News Intel's latest lower-powered CPUs give Ryzen rivals a run for their money — Core i9-14900T beats Ryzen 9 7900 in Geekbench 6 benchmark

Page 3 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.

bit_user

Titan
Ambassador
TAU is just the the interval the CPU checks to see what conditions there are, if everything is fine nothing will change.
Again, the doc said.

"PL2- SoC opportunistic higher Average Power with limited duration controlled by Tau_PL1 setting.
the larger the Tau, the longer the PL2 duration."

So, if by "everything is fine", you mean that power is <= PL1, then yes. Also, I'm sure it's not a simple polling interval. It's a threshold or window size. The underlying polling is probably happening at a few kHz.

You are conditioned to think that TAU means that the CPU is forced to reduce clocks/power because that's the only type of benchmarks and reviews we ever get to see.
No, because it was explained that way by an in-depth interview Ian Cutress / Anandtech did with a senior technical figure from Intel:

95WTau.png

Source: https://www.anandtech.com/show/1458...ng-an-interview-with-intel-fellow-guy-therien

That's the conditioning, but the main reason I think it's still an accurate description is that the very document you referenced explains it exactly so.

In case of the PL2 TAU ,if it is ever enabled,
Where, in the doc, do they define a "PL2 Tau"?

What they describe is a situation where PL1 is increased to PL2. If both power limits are the same, then EWMA has no effect.

Also why are you linking tpu that shows a PL2 TAU number to justify anything about PL1 TAU, you were strictly against that when I showed numbers for different things.
The default values of 28 seconds is the same in each case. I think they simply mis-labeled the parameter. My point was that 28 is indeed a time value and not Wattage.
 
No, because it was explained that way by an in-depth interview Ian Cutress / Anandtech did with a senior technical figure from Intel:
No, it wasn't explained that way, it had that one picture and an interview and your confirmation bias made you believe that the picture makes you right, you didn't even bother to scroll down a bit to see the second picture that shows the PL2 forever.
Here's what he said:
This PL2 limit, and the power budget, is configurable.
on all client parts,
you can configure both the period over which TDP is enforced and the height of the power limit to which you can burst.
Then, after a time, it’s looking at the time over which it needs to enforce that average power, which we call Tau (τ).
When Tau is reached, the algorithm kicks in and sees if it needs to reduce the power that’s being consumed, so it can hit can stay within the TDP limit.
But a partner can clearly design a system to have a three minute turbo window over which the TDP is enforced. So depending on the system capability, we have the TDP, and then the associated Turbo parameters (PL2 and Tau) can be configured to meet the capability of the system.
PL2Forever_575px.png

and such, and they are implementing their own values for that PL2 limit and also the turbo window – they might be pushing these values up until the maximum they can go, such as a (maximum) limit of 999 W for 4096 seconds. From your opinion, does this distort how we do reviews because it necessarily means that they are running out of Intel defined spec?

Guy Therien:
Even with those values, you're not running out of spec, I want to make very clear – you’re running in spec, but you are getting higher turbo duration.
 

bit_user

Titan
Ambassador
No, it wasn't explained that way, it had that one picture and an interview and your confirmation bias made you believe that the picture makes you right, you didn't even bother to scroll down a bit to see the second picture that shows the PL2 forever.
You accusing me of confirmation bias is a bit much.

What he said was 100% consistent with the image I included. He confirmed that Tau is the time limit for PL2. So, no discrepancy there.

And, since you seem to have forgotten, that's the only reason I cited that article. You accused me of having learned about Tau from some reviewers (with the implication presumably being that they're unreliable sources), so I just cited that 2019 interview, where the relation of PL1, PL2, and Tau is clearly explained.

"This PL2 limit, and the power budget, is configurable. In fact, on all client parts, you can configure both the period over which TDP is enforced and the height of the power limit to which you can burst."​
First, I'd point out that the discussion happened at 2019 Computex (i.e. January, 2019). So, the client parts he was referring to are Coffee Lake. We can't necessarily apply what he said to all future processors.

Next, consider that without overclocking, Coffee Lake would hit a practical maximum amount of power. So, you can set PL2 as high as you want, but that doesn't mean it'll hit that limit, at stock clocks. And if you overclock, well... that's out of warranty, so they don't care.

But, we're not really here to discuss PL2. Whether you can set a higher PL2, on the i9-14900T is irrelevant to the concerns I raised. IMO, 106 W is bad enough. In fact, the same is true of Tau. Bursting to 106 W for 28 seconds is more than enough to be obnoxious, in a mini-PC chassis, while not preventing a noticeable performance drop when gaming or doing other sustained work.

Finally, there's the matter of PL1. If you can't increase PL1, on the i9-14900T, then you don't get unlimited turbo. Tau is definitely limited at 448 seconds (if that much), and it seems the way people go about bypassing that is actually by just increasing PL1 to match PL2.

"they might be pushing these values up until the maximum they can go, such as a (maximum) limit of 999 W for 4096 seconds."
This part was Ian Cutress talking, not the interviewee. I don't know if Ian took those values from an actual example, or if he was just using extreme, imaginary values to make a point.

So, ultimately, the only point up in the air is whether the PL1 of i9-14900T can be increased beyond 35 W. If so, how much? Depending on this point, my statement about performance dropping on sustained workloads (which includes gaming) may not apply. However, I still contend that these T-series CPUs are clearly aimed at SFF form factors where bursting to 106 W for more than a couple seconds is likely to be extremely obnoxious.
 
You accusing me of confirmation bias is a bit much.

What he said was 100% consistent with the image I included. He confirmed that Tau is the time limit for PL2. So, no discrepancy there.
He confirmed that TAU is the time period after which the system checks the AVERAGE power used by the system, if the AVERAGE has met whatever WAS SET as PL1, only then will it stop boosting up to PL2.
That's a huge giant difference from what you are trying to peddle, that it is a binary one time cut off switch.
Then, after a time, it’s looking at the time over which it needs to enforce that average power, which we call Tau (τ).
When Tau is reached, the algorithm kicks in and sees if it needs to reduce the power that’s being consumed, so it can hit can stay within the TDP limit.
So, ultimately, the only point up in the air is whether the PL1 of i9-14900T can be increased beyond 35 W. If so, how much? Depending on this point, my statement about performance dropping on sustained workloads (which includes gaming) may not apply. However, I still contend that these T-series CPUs are clearly aimed at SFF form factors where bursting to 106 W for more than a couple seconds is likely to be extremely obnoxious.
Historically, Intel mentioned the Tau parameter in relation to PL1/PL2. However, with the debut of the 12th Gen Processors, PL1 and PL2 have identical values. Consequently, it's redundant to define the duration the CPU will sustain at PL2 before transitioning to PL1 since both the long and short-duration power values are now the same.
 
However, I still contend that these T-series CPUs are clearly aimed at SFF form factors where bursting to 106 W for more than a couple seconds is likely to be extremely obnoxious.
It will never happen in such systems because they won't allow the CPU to boost to 106W even for a second.
That's why we have configurable PL in the first place, any OEM worth anything will set PL to whatever their cooling system can handle at the noise level they decide.
But even without the OEM setting anything the CPU will not boost to 106W if there isn't enough cooling there to support it.
 

bit_user

Titan
Ambassador
He confirmed that TAU is the time period after which the system checks the AVERAGE power used by the system, if the AVERAGE has met whatever WAS SET as PL1, only then will it stop boosting up to PL2.
You realize that's actually worse for you, right? If the average over 28 seconds needs to be 35 W, and you start from about 0 W, then you can only boost to 106 W for a mere 9.25 seconds!

"Historically, Intel mentioned the Tau parameter in relation to PL1/PL2. However, with the debut of the 12th Gen Processors, PL1 and PL2 have identical values."
That's referring to K-series, only.

It will never happen in such systems because they won't allow the CPU to boost to 106W even for a second.
That's why we have configurable PL in the first place, any OEM worth anything will set PL to whatever their cooling system can handle at the noise level they decide.
Then that invalidates the benchmark results, in the article, for any such systems.
 

Ogotai

Reputable
Feb 2, 2021
361
240
5,060
because they won't allow the CPU to boost
um who wont allow it ? intel ? thats a laugh and a joke... intel does nothing to the mobo makers for how they run their chips, allowing them to do what ever it takes to keep their cpus competitive, while using as much power ad they need to, and now you think intel wont allow it with the t series ? come on... imtel will allow it, as long as it makes their cpus look good.
 
um who wont allow it ? intel ? thats a laugh and a joke... intel does nothing to the mobo makers for how they run their chips, allowing them to do what ever it takes to keep their cpus competitive, while using as much power ad they need to, and now you think intel wont allow it with the t series ? come on... imtel will allow it, as long as it makes their cpus look good.
The system itself will not allow it, power limit does not trump thermal throttling.
A system that is made for 35W and has cooling for 35W, let it be 50W , will not have any thermal headroom for the CPU to reach anywhere close to 100W ,it will thermal throttle to that 35-50W whatever the cooling can handle.
 
You realize that's actually worse for you, right? If the average over 28 seconds needs to be 35 W, and you start from about 0 W, then you can only boost to 106 W for a mere 9.25 seconds!
Did your dementia kick in? We just had that conversation one of these days.
You never start from 0, the CPU needs a certain amount of voltage to even run and that corresponds to a certain amount of watt.
Remember?
v2-3246ea92d494fffe1ea26f30bc9d9d79_720w.webp

I wouldn't be surprised if that 35W is the lowest the CPU can ever operate at.
If the average over 28 seconds needs to be 35 W, and you start from about 0 W, then you can only boost to 106 W for a mere 9.25 seconds!
The average power draw AFTER the amount of TAU not during the whole period of TAU, and yes that's my whole point if you have that CPU setup with an TDP of 35W and the cooling to match then it will almost never boost.
That's referring to K-series, only.
Only in your mind.
"However, with the debut of the 12th Gen Processors"
Not 12th gen K processors, not high-end processors, not some processors.
Then that invalidates the benchmark results, in the article, for any such systems.
Well, it's not an benchmark OF any such system.
It's a freely available CPU that anybody can buy and put into any system they want to.
 

bit_user

Titan
Ambassador
Did your dementia kick in?
I'm sure you know the site's policy on personal attacks.

You never start from 0, the CPU needs a certain amount of voltage to even run and that corresponds to a certain amount of watt.
I used a hypothetical example of starting at zero W, for the sake of simplifying the math.

It also happens to benefit the amount of boost duration, compared to a real figure for the package power of the CPU at idle. If you'd rather use that number we could work it out and the boost duration would probably fall below 9 seconds, given your definition of Tau and a value of 28 seconds.


The average power draw AFTER the amount of TAU not during the whole period of TAU,
If we focus just on what the Gen 13 Reference says:

"Turbo Time Parameter (Tau): An averaging constant used for PL1 exponential weighted moving average (EWMA) power calculation."
(Pg. 76)

So, Tau defines the averaging interval. Now, I used a simplistic example, but they describe it as an exponential weighted, moving average. So, it's really more of an IIR. This also explains why they indicate that you need to wait about 3x Tau for the algorithm to settle out, when adjusting Tau on the fly.

This diagram is quite busy, but there's a curved, dotted line in the lower-left which reflects the approximate average power, under these circumstances. The shape of the curve is defined by the precise EWMA formula (which I haven't been able to find), but the width correlates directly to Tau. Had the average used a simple box filter, the dotted line would be straight.

"However, with the debut of the 12th Gen Processors"
Not 12th gen K processors, not high-end processors, not some processors.
If it were true, then you could cite where the reference doc says this. However, the reference doc says no such thing.
 
Last edited:
I'm sure you know the site's policy on personal attacks.
It is a sincere concern since we just talked about that.
It also happens to benefit the amount of boost duration, compared to a real figure for the package power of the CPU at idle. If you'd rather use that number we could work it out and the boost duration would probably fall below 9 seconds, given your definition of Tau and a value of 28 seconds.
I already said, soooooo many times, that it will never even once go to 106W if the system doesn't support it.
If it were true, then you could cite where the reference doc says this. However, the reference doc says no such thing.
It doesn't say anything to the contrary either...otherwise you would have cited that.

Look at this video, the CPU, a 13900H and not a 14900t, but unless you can find any info that says that the turbo method changed between 13th and 14th then this still is valid, draws around 20W at idle, when starting the stress test it goes to 45W first, the official PBP of that CPU, for a beat and only then does it calculate that it has some headroom left and goes above that to 65W which is either the TDPup that was set in the bios by selecting performance mode or it is the thermal limit because one of the cores is at 90 and the OEM could have set 90 as the limit to not melt that little thing.

TL;DR
It does not just go from "zero" to 115W for TAU amount of seconds just because that's the official PL2...

Also it shows average total system power at the end being 82W for gaming which means that it stayed at 65W for the CPU for the duration of test.
View: https://youtu.be/uB2PKwC4ZPs?t=281


Also here a different video that shows the 13900h again, with him freely changing the two PL in XTU.
 

bit_user

Titan
Ambassador
It is a sincere concern since we just talked about that.
For future reference: alleging someone is suffering from a medical condition is not an appropriate way to express concern. Future incidents will be reported and the mods can adjudicate as they see fit.

There's usually more than one explanation for things. Maybe something was discussed and one party forgot, but maybe the parties thought they were discussing different things. You can remind them or ideally provide a link to the earlier conversation, which would be the best/quickest way to clear up the matter. Simply making an allegation, without evidence, provides no way for the person or other individuals to know if your claim has any merit. Furthermore, linking to the earlier discussion lets them refresh themselves on what was said, without you having to repeat yourself (in case you're right that they simply forgot).

In many cases, we'll discuss something and it doesn't end in agreement. So, the mere fact that you previously made some claim isn't sufficient cause to assume I'll take it as given, the next time around.

I already said, soooooo many times, that it will never even once go to 106W if the system doesn't support it.
I can point you to plenty of mini-PC reviews where reviewers complained about loud fans. So, if you're depending on the system integrator to prevent it from boosting to the point where fan noise becomes annoying, I think that's unreliable at best.

Heck, even a Dell compact desktop I use at work gets annoyingly loud, when you hit its "65 W" i9 CPU with an all-core load and it's not that small. Fortunately, I was able to place mine in the server room and just connect to it from my laptop.

It doesn't say anything to the contrary either...otherwise you would have cited that.
Where it lists PL1 = PL2 is on Page 98, for the "8P + 16E 150 W" model. That should correspond to the i9-13900KS. It also lists them as equal for "8P+16E Core 125W Extreme Config", which I guess is a maximal configuration of the i9-13900K.

Beyond that, it's full of discussions about PL1 vs. PL2 as separate quantities. If they were now the same, they could just get rid of all that.

Look at this video, the CPU, a 13900H and not a 14900t
That's an issue The i9-13900H supports Assured Power, while the T-processor does not. They're also different dies, with the former being a P-series die and the latter being a S-series die.

draws around 20W at idle, when starting the stress test it goes to 45W first, the official PBP of that CPU, for a beat and only then does it calculate that it has some headroom left and goes above that to 65W which is either the TDPup that was set in the bios by selecting performance mode or it is the thermal limit because one of the cores is at 90 and the OEM could have set 90 as the limit to not melt that little thing.
It sounds like you're talking about cTDP functionality, which is a separate thing. It's briefly described on pages 78-79 of that document.

TL;DR
It does not just go from "zero" to 115W for TAU amount of seconds just because that's the official PL2...
The official diagram from Intel's reference doc shows exactly that. So, I'd suggest looking at a non-laptop processor.

In my own experiments, I see exactly this behavior on a Dell Precision desktop, with a i9-12900.

Also it shows average total system power at the end being 82W for gaming which means that it stayed at 65W for the CPU for the duration of test.
If you have a dGPU, then gaming on an i9-14900T at 35 W would definitely make it a bottleneck.
 
Last edited:
I can point you to plenty of mini-PC reviews where reviewers complained about loud fans. So, if you're depending on the system integrator to prevent it from boosting to the point where fan noise becomes annoying, I think that's unreliable at best.

Heck, even a Dell compact desktop I use at work gets annoyingly loud, when you hit its "65 W" i9 CPU with an all-core load and it's not that small. Fortunately, I was able to place mine in the server room and just connect to it from my laptop.
How is that different depending on the CPU used? An OEM using a crap cooler isn't any indication on any characteristic of the CPU.
Although, you can tune down this 14900t to 35W and it will do a lot of work at that 35W while no desktop CPU from AMD will even turn on at 35W so if you have a loud fan still depends on how much you personally want to push a little system like that .
Where it lists PL1 = PL2 is on Page 98, for the "8P + 16E 150 W" model. That should correspond to the i9-13900KS. It also lists them as equal for "8P+16E Core 125W Extreme Config", which I guess is a maximal configuration of the i9-13900K.
This means what it says, these are just default settings and not enforced, they are NOT characteristics of the part.
The k CPUs are defaulted for pl1=pl2 but you can still reduce pl1 and pl2 if you want to, hence extreme, and you can still change pl1 and pl2 on lower models.
7 This is a hardware default setting and not a behavioral characteristic of the part.
jZA2aDZ.jpg

Beyond that, it's full of discussions about PL1 vs. PL2 as separate quantities. If they were now the same, they could just get rid of all that.
They could also not even publish a datasheet in the first place, it's full of discussions to provide some starting point for OEMs.
The official diagram from Intel's reference doc shows exactly that. So, I'd suggest looking at a non-laptop processor.

In my own experiments, I see exactly this behavior on a Dell Precision desktop, with a i9-12900.
Note 6 which is also applied to all CPUs.
If your 12900 just boosts up to max turbo then turbo boost 2 is disabled and it just runs at the max power it can reach all the time, or if it clocks down then speed step is still enabled but under any load it boosts to full.

6 The Processor will be controlled to a specified power limit as described in Intel® Turbo Boost Technology 2.0 Power Monitoring on page 47. If the power value and/or 'Turbo Time Parameter' is changed during runtime, it may take a short period of time (approximately 3 to 5 times the 'Turbo Time Parameter') for the algorithm to settle at the new control limits.
Intel® Turbo Boost Technology 2.0 Power Monitoring When operating in turbo mode, the processor monitors its own power and adjusts the processor and graphics frequencies to maintain the average power within limits over a thermally significant time period. The processor estimates the package power for all components on the package. In the event that a workload causes the temperature to exceed program temperature limits, the processor will protect itself using the Adaptive Thermal Monitor
If you have a dGPU, then gaming on an i9-14900T at 35 W would definitely make it a bottleneck.
Yeah, no kidding. And gaming on the iGPU is limited as well, the point was that it was sustaining high power on the CPU even though it's not an unlocked part.
 

bit_user

Titan
Ambassador
If we focus just on what the Gen 13 Reference says:

"Turbo Time Parameter (Tau): An averaging constant used for PL1 exponential weighted moving average (EWMA) power calculation."
(Pg. 76)

So, Tau defines the averaging interval. Now, I used a simplistic example, but they describe it as an exponential weighted, moving average. So, it's really more of an IIR. This also explains why they indicate that you need to wait about 3x Tau for the algorithm to settle out, when adjusting Tau on the fly.

This diagram is quite busy, but there's a curved, dotted line in the lower-left which reflects the approximate average power, under these circumstances. The shape of the curve is defined by the precise EWMA formula (which I haven't been able to find), but the width correlates directly to Tau. Had the average used a simple box filter, the dotted line would be straight.
Here's data I collected from two runs on that i9-12900 machine. I used a MSR tool to confirm that its PL1, PL2, and Tau are indeed set to 65 W, 202 W, and 28 seconds, as specified on ark.intel.com (and in the reference manual):

The first run is PoV-Ray, running benchmark.pov at quality=9 and targeting a 4k resolution:

WSxwZ78.png

If you zoom it, you can see that there's a brief, single-threaded setup phase, after which it boosts in the range of 164 to 176 W, for a total of 12.8 seconds. It then drops to a hair below 65 W, where it stays for the rest of the run. I think the reason power increases during the boost period is probably due to the complexity increasing as the render gets further into the frame.

The second run is using a program called Stress-NG, with cpu-method=fibonacci and 24 threads, which I stopped a short while after the power dropped to 65 W. This was the CPU method that used the least amount of power. I used it to confirm the boost duration is prolonged at lower power levels.

WhkRtfK.png

In this case, you can see it ranges from 116.5 to 123.4 W, over a period of 21.3 seconds. Then, it drops to just below 65 W. It's clear that temperature has nothing to do with the drop, since it was about 90 C, in the above case, and only about 81 C in this case.

I think these demonstrate the behavior of the EWMA algorithm, exactly as described above. I wonder if the boost duration only really hits 28 seconds at 1.25 * PL1 (i.e. 81.25 W). I recall seeing this formula in the doc, but I didn't really register what it was saying.
 
Last edited:
COavhwQ.png

If you zoom it, you can see that there's a brief, single-threaded setup phase, after which it boosts in the range of 164 to 176 W, for a total of 12.8 seconds. It then drops to a hair below 65 W, where it stays for the rest of the run. I think the reason power increases during the boost period is probably due to the complexity increasing as the render gets further into the frame.
Why are you describing the series 2? Which I would guess is the second consecutive run of the same bench?
Series 1 shows the power only drops to about 100W after TAU and it goes up to about 145W , it stays above 100W for the whole 300 secs with an TAU of 28 sec, it does not stay stuck at 65W until it runs for the second time.
 

bit_user

Titan
Ambassador
Why are you describing the series 2? Which I would guess is the second consecutive run of the same bench?
Series 1 shows the power only drops to about 100W after TAU and it goes up to about 145W ,
Apologies, but it seems I hit a very strange bug in MS Excel. Whenever I copy the plots, the series lose their names! Furthermore, the formatting of the legend resets. I didn't believe it myself, until I tried several times - including saving the file & re-opening. I added another callout to each graph, which is visible in the copy, so I know it wasn't just using old data from the clipboard. I worked around it by using a screen shot.

Also, I should clarify that I included power and temperature data in the same plot. There's a set of Y-axis labels on both the left & right sides of the plots, as I've noticed professional reviewers tend to do, when including multiple data types in the lame plot. The left axis labels apply only to the yellow series (Package Power), while the right axis labels apply only to the orange series (Package Temperature). I added an additional call-out to each graph, in hopes of making it a little clearer that orange series are on a different Y-scale.

Edit: In my re-uploaded images, I initially had the series labels swapped. Those were only visible for a couple minutes, so you hopefully missed those. The correct images are now showing. Maybe refresh your browser window, just to be sure.

Original post updated (revision 3 now complete). Please re-read.
 
Last edited:
Apologies, but it seems I hit a very strange bug in MS Excel. Whenever I copy the plots, the series lose their names! Furthermore, the formatting of the legend resets. I didn't believe it myself, until I tried several times - including saving the file & re-opening. I added another callout to each graph, which is visible in the copy, so I know it wasn't just using old data from the clipboard. I worked around it by using a screen shot.

Also, I should clarify that I included power and temperature data in the same plot. There's a set of Y-axis labels on both the left & right sides of the plots, as I've noticed professional reviewers tend to do, when including multiple data types in the lame plot. The left axis labels apply only to the yellow series (Package Power), while the right axis labels apply only to the orange series (Package Temperature). I added an additional call-out to each graph, in hopes of making it a little clearer that orange series are on a different Y-scale.

Original post updated. Please re-read.
I don't want to believe that those are temps, these numbers are way too high, if you are just taking CPU temps then it's not even measuring the hottest core temp, you are definitely hitting thermal throttling which would explain why you are not getting Pl2 for even the duration of TAU.

I can't believe that you saw temps in the 90ies with a 65W power draw and thought to yourself, yeah, that's fine....nothing to see here.
(now I imagine you as the little girl that is smiling evily with the house burning down behind her)
 

bit_user

Titan
Ambassador
I don't want to believe that those are temps,
I momentarily had the labels switched around, in my first update. Please re-check, but it sounds like you figured out the orange series are the temperatures.

you are definitely hitting thermal throttling which would explain why you are not getting Pl2 for even the duration of TAU.
It's not hitting thermal throttling. I later tried increasing PL1 and was able to reach 100 C. Plus, thermal throttling just backs off slightly. Unless you're way over, it doesn't drop power like what we're seeing.

Furthermore, if you see the fibonacci test case, power gets cut when temperature is only 83 C. I added a call-out, so you can clearly see.

The point you see to be missing is that Tau is just a constant characterizing the width of an IIR filter. It's not a hard threshold! Remember, the average power is computed via EWMA (Exponential Weighted Moving Average). The actual shape of the window function should be something like this:

I can't believe that you saw temps in the 90ies with a 65W power draw and thought to yourself, yeah, that's fine....
It's a Dell compact desktop that won't allow you to configure the higher-TDP CPU models. So, it's clearly designed only for a 65 W TDP. To include a beefier cooling system would be a waste of money for them.

(now I imagine you as the little girl that is smiling evily with the house burning down behind her)
This is not a constructive comment.
 
Last edited:

bit_user

Titan
Ambassador
To benefit our understanding of the subject, I tried a further experiment (not shown), where I set a custom PL2 of just 82 W. In this test, boost was sustained for 43.7 seconds! I expected it should be possible to see boost durations longer than Tau, but I wasn't sure what was the point when boost would equal Tau.

It also occurred to me that another interesting consequence of the EWMA method is that you can actually increase the boost duration simply by increasing PL1 and not even touching Tau or PL2!

Another experiment I tried was to see what the boost duration would be with PL1=35 W and PL2=106 W. Assuming the EWMA formula used in the Gen 14 CPUs is the same as Gen 12, the boost duration should exactly match what I measured (below)!

cW2QzYp.png


So, we see that the T-series should boost for just 12.8 seconds, on workloads involving more than just a couple threads.
 
It's not hitting thermal throttling. I later tried increasing PL1 and was able to reach 100 C.
Well, this is obviously a lie since you are so adamant about it being impossible to increase PL1 on locked CPUs....
The point you see to be missing is that Tau is just a constant characterizing the width of an IIR filter. It's not a hard threshold! Remember, the average power is computed via EWMA (Exponential Weighted Moving Average). The actual shape of the window function should be something like this:
That's what I have been saying for so long.
But just in the same way that EWMA can reduce the power so can overheating.
TAU is just the the interval the CPU checks to see what conditions there are, if everything is fine nothing will change.

It also occurred to me that another interesting consequence of the EWMA method is that you can actually increase the boost duration simply by increasing PL1 and not even touching Tau or PL2!
It was said from post #14 and it only took you 3 pages this time...
It was a whole thing 2 years ago, Pl1=Pl2=241W ,they stopped using TAU in form of a timed limit, now turbo boost 2 just boosts as far and as long as thermals and average power draw are within the set limits.
Any CPU that has turbo boost 2 has no TAU in seconds anymore.
12th-Gen-Blueprint-Series-Presentation1-27-1.png
 

bit_user

Titan
Ambassador
Well, this is obviously a lie since you are so adamant about it being impossible to increase PL1 on locked CPUs....
No, I never said that. I asked you for evidence, one way or another, which you failed to provide.

It was @TJ Hooker who finally provided a reference to boosting the PL values on non-K parts. However, as that didn't say anything specifically about T-series parts, the question still remains whether their PLs can be altered and to what extent.

That's what I have been saying for so long.
Not exactly. You claimed that a boost duration shorter than Tau was evidence of thermal throttling. If you understand how EWMA works, then you wouldn't say that, when a large ratio exists between PL1 and PL2.

But just in the same way that EWMA can reduce the power so can overheating.
Not in the same way. Like I said, I got the CPU to bounce off 100 C, and what that looked like was very different. It basically oscillated between 99 and 100 C, only backing off power consumption by a little bit. Conversely, when EWMA crosses the PL1 threshold, the above plots shows how power consumption plummets and temperatures quickly follow.

It was said from post #14 and it only took you 3 pages this time...
No, what you said in that post isn't accurate. You can't claim that PL1=PL2 and maintain that Tau still does anything. Its function is primarily defined in terms of PL1, and if PL2 is the same as PL1, then the CPU won't be boosting above PL1 long enough for Tau to have any effect.

I agree that Tau isn't strictly a number of seconds, but it's still a time constant and Intel still labels its units as seconds. Furthermore, if you start from idle and use a certain ratio of PL1:PL2, you will indeed experience boost for approximately that number of seconds. It's still an open question what that ratio is...
 
Last edited:
Not exactly. You claimed that a boost duration shorter than Tau was evidence of thermal throttling. If you understand how EWMA works, then you wouldn't say that, when a large ratio exists between PL1 and PL2.


Not in the same way. Like I said, I got the CPU to bounce off 100 C, and what that looked like was very different. It basically oscillated between 99 and 100 C, only backing off power consumption by a little bit. Conversely, when EWMA crosses the PL1 threshold, the above plots shows how power consumption plummets and temperatures quickly follow.
Because you increased PL1 and you either hit the exact spot or thermal throttling reduces it to exactly what the cooler can barely support...
Keep the PL the same but put the cooler to the lowest rpm and lets see what happens then, power will always drop to what the cooler can support, that's just how physics works, the CPU will not wait to get fried just because TAU isn't over yet.
No, what you said in that post isn't accurate. You can't claim that PL1=PL2 and maintain that Tau still does anything. Its function is primarily defined in terms of PL1, and if PL2 is the same as PL1, then the CPU won't be boosting above PL1 long enough for Tau to have any effect.
I didn't maintain that TAU does anything when PL1=Pl2, I said the exact opposite of what you are trying to convince yourself that I said.
I said that TAU isn't a thing at all anymore, and the datasheet agrees since they just call these default settings but not hardware characteristics.
It was a whole thing 2 years ago, Pl1=Pl2=241W ,they stopped using TAU in form of a timed limit, now turbo boost 2 just boosts as far and as long as thermals and average power draw are within the set limits.
Any CPU that has turbo boost 2 has no TAU in seconds anymore.
 

bit_user

Titan
Ambassador
Because you increased PL1 and you either hit the exact spot or thermal throttling reduces it to exactly what the cooler can barely support...
Yes, it's a gradual backoff. Not like the sudden dropoff in power, shown by every plot I've posted so far.

Just so we're completely clear, here's what it looks like when the CPU hits 100 C. For this test, I set PL1=125W and PL2=241W (not that it matters, because the workload can't even reach the default PL2 of 202W).

cRkXEdv.png


And here's the same exact workload, on the same exact system, where the only difference is it's running at the stock limits of PL1=65W and PL2=202W. I made sure the axis scales are exactly the same, too.

40Qd3me.png


I didn't maintain that TAU does anything when PL1=Pl2, I said the exact opposite of what you are trying to convince yourself that I said.
I said that TAU isn't a thing at all anymore, and the datasheet agrees since they just call these default settings but not hardware characteristics.
Tau is certainly a thing, as the graphs I've posted clearly show! The data I presented is consistent with the EWMA algorithm described in the doc.

The only way to make Tau not a "thing" is if you set PL1 high enough that the system is either frequency-limited or thermally-limited beneath it. In such a configuration, it simply becomes irrelevant.

Since this thread is about a CPU with a PL1 of 35W, Tau is absolutely relevant (unless you put it in such a trash PC that its cooling can't even dissipate 35 W of heat).
 
  • Like
Reactions: Ogotai

DaveLTX

Commendable
Aug 14, 2022
104
66
1,660
Funny thing is, you're the only one talking about a 7950X. The article didn't mention it, and neither did I. I wouldn't recommend that CPU for someone looking to build a 45 W system or below. AMD doesn't even sell a non-X 7950. The highest their 65 W CPU range goes is the 7900.

Injecting an unrelated topic into the discussion is called whataboutism and it's a red flag that someone is posting in bad faith, or is at least working an agenda.


Maybe, but that's news to me. Got proof?


If you're saying the original benchmarks referenced by this article could've hit thermal throttling before Tau, I suppose that's possible. However, that doesn't mean Tau won't hit and Tau-throttling is far worse than thermal throttling (unless you have a truly rubbish heatsink, which it's very unlikely they do).
I install 13400/13700 on asrock B760 boards and they allow me to boost to almost any wattage I set. Only limit are the max clocks intel bakes in

12400 of course won't turbo any higher but that's 12400 being set low

But, non K parts have turbo timers. Which the motherboard can ignore but OEM boards which usually have the T parts (intel hasn't sold T parts outside OEM system) absolutely respects turbo timers. The turbo timer is pretty much Tau in behaviour.


And of course the systems getting T parts usually have cooling systems and power delivery that will not support max boost either. Probably more on the level of 70W
So for those parts either of 3 things will happen
PL2 expires after 28 seconds if it maintains 100c
PL2 expires because it went over 100c
Throttling occurs separately because of power limit alongside the other two because of small external power supply (see where T parts usually end up) and small VRM because... Why do you need a massive vrm setup to support a "35W" tdp proc?
 

DaveLTX

Commendable
Aug 14, 2022
104
66
1,660
Anyway all this armchair arguments are pointless because the 35-45W market are all serviced by 7840H/S & 155H anyway
Unless they have a specific requirement to have a upgradable CPU. In which case is extremely moot when Intel kills the socket within Alder/Raptor which is essentially the same

And also OEM bioses usually only support the generation they were designed in.
So if it was designed for 12th gen T then it will support only 12th gen T CPUs for instance
 
  • Like
Reactions: bit_user
Status
Not open for further replies.