Low-power T processor vs. Unthrottled CPU for Plex Encoding


Recommended Posts

... PMS throttles the load and produces a steady stream matching the framerate of the outputted video. It does not go ahead any more than the buffering requires ...

 

This comment clarifies what you've been trying to say.    Given the effective "governor" of PMS, there's no real advantage of a higher-performing CPU ... so except for the brief jump while it fills the buffer more quickly, the processing done by either CPU will indeed be about the same.  It's not at all clear that a lower TDP CPU would use use less power than an unthrottled CPU at the same processing level, but I do understand what you're saying.    Short of doing the exact same task on the exact same system with two different CPU's installed and monitoring the total power consumption, I doubt you'll ever really "know" the answer.

 

Finally, I agree this is a big divergence from the intent of this thread  :)

... so the discussion should be done in a different thread if you want to continue.

 

My (final) nickel's worth:  Since PMS effectively throttles the CPU anyway, I see no reason to use the lower TDP units ... in the scenario you've outlined, they'll be running at low power anyway;  and the extra headroom they provide just may be useful at some point, should you decide to stream more simultaneous streams.

 

Link to comment

In the Plex transcoding scenario PMS throttles the load and produces a steady stream matching the framerate of the outputted video. It does not go ahead any more than the buffering requires and thus neither processor will drop back to idling sooner than the other. The T-model will have a higher average CPU usage. The difference in CPU usage is exactly the same as the difference of the Passmark values of these two processors. I have benchmarked and verified this. The graph in my previous post indicates that there is a very short buffering with full load and then it goes to throttling with lower CPU load. When I had a more powerful CPU installed the CPU usage in the throttled state was lower and difference was relative to processing power (=Passmark value) of the two CPUs.

 

Fair enough.  So in that scenario, the T-model would run at higher CPU utilization (essetially whatever clock speed was necessary to maintain that load).  The standard CPU should run at the exact same clock speed to maintain the exact same load.  It just so happens that this clock speed is further away from it's max clock speed than it would be for the T-model (signified by lower CPU utilization).

 

i3-4130T - idle 800MHz, max clock 2.9GHz

i3-4130 - idle 800MHz, max clock 3.4GHz

 

So if both CPUs are running at the same clock speed to transcode the video in the above scenario (which they should be), they will both be using the same amount of power and generating the same amount of heat.  To your point regarding the T-model not causing any higher noise (due to fan speed) I agree that should be true, but only in a low utilization steady stream processing job, such as Plex on-the-fly transcoding.  For any other job in which the standard CPU clocks above 2.9GHz the fan will need to spin faster and therefore generate more noise.

 

These 2 CPUs should consume the same energy, generate the same heat, have the same fan speeds, etc up to 2.9GHz.  At 2.9GHz the T-model is capped and can go no higher.  The standard model can continue to clock up to 3.4GHz, obviously consuming more power, generating more heat, higher fan speeds, etc to get there.  Of course the standard model can also perform more work than the T-model at clock speeds above 2.9GHz.  In the end, these two CPUs are identical, the ONLY difference between them is the T-model has had its maximum clock speed capped at 2.9GHz to limit the maximum amount of heat it can generate.

 

All that said, I fail to see how this makes one CPU more energy efficient than the other, but perhaps that's not the point you're discussing with this specific example.  The conversation was about whether or not T-model CPUs were more energy efficient than their standard counterparts.  If it deviated to noise and fan speeds then I apologize as I was unaware.

 

I too feel that this not going to be solved, I've also exhausted my wording. If there is a need to come back to this, I think it's better to be done in a separate thread.

 

I totally understand, garycase is a moderator so perhaps he can split it off into it's own thread if he'd like.

Link to comment
My (final) nickel's worth:  Since PMS effectively throttles the CPU anyway, I see no reason to use the lower TDP units ... in the scenario you've outlined, they'll be running at low power anyway;  and the extra headroom they provide just may be useful at some point, should you decide to stream more simultaneous streams.

 

Agreed.  It's also useful for Plex Sync, whereby Plex pre-transcodes the content for storage on the target playback device for off-line viewing.  In these cases, Plex pegs the CPU to near 100% utilization until the transcode is finished.  Of course, there are also myriad other uses for an unRaid server where you'd want the extra processing power.

Link to comment

All that said, I fail to see how this makes one CPU more energy efficient than the other, but perhaps that's not the point you're discussing with this specific example.  The conversation was about whether or not T-model CPUs were more energy efficient than their standard counterparts.  If it deviated to noise and fan speeds then I apologize as I was unaware.

You are right, we had two different subjects quite close to each other. The subject "will standard model generate more heat than a T-model when performing Plex transcoding" was forked from the efficiency discussion. The outcome of this part was that if there is no efficiency difference then a standard model will not generate more heat/noise while performing Plex on-the-fly transcoding, it will simply run with a lower CPU usage.

 

But the energy efficiency difference issue is still open. I made an order for a i3-4130 which unfortunately was not in stock at local provider and will take 1-2 weeks. I really wouldn't bother with this if I wasn't going to build quite a few of these systems. As I would like to build as energy efficient systems as possible I want to know whether the low-TDP models provide that or not.

Link to comment

You are right, we had two different subjects quite close to each other. The subject "will standard model generate more heat than a T-model when performing Plex transcoding" was forked from the efficiency discussion. The outcome of this part was that if there is no efficiency difference then a standard model will not generate more heat/noise while performing Plex on-the-fly transcoding, it will simply run with a lower CPU usage.

 

But the energy efficiency difference issue is still open. I made an order for a i3-4130 which unfortunately was not in stock at local provider and will take 1-2 weeks. I really wouldn't bother with this if I wasn't going to build quite a few of these systems. As I would like to build as energy efficient systems as possible I want to know whether the low-TDP models provide that or not.

 

My mistake, I failed to note when the discussion branched. 

 

As for the energy eficiency issue, in my mind it's quite settled as I see no difference in efficiency, at least with the data at hand.  That said, I look forward to seeing the results of your testing and some good data one way or the other.

Link to comment

You'll need to run several very controlled test cases to answer your efficiency question.

 

You can do these now for the T model;  then swap the CPU and see what happens with the exact same loads with the unthrottled model.

 

Assuming you have a Kill-a-Watt, you can set it to measure total KwH for the tasks; but you'll need to be sure you're very consistent at zeroing it and reading it at the same point within the process for both setups.

 

Link to comment
  • 1 month later...

Took some time until I had a chance to replace the i3-4130T with a standard i3-4130 and run some tests with both of them.

 

The goal of these tests was simple; to find out if there is an energy efficiency difference between the standard and T-model processor. In other words, will you get lower power consumption when performing the same task. Initially I thought I would be running a transcoding job on both processors fitted in a system otherwise identical. However I did run into problems due to the low precision of my el-cheapo kill-a-watt meter. It measures power consumption by 0.1W but energy consumption only by 0.1kWh. Since the consumption is in 50W region it would have taken too long and would have been too hard to arrange comparable situation.

 

So I decided to see what I could come up using synthetic tests. Below is a chart containing the comparison between i3-4130 and i3-4130T. The system draws about the same ~30W at idle on both processors. When load is introduced (100%, 50% and 25%) there is quite an interesting difference on the power draw. If the two processors would have an identical efficiency, the ratio on different load levels should be close to of the Passmark values. I've also calculated an efficiency ratio from the 100% load figures which tells how many Passmarks you get for each watt.

 

I'm yet to decide whether we should be using the the total system consumption or the difference between the idle and 100% load. My feeling is that it should be the difference figures. Anyhow based on these figures it seems clear that the T-model is more energy efficient. Either 10% or 42% depending how you look at it. Since the two models cost at least here the same, it would not make sense to buy the standard model unless you absolutely needed the extra 18% of processing power. Even in that case you could take a look on the next T-model.

 

So what say you? Does this make sense to you?

 

Intel_i3_4130_vs_4130T_energy_efficiency.png?3815

 

Link to comment

I guess I'll chip in and say I don't agree for all the reasons I previously discussed in this thread. Nothing in this test accounts for the amount of work that was done during the time-slice studied.

 

If you run these tests on a finite workload and measure the total amount of power consumed during a set time frame the total power consumed should be roughly the same between the two CPU's.

 

An example. These are just made up numbers but will illustrate my point. A transcoding job that takes the 4130 CPU 90 seconds to complete, it takes the 4130T 120 seconds to complete. If you measure the total power draw of both systems for say 180 seconds, and both systems perform the transcode somewhere within those 180 seconds, the total power consumed by both systems should be roughly equal as the 4130T spends 30 seconds more at 100% load (49W) while the 4130 has already returned to idle (30W). Yes the 4130 consumes more at 100% load, but that is totally negated by the fact that it finishes the job and returns to a low-power idle state sooner than the 4130T.

 

EDIT: typos. Damn iPad.

Link to comment

I guess I'll chip in and say I don't agree for all the reasons I previously discussed in this thread. Nothing in this test accounts for the amount of work that was done during the time-slice studied.

 

If you run these tests on a finite workload and measure the total amount of power consumed during a set time frame the total power consumed should be roughly the same between the two CPU's.

Ah, didn't explain the test arrangement properly. I just introduced steady synthetic load which caused CPU utilisation at 100%, 50% and 25%. This also caused very steady power draw and these are the watt figures in the table. 25% and 50% are there just for comparison, all the calculations are based on the 100% CPU utilisation since that is something comparable to the Passmark values; I would at least expect the pure CPU benchmark would cause 100% load.

 

Now the 4130 gets a Passmark result of 4900 using 64W (total system draw) and 4130T gets 4123 using 49W. If you now divide the Passmark results with the power usage, you get the efficiency. And 4130T has  ~10% better efficiency (or 40% if you look at the difference figures).

 

It seems odd that both Intel and AMD do not publish energy consumption figures, neither idle or load. These figures would indicate that they do it on purpose...

 

 

Link to comment

Ah, didn't explain the test arrangement properly. I just introduced steady synthetic load which caused CPU utilisation at 100%, 50% and 25%. This also caused very steady power draw and these are the watt figures in the table. 25% and 50% are there just for comparison, all the calculations are based on the 100% CPU utilisation since that is something comparable to the Passmark values; I would at least expect the pure CPU benchmark would cause 100% load.

 

Right, so how did you account for the amount of work that was done by each processor while it was at 100% load for a specified amount of time?  Did you run a specific test suite?  If so, how long did it take each CPU to finish the specified test suite?  If we are measuring efficiency here, how many watts the CPU consumes at 100% load is only half of the equation.  The amount of work it did during that time is the other half.

 

Now the 4130 gets a Passmark result of 4900 using 64W (total system draw) and 4130T gets 4123 using 49W. If you now divide the Passmark results with the power usage, you get the efficiency. And 4130T has  ~10% better efficiency (or 40% if you look at the difference figures).

 

You get what YOU are calling efficiency, expressed as Passmark/W.  I don't really consider that an accurate expression of efficiency (as discussed earlier in this thread).  If you do, then you already have your answer.  After all, most synthetic benchmarks have little to do with real world usage.  All you did was set the CPU utilization to a certain percentage and measure the power draw.  That says nothing of the amount of work done by said CPU.  A 4130 can do more work per given time-slice at 100% utilization than can a 4130T.

 

It seems odd that both Intel and AMD do not publish energy consumption figures, neither idle or load. These figures would indicate that they do it on purpose...

 

That would be hard to do as actual workloads would have to be taken into account, not to mention various workloads (typical server, typical desktop, etc).  And what's typical for one isn't typical for another.  Besides, the marketing department would likely reject the idea outright for numerous reasons.

Link to comment

 

Right, so how did you account for the amount of work that was done by each processor while it was at 100% load for a specified amount of time?  Did you run a specific test suite?  If so, how long did it take each CPU to finish the specified test suite?  If we are measuring efficiency here, how many watts the CPU consumes at 100% load is only half of the equation.  The amount of work it did during that time is the other half.

 

You get what YOU are calling efficiency, expressed as Passmark/W.  I don't really consider that an accurate expression of efficiency (as discussed earlier in this thread).  If you do, then you already have your answer.  After all, most synthetic benchmarks have little to do with real world usage.  All you did was set the CPU utilization to a certain percentage and measure the power draw.  That says nothing of the amount of work done by said CPU.  A 4130 can do more work per given time-slice at 100% utilization than can a 4130T.

Passmark is actually very accurate when it comes to the real life scenario from which this discussion started from, Plex transcoding:

http://lime-technology.com/forum/index.php?action=post;quote=272129;topic=30355.0;last_msg=279273

 

Few months ago I tested CPU utilisation for Plex transcoding using three different CPUs and Passmark results were within 1-3% accuracy. That time I had a very controlled environment running the same movie transcoding job and recording the CPU utilisation for each 5 second time period. The results were very clear and repeatable; you can directly derive CPU's transcoding capability from it's Passmark result. I would expect the same to apply to any processor only test scenario.

 

I admit that I do not know whether running Passmark causes similar utilisation and power draw as my synthetic 100% load test. I have just assumed that it must since it's trying to max out the performance from the processor. I feel this to be a fair assumption.

 

Why do you find it hard to believe that there could be an efficiency difference between processors when the same has been always between processor families? I was also surprised about the large gap (yep, I'm now convinced that the difference is closer to the 40% at full load) but otherwise I don't see a major problem in the logic.

 

As a totally different matter, when you look at the 100% utilisation power usages compared to the idle usage (34W for 4130 and 20W for 4130T), they are both almost exactly 60% of the related TDP values. Seems like a pretty large safety margin for me, but of course we do not know the actual idle power usage for the processors only (30W - the rest of the system). And we haven't even started to talk about the efficiency of the PSU which at these low levels isn't that good either though it's a a 80 Gold certified one.

 

I also measured a peak usage of 100W on system startup. This was with 4 x WD Green 3TB and 2 x WD Red 4TB. I guess we could easily do with the 300W Silverstone SFX in the Lian Li PC-Q25.

Link to comment
  • 4 months later...

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.

Guest
Reply to this topic...

×   Pasted as rich text.   Restore formatting

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.