Slashdot Mirror
SpeedStep On Your Desktop - Intel's Prescott-2M
Kez writes "Intel's Prescott core has undergone a few changes, and the latest version - Prescott-2M - includes new features, one of which is Enhanced SpeedStep technology. Given the jokes about the heat that the Prescott gives out, Intel had to act. It was inevitable that a power (and heat) saving technology such as SpeedStep would find its way into desktop PCs. HEXUS.net has an article looking at the new Prescott-2M based Pentium 4 660 and Extreme Edition 3.74Ghz CPUs, examining their new features and performance."
Slashdot
Dupes on your desktop
Anandtech has another article up, but it emphasizes the increase in L2 cache and the effect this has on performance.
This is a free program that lets you control speedstep in XP (something you could do with windows 2000). I have my laptop set for full performance when on AC and Max Battery when unplugged.
Can I interest you in an article from someone who knows WTF they are talking about?
AnandTech
I don't know what the Hexus kid was on - but I feel safe trusting my reviews to people who have trouble writing big words!
From Page 2: Being LGA775 CPUs, the new processors all look the same. Being press samples that I get the privilege of testing, they're also unmarked with any meaningful information bar the slightly exciting Intel Confidential. So I draw on them. Not quite the Mona Lisa in miniature, mind you, rather an idea of what it is. Any retail example you purchase will be umblemished with my scriblings.
Athlon 64s all have PowerNow technology and have been able to do this from the start to cut power consumption drastically.
Considering that all the sites are now talking about Intel finally catching up by having a similar feature enabled by default, I'd say that it wasn't available before now.
Anyway, power consumption tests on these new Intel processors on other reviews (Tech Report) show that this technology is only useful when you aren't doing any work at all on the processor, when you do stuff, the Intel 6xx processor jumps to 50W-70W higher than an Athlon64 90nm under the same load. So if you are folding or SETIing or whatever, 24 hours a day, and your electricity is 10 cents a unit, you are talking up to $62 more a year in electricity bills.
If you keep a system for three years, a P4 will cost $180 more to run than an A64, and that is certainly something that should be factored into the purchase price for people who like their systems to keep on doing stuff. If you leave it idle overnight, then the cost difference will be a lot less of course, or if your overnight electricity is a lot cheaper then folding at night only is a good choice.
I don't know about Intel CPUs but AMD's standard Windows Cool'n'Quiet driver ignores tasks with "idle" priority. As a result it does not raise the Vcore or clock speed just for SETI@Home unless you tweak the task priority.
I think this is a good choice. This way you can safe energy and donate CPU power at the same time.
Today's experiment
Another Intel 600 series Pentium 4 review from Hardware Analysis.
While I recognize the tone of your comment was meant to be funny, I don't understand the many comments that seem to suggest this type of power management will cause a drop in performance. From my understanding, Intel's Pentium M processor has been using this technology for some time and often performs better than a similarly clocked Pentium 4. Most people simply don't use the full capabilities of their processor most of the time. An architecture that takes advantage of this in order to comsume less energy and run more quietly just makes sense, and there is no reason to believe demanding consumers will take a performance hit.
Simple: Most algorithms are not easily split into independant parallel threads. If they could be, then previous (smaller) attempts to integrate vector processing (like MMX and SSE) would have had larger impacts than they did. Plus, many of the obvious vector operations have in fact been offloaded - to the GPU.
90nm AMD64 appears to use under 40W even at 3500+ rating, at full load. This has been confirmed on many sites.
Simply put, it's because effective parallel code isn't simple to produce.
I have a Pentium-M laptop and I will explain how SpeedStep works in practice.
SpeedStep itself allows the CPU to change frequency (and voltage along with it) to save power. For example, my 1.7 GHz Pentium-M can run at 600, 800, 1000, 1200, 1400, 1500, or 1700 MHz. This is usually controlled by turning up the speed when the CPU usage is high and turning it down when the usage is low. Thus, if you are playing Doom, it will run as fast as it needs to to keep up. These "performance" states are described as "P-states" in the ACPI spec (they run from P0 [full speed] to P7 [slowest speed] on my computer)
There is a second, older type of throttling. that is just called throttling, and described as "T-states" in the ACPI spec. On older Pentium III's, there are two states, T0 (on all the time) and T1 (off half the time). On my Pentium-M, it goes from T0 (on all the time) to T7 (on 1/8 of the time) in 1/8 time increments. This throttling is much less efficient (it has to start and stop the CPU constantly, and still runs the clock at 1.7 GHz when it's on) but is used for a different purpose, as you will see.
ACPI "thermal zones" are objects that consist of a temperature, and "trip points" (temperatures) that trigger "active" and "passive" cooling. Active cooling is set at a lower temperature, and is linked to a fan object that the OS should turn on. Passive cooling is set at a higher temperature, and is linked to the processor object. When the temperature passes the "passive" threshold, the CPU is throttled using T-states.
I actually have a bunch of data about CPU speed (P-state) and throttling (T-state) versus temperature and power usage, and I can tell you that both types of throttling save battery power and run cooler. However, P-states are much more efficient. If you take a 1.7 GHz processor and run it at P0 and T7 (1.7 GHz on 1/8 of the time = about 215 MHz) it runs almost 20 degrees hotter and uses up about 5 Watts (the lowest usage I recorded was about 12.5 Watts, so that's a large fraction) more than running at P7 and T0 (600 MHz on all the time). It's also 1/3 the speed. So basically, P-states are much more efficient, but T-states are what is tied to cooling, probably because they existed first.
The unfortunate problem here is that P-states are much more efficient, but traditionally P-states are tied to usage and T-states are tied to temperature. It is often suggested to use T-states once you are in the bottom P-state (i.e. go 1700*8/8, 1500*8/8, ... 800*8/8, 600*8/8, 600*7/8 ... 600*2/8, 600*1/8) but frankly that doesn't save much power, and does hurt the responsiveness of the computer. (It needs a certain minimum speed to be able to speed itself back up in time to not look laggy...)
The best thing to do if your processor thermal-throttles itself is to 1. cool it better (perhaps attach an air conditioner to the side of the case?), 2. turn down the speed--voluntary throttling may sound like a waste, but it keeps the temp down better than letting the OS throttle it, and it gives better performance, or 3. get a Pentium-M.
I haven't actually been able to compare my Pentium-M to a Pentium 4, since I avoid those like the plague (1.5 hours battery life? how about 5?), but I can say that subjectively, it's quite snappy (thanks to 2 MB L2 cache and Linux's good disk caching) and doesn't show its slower clock except in raw processing work. (If anyone wants me to time a kernel build, email me and I'll do it, you pick the version and .config...)
Tired of free iPod sigs? Subscribe to my blacklist
Actually what's more interesting than the SpeedStep thing, is the fact that the 600 series uses 25W less power than the 500 series on full load ( that's something when you consider that there is 1M more memory on the new chip).
Up til now Intel's 90nm process was a huge failure because of the heating problems and forced Intel to abandon their plans to hike speeds above 3.8GHz