The purpose of this thread is to discover cheap ways of running BOINC as much as possible. That could mean finding the lowest wattage CPU, or the cheapest PC just for BOINC.

A friend sent me this link to a $200 desktop with a VIA C7-D 1.5 GHz CPU, which uses 12 watts! $200 PC

Now that's computing on the cheap! That's 128 MHz/watt. Very good.

Here's how to calculate how many MHz you get per watt of power:

[CPU SPEED IN GHZ] x 1024 / watts = MHz per watt
If the CPU is only in MHz, just do this: MHz/Watts = MHz per watt.

Obviously the more MHz per watt, the better the value (more speed for less money per CPU).

Don't know the wattage of your CPU? Check out this chart: CPU Wattage

Clock speed is not everything in crunching. In fact I would recommend strongly against any of the Via CPUs. They tend to be much slower in BOINC projects than the raw speed would indicate.

---

BOINC WIKI

BOINCing since 2002/12/8

Clock speed is not everything in crunching. In fact I would recommend strongly against any of the Via CPUs. They tend to be much slower in BOINC projects than the raw speed would indicate.

You indicated that clock speed is not the only factor involved in determining a CPU's ability to crunch BOINC work units. Can you detail what other factors are involved in negating the importance of a CPU's clock speed?

Perhaps explaining this would also explain why you believe a VIA CPU is "much slower" than the clock speed indicates.

I'm glad you brought this up, because although my post was about finding CPUs with the most MHz per watt, you brought up the point of how the clock speed may not be as important as my post may have suggested.

Your post now has me wondering what features of a CPU might make it better suited for crunching BOINC work units, other than raw clock speed.

For most BOINC projects an effcient FPU is a prime consideration.

If the on-dye cache is large enough to hold the project's working set it can make a huge difference.

Number of cycles per instruction.

I have a VIA at 800mhz it takes about twice as long to process tasks as my 500mhz PIII did. The CPU may have changed since that one. However they are designed for low heat generation not intensive processing. It is a good processor for surfing the web and embedded devices though.

My best computer for crunching is a intel core 2 running at 1.6ghz. It beats out a P4 @ 2.0ghz, a celeron @ 2.8ghz and an AMD X2 @ 1.8ghz.

---

BOINC WIKI

BOINCing since 2002/12/8

The more processor cores the better. At the moment the Q6600 might be the cheapest processor when you look at how much computer power you get?

It'd be many many times faster than the Via. As Keck says, MHz is very misleading.

---

For most BOINC projects an effcient FPU is a prime consideration.

If the on-dye cache is large enough to hold the project's working set it can make a huge difference.

Number of cycles per instruction.

I have a VIA at 800mhz it takes about twice as long to process tasks as my 500mhz PIII did. The CPU may have changed since that one. However they are designed for low heat generation not intensive processing. It is a good processor for surfing the web and embedded devices though.

My best computer for crunching is a intel core 2 running at 1.6ghz. It beats out a P4 @ 2.0ghz, a celeron @ 2.8ghz and an AMD X2 @ 1.8ghz.

In layman's terms, can you explain how I would go about looking for a CPU with these features?

How can I find out if a CPU has an efficient Floating Point Unit?

By "on-dye" cache, do you mean the registers, or L1/L2 cache? How big should they be for BOINC?

This is where I get a little lost: CPI or Cycles Per Instruction. I found an explanation on Wikipedia.org but could you tell me how I can apply this to BOINC?

By the way, I know what you mean when you said your Intel Core 2 1.6 GHz is actually better at crunching BOINC than higher-clocked CPUs. I too have a Pentium 4 at 2.6 GHz and it's actually slower than my laptop with an Intel Core Duo 1.73 GHz. I wish I knew how this is possible. Maybe this thread will help explain that.

In layman's terms, can you explain how I would go about looking for a CPU with these features?

Not that simple. For the VIAs compared to other Boinc cruncher, according to the computing efficiency, you can try to look at them on the statistics sites, like BoincStats:

Host CPU breakdown, sorted on average credit per CPU second (Total 3529 CPUs):

Here are the VIAs' values to look for (few examples):

<pre style="white-space:pre-wrap; ">Pos.  CPU                               Average credit per CPU second
0039  Intel Core2 Quad Q6600 @2.40GHz         0.005240 
0076  Intel Core2 Quad CPU @2.40GHz           0.004513 
0096  Intel Core2 CPU 6600 @2.40GHz           0.004246  
2241  VIA/IDT Unknown                         0.000835 
3115  VIA C3 Nehemiah+RNG+ACE                 0.000417  
3165  VIA C3 Ezra                             0.000389  
3225  VIA C3 Nehemiah+RNG                     0.000358  
3393  VIA C3 Samuel                           0.000201  
3447  VIA Samuel ("CentaurHauls" 686-class)   0.000143  </pre>

(Is the mentioned VIA C7-D 1.5 GHz here (with some other name) or not yet?)

Off course, the power consumption is not mentioned here. Would be interesting to see in the table :-)

Peter

A Pentium runs faster than a Celeron at the same clock speed. It may have some advanced instruction sets that the Celeron doesn't have, so an application that uses them would have to fallback to the "normal way" of doing those calculations, which is slower (the whole point of those advanced instruction sets is doing it faster). It may have a bigger cache. Or better algorithms for choosing what gets cached, or higher cache-CPU bandwidth. Definitely has much better "pre-fetch" systems, where it can predict what data it will need and fetch it into the cache whenever it can, before it's needed, and *while* it's computing something else. And definitely has a better instruction pipeline.

(don't even bother correcting me, I *know* I'm wrong in at least *something* I said above :P)

It's *so* much more than just clock speed...

EDIT: also, you need to take into account the power consumption, not only the computer price. Electricity has to be paid for too. :)

---

EDIT: also, you need to take into account the power consumption, not only the computer price. Electricity has to be paid for too. :)

My original post was as much about power consumption as it was about CPU speed.

This is the flaw in your question :

The purpose of this thread is to discover cheap ways of running BOINC as much as possible. That could mean finding the lowest wattage CPU, or the cheapest PC just for BOINC. ...

This part cannot be answered because BOINC itself has fairly low requirements.

VIA CPUs can be very efficient (CPU power vs. power consumption) under BOINC, but only combined with certain projects. The project clients are the ones that have certain needs, not BOINC.

They will be lost completely with QMC@Home, Einstein or CPDN.

They will be competetive on SIMAP.

So if your main project would be SIMAP and SIMAP would be a full time project, a C7 box would be a fairly good choice.

If you find a second project that has similar requirements as SIMAP for the time when SIMAP has no WUs, a C7 will be a cheap cruncher (energy-wise).

p.s.: Yoyo@home might be such a secondary project with not too high requirements. I will ask.

By the way, I know what you mean when you said your Intel Core 2 1.6 GHz is actually better at crunching BOINC than higher-clocked CPUs. I too have a Pentium 4 at 2.6 GHz and it's actually slower than my laptop with an Intel Core Duo 1.73 GHz. I wish I knew how this is possible. Maybe this thread will help explain that.

The core and core2 architectures have much better CPI ratios than the P4 architecture. They also have shorter pipelines so a pipeline stall doesn't cost as much. I think they also have L3 cache on the dye, the P4 either does not have L3 cache or it uses main memory for it. That is why the core/core2 is faster, without even considering the extra CPU(s) the core chips usually have.

As Ananas said it also depends on the project. There are a few that do not depend heavily on the FPU. On these projects a VIA CPU is respectable.

CPI applies to all processing tasks not just BOINC. In simple terms CPU A takes one cycle for a simple add and CPU B takes 2 cycles for the same add. If CPU B is clocked twice as fast as CPU A then the processing speed will be roughly equal.

---

BOINC WIKI

BOINCing since 2002/12/8

Hi,
requirements for yoyo@home are very less, the cruncher application does require less memory and does mainly integer operations. I don't know which kind of processor is best for it. I think mainly a lot or all math projects do mainly integer operations and uses less memory.
yoyo

---

Germany biggest distributed computing community Rechenkraft.net

Actually the Yoyo@home math application knows the VIA C7 since version 2.9013-498 which means they can choose the best method for it without benchmarking. The efficiency of distributed.net is very good because they have several math cores in their application and can adapt ideal to all kinds of processors (including the SPEs of Cell!).

I just found the perfect link! It's all about BOINC performance - how to increase it and what hardware features make BOINC crunch faster.

BOINC PERFORMANCE FAQ

the crunching speed is also heavily project dependant. For example, Rosetta doesn't make use of SSE extensions while some projects do, as well as their general FPU and cache dependancies etc.

However, if you want to get good performance per watt, i don't think much analysis would be needed to pick out the Q6600 as mentioned. It can also be overclocked, although OCing increases the power consumption (not linearly) and so there is a point where the gains from OCing will be relatively small in comparison to the increased energy needed to run at that speed.

Another *free* option/factor that can make a big difference to the efficiency is undervolting. If you're running a CPU at stock speeds then reducing the voltage of the CPU can reduce it's power requirements massively. I reduced my laptop CPU (1.86GHz Pentium-M (Dothan core)) from 1.325V down to 1.053V using RMClock and the battery lasts around a third longer (while running Rosetta).

Crunchers also don't need hard drives - you can net boot, or my preference is to boot from compactflash cards. My media centre boots Windows MCE 2005 from a 2GB CF card - TV is recorded to the server over the network.

Other things that help reduce electrical consumption are onboard video, no keyboard/mouse, good airflow (less electrical resistance at lower temperatures), few fans, and not running where air-con is needed!

HTH
Danny

---

the crunching speed is also heavily project dependant. For example, Rosetta doesn't make use of SSE extensions while some projects do, as well as their general FPU and cache dependancies etc.

However, if you want to get good performance per watt, i don't think much analysis would be needed to pick out the Q6600 as mentioned. It can also be overclocked, although OCing increases the power consumption (not linearly) and so there is a point where the gains from OCing will be relatively small in comparison to the increased energy needed to run at that speed.

Another *free* option/factor that can make a big difference to the efficiency is undervolting. If you're running a CPU at stock speeds then reducing the voltage of the CPU can reduce it's power requirements massively. I reduced my laptop CPU (1.86GHz Pentium-M (Dothan core)) from 1.325V down to 1.053V using RMClock and the battery lasts around a third longer (while running Rosetta).

Crunchers also don't need hard drives - you can net boot, or my preference is to boot from compactflash cards. My media centre boots Windows MCE 2005 from a 2GB CF card - TV is recorded to the server over the network.

Other things that help reduce electrical consumption are onboard video, no keyboard/mouse, good airflow (less electrical resistance at lower temperatures), few fans, and not running where air-con is needed!

HTH
Danny

Thank you Danny, for your most relevant message! You've made lots of good points.

How can I determine a project's FPU and cache dependancies?

You mentioned the Q6600 as having good performance per watt. I did a comparision of the Q6600 with the Core 2 Duo E4500, and the E4500 actually gives you more speed per dollar and per watt.

E4500 2.20 GHz:
65 Watts, which comes out to be 0.0338 GHz per Watt;
$125, comes out to be 0.0183 GHz per Dollar.

On the other hand, the Q6600 at 2.4 GHz:
95 Watts, comes out to be 0.0252 GHz per Watt;
$280, comes out to be 0.0085 per Dollar.

So with those criteria, it looks like the E4500 gives you more speed per initial dollar investment and for long term power requirements than the Q6600.

The whole point of me calculating this and doing CPU comparisons is to give people the information they need in order to build the cheapest possible PC for running BOINC on. So I hope they find this useful.

You said crunchers don't need hard drives. Interesting! Did you mention this to point as another way of reducing costs per PC?

But the Q6600 crunches 4 models at a time, as against 2 for the E4500.

E4500 2.20 GHz:
65 Watts, which comes out to be 0.0338 GHz per Watt;
$125, comes out to be 0.0183 GHz per Dollar.

On the other hand, the Q6600 at 2.4 GHz:
95 Watts, comes out to be 0.0252 GHz per Watt;
$280, comes out to be 0.0085 per Dollar.

The Q6600 basically consists of two E6600s in a single package, so as Les says it can do twice as much work as a dual core processor.

If I recall correctly, the E4500 has a small cache and a slower front-side-bus, and half the cores of the Q6600. So in reality it has about a third of the throughput of the Q6600. And you're not taking into account the electrical overhead of the rest of the system it fits in.

---

E4500 2.20 GHz:
65 Watts, which comes out to be 0.0338 GHz per Watt;
$125, comes out to be 0.0183 GHz per Dollar.

On the other hand, the Q6600 at 2.4 GHz:
95 Watts, comes out to be 0.0252 GHz per Watt;
$280, comes out to be 0.0085 per Dollar.

The Q6600 basically consists of two E6600s in a single package, so as Les says it can do twice as much work as a dual core processor.

If I recall correctly, the E4500 has a small cache and a slower front-side-bus, and half the cores of the Q6600. So in reality it has about a third of the throughput of the Q6600. And you're not taking into account the electrical overhead of the rest of the system it fits in.

I purposely left out the electrical overhead of the PC because I can't possibly know what system people will put these CPUs in, but I do know what power the CPU uses, so that's what I talk about.

I have now taken into account the combined GHz of the cores of both CPUs (the Core 2 Duo and Core 2 Quad), and the Core 2 Quad gives you more GHz per Watt, but slightly less GHz per Dollar.

So it turns out the Core 2 Quad gives you more value, but at a higher initial cost and at higher wattage.

So with that, it comes down to a personal choice between the two.

Core 2 Duo: 4.4 GHz combined cores:
0.0676 GHz per Watt;
0.0352 GHz per Dollar.

Core 2 Quad: 9.6 GHz combined cores
0.1010 GHz per Watt;
0.0342 GHz per Dollar.

Why don't you do the calculations for the FSB and cache and compare the two for the dollar amount of each?

Core 2 Duo: 4.4 GHz combined cores:
0.0676 GHz per Watt;
0.0352 GHz per Dollar.

Core 2 Quad: 9.6 GHz combined cores
0.1010 GHz per Watt;
0.0342 GHz per Dollar.

Why don't you do the calculations for the FSB and cache and compare the two for the dollar amount of each?

For the Boinc purposes, it would be simpler to calculate e.g. "RAC per Watt" and "RAC per Dollar".

Let's throw the ball. My C2D T7200 clocked mostly at 1.33-1.66 GHz (trade-off between speed and fan loudness) seem to earn 360 CS daily, or 15 CS hourly. At 1.33 GHz the whole machine consumes between 49-54 Watt (depends on the display backlight off/on), thus 0.3 CS/Watt. For CS/Dollar, put your own numbers in to calculate it.

Peter