Are Not CPU's Supposed to Consume Less Power and Generate Less Heat as they Advance?

I have heard people talk of how the operating frequency of CPU's has not been increasing as greatly recently because, as the device operates at a greater frequency, it generates more heat and consumes more energy, making it too difficult for it to operate faster than a certain speed. Similarly, supporters of the new high-bandwidth memory (HBM) have stated that it is vital, because current DDR shall consume more power as it advances, making it impractical after a certain point.

However, I have long been under the impression that, as integrated circuits advance, they generate less heat and consume less power, a key aspect of Moore's law, so why would the opposite phenomenon be happening? Could someone please explain this to me? Thank you very much.

Why has no one responded to this thread? Is there not anything so say on this subject?

Quote:

---

Originally Posted by DemonDragonJ

However, I have long been under the impression that, as integrated circuits advance, they generate less heat and consume less power, a key aspect of Moore's law, so why would the opposite phenomenon be happening?

---

You are essentially correct, but you are missing a key factor: we keep increasing the transistor count.

Moore's Law was an observation that the number of transistors in a dense IC doubles about every 2 years. This has gradually morphed into "the number of transistors which can be placed in a given space/cost/power envelope doubles about every 18 months."

This has been caused by a number of factors, but the key one is improvements in manufacturing techniques allowing for manufacture of smaller and smaller transistors. Smaller transistors take up less space, consume less power, and (unless fundamental technologies change) cost less to make a given quantity.

So yes, as these technologies advance the power consumption and heat production for a given transistor count decreases. However, for most applications there is a lot more value in increasing the transistor count (and thus the computational power). This is why, if you look at the last 10 years of Intel chips, you will see that the TDP for each CPU tier stays roughly the same from generation to generation, while the computational power of those chips increases.

There is a bit of a different curve visible in the mobile space, however, where all three of these factors are important ("mobile device" here being defined as any device which is designed to operate primarily on battery power). What can be seen in this space is that over time computational power increases, but not as rapidly, because power consumption is simultaneously decreasing.

Another interesting pattern can be seen in desktop CPUs. Ten years ago, most of the upper tier desktop CPUs had TDPs in the 130-150W range. Now, most upper tier desktop CPUs are in the 90-95W range. There are still CPUs available in the 130-150W range (Intel even has a few chips with a TDP of 165W), but these are very high end enthusiast or server chips. What this means for the consumer is that a PC of the same market-relative performance can be much smaller and more silent than could be ten years ago.

The question is also very generalized.

Intel's processor suffixes offer processors in every category, ranging from powersucking heatpigs (-X, -K) to power-optimized (-S, -T) to mobile (-M, -H), to power-efficient (-U, -Y), etc. Many of the low-power parts can approach higher performance limits when needed by ramping up power consumption - this is what Xeon racks are all about.

NVidia and those AMD laggards are also always working on higher power efficiency, more from silicon process shrink than from microarchitecture refinements, although they invariably turn around and use newer higher efficiency into increased logic count because their best-selling products cater to performance-hungry niche markets.

The fine people at Samsung, Hynix, STM, Micron, Infineon, Qualcomm, Freescale, Maxim, TI, etc are also always looking at ways to increase power efficiencies. Some of them produce things like SDRAM which requires greater density and performance, others focus on products which operate low-power consumer appliances. Energy-efficient integrated circuitry is a big market and a growing field. Consumers are beginning to realize that more power and more speed and more capacity are not always better, and manufacturers are always eager to buy cheap low-power components for their products.

Certainly a high-end computing platform can consume more power today than ever before. But it can also do things no computer could ever do before because of the astounding number of circuits and logic gates it contains. If we wanted to rebuild old 6502 CPUs on today's dozen-nanometer silicon we could produce individual parts which consume less electricity than can be measured, but good luck getting such a thing to run any modern operating system or application.

I agree with your question. Do you have ant example of integrated circuit which consumes less power and less heat. What I know is almost opposite.

I'm aware of many newer parts which can easily outperform older parts with less power consumption. One example would be Microchip DSPIC33 parts vs Intel 80386DX parts - the DSPIC is a microcontroller, essentially a microprocessor plus entire multifunctional motherboard chipset - the 80386 is a microprocessor, it requires more off-chip circuitry (motherboard chipset) to function - both are 16-/32-bit parts, both can operate at same frequencies (up to ~40MHz), both can address the same RAM capacity (and this RAM is integrated in-package for some DSPIC33 parts), both can compute up to ~40 MIPS, but the DSPIC has about 400K transistors (in the uP alone) and consumes <50mA@3V (<.15W) vs the i386 with 275K transistors consuming ~5W (not counting the ~10W+ consumed by the chipset). These DSPICs could fully emulate a full 386DX processor and motherboard on a card-sized board powered off an AA battery - and it would probably only need to be that size to accommodate all the ports for USB/SATA/etc. And to hammer it home, this particular DSPIC family is itself somewhat elderly and largely obsolete when compared to todays (or yesterday's) kickass little ARM chips and SoC solutions - now it's the cheap sort of part which might be used to control something simple but mass-produced, like a vending machine or a remote control.

A more obvious example: the average multi-core multi-GHz multi-GB converged smartphone device in your pocket today is entirely powered off a little rechargeable lithium battery - a battery which holds something around 5Wh, usually able to run the device all day - and this is the sort of raw computing and display and storage power which a boxy old Pentium-4 desktop could only achieve with a 450W PSU (plus networking hardware).

You won't see as many differences when comparing incremental processor models, even when they scale down to smaller silicon fab. But the differences are readily apparent and quite profound when viewing technologies separated by half-decade timespans.