# Is exponential growth of computing power still valid for desktop computers?

Is exponential growth of desktop computing power not valid anymore? Today, it looks more like a logarithmic growth.

For example, at the beginning of 1990th a typical desktop CPU was 20..50 MIPS (386,486), then in 2000 it was 1000..3000 MIPS (Pentium III, Athlon), but now they are 10000..40000 MÌPS only (Quad/Dual Cores). It means between 1990 and 2000 we've seen 100-times growth, but between 2000 and 2010 it was only 10-times growth per 10 year.

The same holds if we compare GFLOPS per Chip. In the 1992 Intel 486DX was 0.03 GFLOPS, in 2000 PentiumIII was 2 GFLOPS, and now 2010 most desktop CPUs are of 20..30 GFLOPS.

Now, let's look at the non-volatile storage: <40Mbyte at 1990> vs. <10Gb at 2000> vs. <1000Gb at 2010>. So, 1000-times vs. 100-times growth per decade.

Moreover, a volatile storage: <2Mb at 1990> vs. <256Mb at 2000> vs. about 2Gb (or more, but not 20!) at 2010. So it gives 100 vs. 10-fold growth per decade.

Does this empirical evidence means that the growth of the desktop computing power becomes slower?

-

Well, eventually, yes.

Hmm sounds like Law of Diminishing Returns vs Moore's Law

If I put my economist hat on, I'd say that the Law of Diminishing Returns is eventually going to win. It does in practically every other human endeavour. While I have respect for Gordon Moore's achievements, his law is more of a strategic business vision that can only apply to circuit density for a finite time, and does not describe the world overall. The Law of Diminishing Returns has a better track record.

If I put my IT hat on, it doesn't fit. Cause it gets bigger every year :-)

Danger: This question might be too speculative or contentious for the moderators.

-

I think you're confused about "logarithmic" versus "exponential". Exponential growth is shown as a straight line on a logarithmic scale. I also think you have to factor cost into "power" comparisons. When you do this, you'll see that the growth continues to be exponential.

The number of transistors in processors doubling every two years is what Moore's law originally referred to and it has stayed true so far and will continue to for the foreseeable future. It is likely, however, that some other technology will take the place of transistors at some point.

-
No, I am sure what is log(x) vs. exp(x) :) By the way, Moore's Law is Intel's promotional trick and has nothing to do with this discussion. I don't think that transistors number is more important than computational power for the end user. – psihodelia Feb 12 '10 at 15:42
@psihodelia: Where did the idea of exponential growth of computing power come from if not Moore's Law? – Dennis Williamson Feb 12 '10 at 15:59
By the way, exp(log(x)) == log(exp(x)) – Dennis Williamson Feb 12 '10 at 16:03

I've been holding out on this for a bit but I'll throw in a couple of points.

The average amount of RAM you specify for 2000 is out by a bit, 32\64Meg would have been more accurate and with those values the ratios for 1990-2000 & 2000-2010 are more or less the same.

MIPS & GFlops as measured from general purpose processors are pretty useless measures of computing power. Systems are not designed to maximise those numbers, certainly general purpose desktop type systems aren't.

Hard disk capacity has increased by a factor of 200k (5Meg to 2TB) in the 35 years or so since I first got a system with a hard disk but the random access latency has only improved by a factor of around 10-15 because of mechanical limits and the interface bandwidth has only increased by around a factor of 200 (24 Megabits/sec for the early 90's ESDI to 6Gbps for SATA II). All of this is going to be turned on its head though with SSD's. Ignore the teething problems, people said HDD's would never replace tape too back in the day, but the improvement in latency that SSD's provide will fundamentally change the way storage is built into systems over the next three years. HDD's are not going away but spinning platters will only be used for write-rarely, read occasionally long term storage by the middle of this decade. The difference that will make to day to day computing experiences, once the OS designers stop having to workaround that HDD latency wall that hasn't moved noticeably in 20 years, will be immense. And remember all of that is due to the steady exponential increase in transistor density that has finally begun to brought solid state storage within striking distance of spinning disks.

All of this misses the point substantially though. You can't condense the effectiveness of a modern computer down to a handful of metrics that can be directly compared to those from systems 20 years ago and make any really meaningful comparisons.

One metric I will share with you is that in my field I do a lot of Server consolidation work where I replace large numbers of 3-5 year old servers with far fewer new boxes that are inherently not that much "bigger". We typically consolidate on a 10-15 ratio depending on the load but I'd have no problem saying that I can take 10 5 year old dual socket servers that are moderately busy (~50% utilization) that would have cost ~\$6k new with 1 dual socket (but now 8 core) server that costs ~\$6k and expect it to host all of those without breaking past 50% itself. We also do workstation virtualization work where we will happily double that ratio without being too concerned. I'm absolutely confident saying that a typical system today is between 10 and 15 times better in terms of any metric that counts than a 4-5 year old system from a comparable segment. What's more we're doing all of this with a lower thermal\power draw per device so not only is "red line" performance increasing for general purpose computing tasks but it is steadily improving in terms of energy efficiency at the same time.

-
just FYI, SSDs are good for reading, but offer poor writing performance. Spinning platters are still better. – gbjbaanb Feb 12 '10 at 18:34
That is true today (teething problems as I said) but it also depends on the SSD and whether your OS understands them well enough. A well designed SSD running with an SSD aware controller can sustain 10k small write random IOPS with ease, unfortunately those still cost a couple \$K a pop but don't worry those bugs are being ironed out very fast. – Helvick Feb 13 '10 at 0:25

Moore's Law referred to the number of transistors available. So, getting a couple of data points from Wikipedia, typical iAPX chips at decade intervals

• 1980: 8088 0.029M transistors
• 1990: 80486DX 1.2M transistors
• 2000: Pentium III Coppermine 28.1M transistors
• 2010: Core i7 Bloomfield 781M transistors

so, 1980-1990, factor of 40 1990-2000, factor of 25 2000-2010, factor of 27

If we weasel a bit, and note that the 1980's CPU actually had a second chip for floating point, we get a per decade capability increase (transistor count) of around 25x pretty consistently. So I don't see (yet) transistor growth in the typical desktop computer slowing.

-

It kind of depends what you mean by 'growth of desktop power'. Sure, CPUs are still getting more and more powerful, but the overall performance of them isn't showing through.

One example of this is caches - whilst your CPU can crunch through a billion instructions per second, all it takes is 1 little programming exception to strike all the cache off and make the CPU fetch the instructions and data from main RAM - which is incredibly slow (by your CPUs standards). So unless you can keep your CPUs 'fed' with data, its going to be twiddling its metaphorical thumbs, biting its fingernails and reading the newspaper. And your users will be throwing theiur hands up exclaiming "I bought a super new PC and it still takes ages to do stuff, what is the damn thing doing" (as the disk thrashes and the CPU cache lines slosh between threads, and the northbridge overheats transferring RAM from main memory to CPU L3 cache).

And then we put our data on the other side of the network....

So... yes - they're getting more powerful. No - they're still just as slow as they used to be.

-