean that the thermal interface material between the RAM and the metal isn't working very well. Coolers like the Thermaltake ones, which have thermal tape between module and metal and, of necessity, clips that don't clamp down nearly as hard as many CPU cooler clips (if they clamped really hard they'd be very difficult to attach, and might well damage the RAM), are unlikely to achieve a very good thermal contact.
It's worth noting that lots of electronic devices work fine when they're too hot to keep your finger on. Making them cooler may allow you to run them faster, or it may not. Lots of heat doesn't indicate a superior performance ceiling, any more than lifting the inside wheels on a corner means you're going to get a better lap time.
Since you ask, I don't wonder why some DDR RAM modules come with heat spreaders; given the above numbers, I imagine it's for the same reason that so many cars come with spoilers.
RDRAM needs heat spreaders because of its ability to concentrate the whole module's activity, and heat output, on one chip for a non-trivial period of time. SDRAM spreads its heat naturally.
MORE (drill) POWER!
In this article you said "I have not yet heard anybody extolling the virtues of running your cordless drill from a truck battery, because lead acid batteries don't have memory effect but everything else does, but I'm sure someone has. Probably someone with an impressive collection of trusses."
A friend of mine uses his ute (actually an ancient Land Cruiser, so sort of a light truck depending on who you talk to) battery to run a cordless drill quite a lot. He bought a cheapo 12V cordless ages ago, and through a combination of neglect and heavy use, managed to kill the battery pack within six months or so. His solution was to gut the battery pack, then solder an appropriately long cord and alligator clips into the pack, and he presto he had converted his useless cordless drill into a corded drill.
Apart from an extremely long battery life, there was another unexpected benefit. The drill turned into a serious drilling machine capable of drilling through heavier material and drilling faster under load. We think the reason is that the motor can get all the current it wants from the car battery, but it may have been a bit limited in the load it could place on its battery pack. Perhaps part of the extra torque may be from the 13.8 volts when the vehicle is running, but it still works much better then before when the ute is off. No signs of memory effect as yet :-).
A disbenefit is the reduced portability, but for his purposes that isn't an issue because most of the time he needs a cordless drill he is out in a paddock fixing a tractor or something of that nature, and his ute is parked close by anyway for ready access to the assorted tool boxes on the tray.
Andrew
Answer:
Fair enough. It'll take you a good long while to run even a small car battery
flat with a cordless drill, so this is a perfectly workable proposition.
There are some heavy duty pseudo-cordless tools, by the way, that use a big battery that you're expected to tote around on your belt, or put on the ground while you use them. Same principle, but probably better looking than your friend's solution.
You're also right about the source of the extra power.
At 13.8 volts instead of 12, when the Land Cruiser's alternator is turning, the drill will have about a third more power than it'd otherwise manage, thanks to DC motors being more-or-less straight resistive loads for these purposes, and power thus increasing with the square of the input voltage.
Higher current capacity from the Land Cruiser battery would, indeed, account for the drill's better performance with the engine off. Many cordless tools use cheap and nasty cells in their battery packs which don't have great high current performance; they're miles ahead of alkaline cells, but not too good for the very high current draw of a brush motor at low speed. Nickel metal hydride (NiMH) cells are worse than nickel cadmium (NiCd) cells in this case; drills with cheap NiMH packs are the least gutsy, all other things being equal.
The tool's performance will also drop as the cells age, which they may do quite quickly if the charger's both fast and dumb, as many cheap cordless tool chargers are.
Any car battery that can still start the car will, in contrast, still be around an order of magnitude more capable than it needs to be to run a cordless drill.
Top-class cordless tools with high drain capacity cells would probably perform only somewhat better from a car battery. Nutcases can and do refit dead drill battery packs with expensive Sanyo RC NiCds.
Baby batteries
I need some new button cells for one of those old laser pointers. The old cells are LR44s, but it seems that for some odd reason they're no longer sold in warehouses like they used to be. I did find SR44 and PR44 batteries though. I quickly looked up some specs and the PR seems to be way off from the LR, but the SR is pretty close. Could you tell me what exactly the difference is, and what the effect would be if I run the pointer on SR and PR compared to LR?
David
Answer:
LR/PR/SR designations are IEC identifiers
for batteries. As you may have noticed, button cells often have many other
different identifiers.
LR batteries are zinc manganese dioxide cells, the same chemistry as is used in normal "alkaline" batteries for larger devices. They've got a nominal 1.5 volt per cell output, and slide down to about 1.0 volts in the course of a normal discharge, before they're functionally flat. LR batteries are good for high drain applications like laser pointers, because although their nominal capacity is low, they're better able to supply high currents.
"High" is a relative term, of course; you're still basically beating them to death in a laser pointer, but they're a bit more beat-resistant than the other chemistries. If you want really long battery life in a laser pointer, no button cell will give it to you; get a pointer that runs from a couple of AAs instead.
SR cells are silver oxide, nominal 1.6 volts per cell, with little voltage sag over the life of the cell - a near-flat SR cell will still have a terminal voltage above 1.5 volts. SRs also have about 1.5 times the nominal capacity of the same size of LR cell, and work better at low temperatures, and last longer on the shelf. They don't like high discharge rates, though; SRs are better than LRs for low drain applications like calculators and wristwatches, but worse for laser pointers.
PR cells are zinc air, with 1.4 volt per cell nominal capacity. To activate them, you remove a stick-on seal that stops air getting into the cell. With the seal in place, they'll last for around five years on the shelf, but they'll only last about three months, at most, with the seal removed - and they need a ventilated battery compartment to work at all. They've got two or more times the nominal capacity of silver oxide cells, and a similarly flat discharge curve, but, once again, they don't like high discharge rates. They also don't like low temperatures. You use PRs in things that're on all the time, and have low drain; their principal application is hearing aids.
