Future of RAM

Got this question:

Ed, do you think, with processor speeds increasing to above 2 GHz anyway, that it’s time for AMD to bite the Rambus bullet ? I mean, RDRAM on the P4 works REALLY well (forget the P4 itself, perhaps, but the memory bandwith is fantastic), and the cost is down to $1.00 per megabyte.

This really is two questions:

Do you think they should?

No, I don’t see any particular reason for them to with the current generation of processors, and there are supposed to be quicker DDR solutions available which should take of AMD’s requirement fine for the next few years.

Do you think they will?

Hell, no, unless those quicker DDR solutions fall flat on their face.

The problem is these are not the right questions to ask (unless you’re a Rambus shareholder, maybe).

Memory Bandwidth is Like Cow Dung

Let’s say I have ten tons of cow dung. Do you want it?

Most of you wouldn’t, simply because you live in an urban or suburban area and have no possible use for it. Even if you have a garden, you hardly need that much.

If you’re a farmer, though, this offer sounds much more appealing (well, to a point). But why? Not because you like the smell better; it’s because you can do something useful with it.

Imagine a bunch of city folks feverishly outbidding the farmers for cow crap, and you get an idea of the exaggerated value of
memory performance today.

More does not necessarily mean better. More only means better if you can and do use it, and most software doesn’t.

Going 15 Isn’t Really Better Than Going 10 On The Highway

Main memory is SLOW!!!. To the fastest CPUs out there today, main memory looks like it’s going about 6mph on the freeway. Going to 10 or 15mph on the freeway isn’t going to solve the core problem, which is that CPU have way outrun memory.

CPUs play every kind of trick not to have to go to main memory if at all possible. If you can imagine a helicopter scooping up these cars going at 10mph and whisking them to their destination, you have a rough notion of what CPUs have to do nowadays
to keep themselves occupied most of the time.

The software manufacturers know that, too, and they help that process along by essentially putting their codes in relatively small chunks that the CPUs can wisk away. Sort of like putting Honda Civics rather than big tractor trailers on the road for the helicopter to pick up.

Most importantly, this approach works fine for most current software most of the time. This is the primary reason why DDR doesn’t help much for most applications. Faster memory is still too slow for CPUs. This is a big reason why the current Willamette doesn’t do too hot. For a number of reasons, it depends on accessing main memory more than the Athlon or PIII, and even dual-channel RDRAM is much slower than accessing prefetched data in L1 or L2 cache.

Now yes, there are applications that by their very nature should be able to use all the memory bandwidth they can get. These are the farmers. However, just like farmers in most developed countries, there aren’t too many of them around. Those are the applications where different data must get continually processed, and/or CPU tricks like prefetching don’t work too well. Video editing is a good example of this, games at least theoretically should.

But “should” is not “is.” Unless software is designed to take advantage of fast memory; it won’t take advantage of it. In the case of gaming, some, maybe most games do, but some really don’t. Just look at benchmarking of your favorite game, and see if in fact it does; it’s not automatic.

Bring On A New Stunt Man

The CPU guys are running out of tricks; they’ve used just about all the easy ones that furnish the most gain.

Now it’s the memory guys turn, and that’s just what is happening.

Call it RDRAM, call it DDR; in either case, tricks are being played to make memory faster. They’re just attacking it from different angles, and what people don’t realize is that 1) the different angles eventually converge and 2) neither addresses the underlying problem.

RDRAM depends on fast, narrow transmission. SDRAM depends on slow, wide transmission. What will mostly happen with RDRAM is that the transmission will get wider. With SDRAM, the improvement will mostly be that the transmission will get faster. They will end up at roughly the same point.

This should buy the computer industry a few more years. However, the question that is getting asked more and more is “Why do we need this?” There’s already an answer for some software. There isn’t for most. Not that there can’t be an answer in the future, there just isn’t one now.

Let’s take Word for an example. Do you need a 5GHz processor to run Word? Not as it is right now. But you could add some features to Word that might make it worthwhile; real voice recognition, for instance. Problem with real voice recognition, though, is that not only do you need fast memory, you need a whole lot of it. A Gb of RAM just for that would be a good start. Will people be willing to pay for that much memory just for voice recognition, or would they rather spend less money and keep typing?

It’s going to take quite a bit for most software to be written that can take advantage of faster memory and have the result end up being better than the current helicopter lift. It could well be in at least some cases that it will never be better (at least for the foreseeable future).

At some point, though, the memory guys are going to run out of tricks, too. They probably have enough of them to sustain a 5Ghz processor. Maybe 10Ghz. Above that, they’re going to have to go back to the drawing board.

Not that the drawing board won’t have answers. That’s not the problem. The problem is retooling the whole memory industry to implement whatever comes off that drawing board. The memory folks like evolution, not revolution, for just that reason. A big reason why current memory companies like DDR and don’t much like RDRAM is because DDR is much easier and less costly to implement than RDRAM. But RDRAM looks like a CPU stepping change compared to what the memory guys will eventually have to do.

Some companies are beginning to attack the underlying problem, but whatever ends up coming off that drawing board has to be a radical expensive change, and the longer the memory guys wait, the worse it will be. But they’ll wait.

In the meantime, though, don’t get seduced by sweet-smelling sh**. Make sure it does you some good.

Email Ed

Be the first to comment

Leave a Reply