8

We know that the ARM chip came out of the Acorn computing initiative.

In the book The One Device, we read:

The low-power big thing that the ARM is most valued for today, the reason that it's on all your mobile phones, was a complete accident," Wilson says. "It was ten times lower than Steve had expected. That's really the result of not having the right sort of tools."

The implication is that (a) the lower power design of the ARM was discovered by accident, and (b) the right set of tools would have lead to a predictable outcome.

This seems strange. One would have thought you could predict the power consumption of a device using physics and scientific models.

My question is: Why isn't the approach to lower power CPUs guided by design rather than accident?

2

Design is one factor determining the actual power use of your device - the other is the semiconductor process used and its tolerances and variations over the product lifecycle.

Especially with whatever process is cutting edge at the time, there is no "either it doesnt work or it works and consumes exactly n watts of power in a given state" - more of a continuum between "part either doesnt work at all or is scrap because it uses power above the maximum stated in the specifications" via "this won't overconsume and/or overheat if not used above this clock frequency, sell it as a lower speed part" to "works perfectly, speed is maximum, and power consumption is on the low end". There will be days like "The machinery isn't at its best calibration today, guess we will be making a lot of lower speed parts today..." in production...

A given design of a CPU is often undergoing "die shrinks" over the product lifecycle, which are in the end adaptations to a new and improved semiconductor process. Depending on the design and the processes, a change of process can lower or raise power consumption - eg a process using smaller structures could lower gate capacitances and thus the power consumed when data changes inside the device, but worsen leakage currents that will mean more power is consumed when certain non-changing states of logic levels are present. A design taking advantage of peculiarities of a given process could well end up taking disadvantage of peculiarities of another...

| improve this answer | |
  • Yeah this is what I was looking for in the question - but can't you measure and to some degree take advantage of those quirks in your planning to get a better outcome? – hawkeye Sep 17 '17 at 8:47
  • 1
    "to some degree" - even today, experience values often get the better of planning - compare Intel's Netburst architecture that hit a dead end due to power consumption/heat in practice.... – rackandboneman Sep 17 '17 at 8:52
11

The question statement is faulty. The approach to low power CPU design is guided by tools. Today.

However, those tools did not always exist, or, when they eventually did exist, were too expensive to be used by small chip development teams who could not afford to buy time on a Cray or large mainframe needed to analyze the physical models. (Plus, the models themselves were still being developed for new fabrication processes.)

At the time most 8-bit and early 16-bit microprocessors were designed, Spice, and other circuit simulation tools, IIRC, could simulate on the order of a few dozen transistors, far fewer than needed for a full CPU. So total power estimation was done by engineers with some experience in integrated circuit design, not by tools. Sometimes those estimates were off.

Same with performance estimation (before the days of fast enough logic simulation). So the performance/power estimates could be even farther off until the first chips were tested with real application code.

Note that HP (with far greater engineering resources than Acorn) did design very low-power processors for the first battery powered scientific calculators (HP-35, et.al.), but with magnitudes less raw performance than personal computer microprocessors of that era.

| improve this answer | |
  • 1
    The irony of bringing up Cray (and his 115 kW computer system) in a discussion of low power CPU design brought a smile to my face. :) – Jules Sep 11 '17 at 22:41
  • Apple purchased a Cray (a YMP, IIRC) around the same time they were first experimenting with ARM processor chips. Around that time, Dave Needle (former Amiga engineer) was involved at Apple ifor a pre-Newton ARM post Apple IIgs follow-on product, later cancelled. – hotpaw2 Sep 11 '17 at 22:54
  • I would guess (that's why it's a comment) that the existing tools were geared toward implementing holding state using the "flip-flop with synchronous enable" mechanism, and ARM developers happened to use clock gating in their manual design. – Leo B. Sep 12 '17 at 3:16
  • @hotpaw2 Irrelevant to the question, but that was probably the only Cray with a psychedelic paint job on the casing. Given the price tag, you could order your YMP in any color cabinet you wanted. Apple's was a lurid purple, then overpainted by hand. Source: I got that story (and saw the pictures!) straight from our company rep in the Cray marketing department! – alephzero Sep 16 '17 at 22:40
  • In the 90s HP made 8088 msdos palmtop PCs that could run one month on two AA cells (actually I used one this year, it lasts almost two months with rechargeables AAs) – Emmanuel Oct 10 '17 at 15:20
4

It is, as the citation already points out, if you use professional tools.

Keep in mind that the first ARM developments should be better described as a hobbyists aproach. They wrote their own tools on BBC Mico systems using BASIC and Assembler. It wasn't until they closed in for real production, when the ARM team had to transfer their ideas into professional toolsets.

| improve this answer | |
  • ... and one would dare imagine that the emergent low-power use case was being designed in at least as early as the Apple collaboration, so that probably shut the door to targeted competitors. This I speculate. – Tommy Sep 10 '17 at 15:45
3

When RISC technology was developed, power usage wasn't the same concern as it is today. Smartphones had not been invented yet, electricity was cheaper, Moore's Law still had plenty of life in it, and global warming was not yet on people's minds. Heat management was a concern, but reducing power draw is only one way to prevent overheating. The main goal back then was to design CPUs with increasing clock rates, so that's where most of the effort went. Other achievements were often just happy accidents that occurred while working toward that goal.

| improve this answer | |
  • 3
    From the days of room-filling mainframes, power has been a concern. Every extra kWh costs you - both for the electricity itself and for the extra cooling required. In addition, CPU design, except for some marketers, has not been about fastest clock rate but about getting the most productive work (calculations, decision trees, database manipulation, etc.) in the least amount of time. If it was ever about clock rates then RISC would have won a long time ago. But CISC has benefits too, with a tradeoff of how much can get done in one instruction vs. how many instructions per unit of time. – manassehkatz-Moving 2 Codidact Sep 11 '17 at 5:07
  • 1
    Power was always a concern in making sure you had enough and also enough cooling. Back in the 80's and 90's when I had more contact with mainframes, the cost of power was never really a concern - it was quite a small component of running a large computer system. – JeremyP Sep 11 '17 at 9:36
  • 2
    @manassehkatz There was a period in history when it was all about clock rates for microcomputers. At that time, the x86 architecture was dominant (so RISC never had a chance) and the MHz/GHz number for the processor was a simple easily comparable number for marketing purposes.The whole thing hit the buffers with the Pentium 4 architecture, which many people saw as a cynical attempt to exploit the weakness of GHz as a measure and which demonstrated that raw GHz does not translate to processing power directly. – JeremyP Sep 11 '17 at 9:45
  • 2
    The relative costs of power versus fabrication costs have changed over time. Power and cooling costs (especially of data centers) have gone up, while the price of making fast transistors has gone down by zillions. That greatly changes design trade-offs. – hotpaw2 Sep 11 '17 at 15:45
  • 3
    @manassehkatz - ARM was explicitly designed for use in home and office computer systems, against the background of the rapid expansion of those fields in the early 80s. Power consumption was simply not a factor anybody cared about for those applications at the time. And another explicit design goal of ARM was to increase memory bandwidth over existing PC-targeted processors (e.g. 68000 et al) in order to allow for better graphics performance. With processor speed and bus speed being intrinsically linked at the time, there were only two options to achieve that: wider bus, or faster clocks. – Jules Sep 11 '17 at 22:54
0

What he means is that the tools for designing CPUs were not able to give accurate estimates of power consumption. So when the power use was so low, it came as a surprise.

Modern tools not only estimate power consumption, they help the designer optimise it.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.