Efficiency vs. low-power CPUs

[Glorious1]

I've read the recommendation from several of the glitterati on this board not to get low-power CPUs. The argument is the high-power ones work faster and do more, but they all use the same low power when idle.

I don't have any knowledge to challenge that, but I am curious. An analogy with cars comes to mind. Think of a Honda Fit/Jazz vs. a monster pickup truck or Maserati. Both use no gas when parked, but the truck and Maserati either carry a lot more or get you there a lot faster.

Of course, the truck and sports car use a lot more gas than the Honda Fit to get you from point A to point B. And it's the gas used per distance traveled that we're concerned about, not how much is used while parked.

So the question is really about efficiency. Do low-power CPUs, or any of them, do calculations and push bits around more efficiently than high-power CPUs? Granted that a given task would take longer, but do they use less power to get it done?

 

[titan_rw]

I see two different things here.

1: Thermally limited CPU's. (T suffix?). Car analogy: Engine was designed to put out 200hp, but to do so it needs to put out 200 'units' of heat. Specialty car maker is making a really small car and can only manage a radiator / cooling system to handle 100 'units' of heat. Engine is limited to 100hp so as to be within thermal design limits of the rest of the car. This engine isn't any more efficient. Will use just as much gas idling. And may take longer to get real work done (assuming 100 hp isn't enough to travel through the mountains. It's max fuel usage is less, but it'll take longer to get there, so net result is about the same). Back to PCs: If your cpu heatsink / fan / case design can handle the heat output of the 'normal' cpu, why bother with the limited one?

2: Low power 'efficient' cpus like the intel avoton. As far as I'm aware they achieve a better 'performance per watt' metric, but have a lower total performance potential than regular socket 1150 cpus. Car analogy: Moped / scooter I guess? Uses little gas while idling. Still does city speeds. Can't travel on freeway or haul very much cargo. (PC: Lacking features like HT, AES, etc. Takes longer to do same tasks).

 

[DKarnov]

The key point lies in how the 'low power' variants actually differ from the conventional parts:

http://www.anandtech.com/show/8774/...ew-core-i3-4130t-i5-4570s-and-i7-4790s-tested

Simply put, T is used for 35W/45W models and S is for 65W models. The S models tend to have similar single core performance to their normal cousins, but reach 65W by reducing the multicore frequency when the CPU is loaded. The T models reduce both single core and multicore frequency to hit their 35W or 45W, but tend to be a little more esoteric in their position in the stack.

There is no mechanical difference; only changes in the frequency vs power control. So at idle or low load, the chips behave exactly the same; it's only when high load is approached that the clocks are dropped. All multicore Intel chips tweak their core freqs based on total load, the S / T chips are just more aggressive about it. It is vital to understand this is not about drawing less power, it is about generating less heat. The S / T chips are for embedded / super compact applications (automotive, industrial, display etc) and essentially no NAS should ever be so compromised as to need them (because your hard drives would fry in short order.)

It is possible for chips to be more efficient. The common methods are:

- Process shrink
- Improved processing pipelines (think better branch prediction etc)
- More advanced power management (ex. Haswell's new C states, although this goes pretty deep.)

The first one and the second two essentially represent the Intel tick / tock development cycle. Thus, all other things (clock, number of cores etc) a newer generation chip will be more efficient for a given processing task. The S and T chips are at base just Haswells, so they're the same as regular Haswells efficiency wise; we'll need desktop Broadwells to see real improvement.

 

[Ericloewe]

Ah, but your analogy is flawed!

You're comparing "small" cars with "large" cars. The recommendations you've read are actually between "big car with the engine artificially limited" and "big car with engine not artificially limited".

An Atom is more efficient than a Xeon (well, probably at least for low loads), but if you take the exact same die and give it two different microcodes, the low-power one won't be more efficient as long as both manage the workload (it gets murkier when the non-limited one gets done more quickly).

Before we look at a concrete example, important information I just discovered:

The regular Xeon E3-1220 v3 has 8MB of L3 cache - yet the low-power Xeon E3-1220L v3 only has 4MB! You don't magically save 67W.

Now, back to business:

Xeon E3-1240 v3: Base clock is 3.4GHz, Boost is 3.8GHz
Xeon E3-1240L v3: Base clock is 2.0GHz, Boost is 3.0GHz

Both of these have the exact same die! Of course, it's possible that Intel harvests processors that behave better at low loads for the Ls, but this is unconfirmed and is most likely negligible.
It's important to keep in mind that "Base Clock" does not mean the CPU will spend most of its time at that clock - both will Idle down to nearly nothing.

 

[mjws00]

That's exactly why there is little gain. The CPU's are not more efficient. The bits are pushed the same way. They just throttle back when asked to worked hard so they don't draw as much power. So there is a tiny potential gain in terms of heat produced, and also advantages controlling current draw if we are on battery. None of these benefit a NAS workload.

EDIT: LOL. Tons of answers before I hit post.

 

[Ericloewe]

EDIT: LOL. Tons of answers before I hit post.

Join the club - that was fast!