Dirt ChipWe have come to the days that it is not absurd to buy a high-end CPU/GPU just to get the better chip bin and then downclock and/or undervolt it to hit the best heat/pref point.

Get a 7950X, shut down 4 cores, limit to 105-95w and enjoy most of the pref at half the watteg and $ cost.

:ohwell:

Solid State BrainModern CPUs are the most efficient (joules for the work accomplished) at the frequency that minimizes operating voltage. Of course, to lower voltage you have to decrease frequency and by extension, power.

To help understanding the idea, here is a graph I made with my i7-12700K at default frequencies, but optimized voltages (my motherboard gives plenty of voltage with default settings). With the settings it has at the moment, the CPU won't require more than 150W, but under overclocked conditions it could take up to 250W or more if cooling allows. Two things are clear from the graph: increasing power yields progressively diminishing performance returns and increases the energy (Joules) required to perform the same work (here, a render test similar to that used in TPU reviews).

All CPUs both from Intel and AMD will exhibit a behavior along these lines.

ratirtGreat but that is not the point we've been arguing about. everyone knows about the efficiency and power a CPU requires to operate. We are talking about limiting a 170W cpu to 65w for efficiency but this will drop the performance very much and it may not be worth paying so much for performance which will not be there any longer.
The other thing is, we are not talking hear about tweaking the CPU but limit its power. That is also different.
Like i said, I undervolted my 6900xt and it runs cooler, uses less power but the performance stayed exactly the same.
Than you have a CPU which uses 170w and you cap it at 65w loosing lets say 30% of performance, since you just want to have a efficient CPU for which you pay a lot money. Now, you pay for performance or for the tweaking? I don't see a gain here but regression in performance instead by lowering power usage exponentially

Asni@W1zzard Can you please add the 7950x with HT off in you cpu test suite?

W1zzardHmm, what would that answer? Nobody will run them like that?

ratirtGreat but that is not the point we've been arguing about. everyone knows about the efficiency and power a CPU requires to operate. We are talking about limiting a 170W cpu to 65w for efficiency but this will drop the performance very much and it may not be worth paying so much for performance which will not be there any longer.
The other thing is, we are not talking hear about tweaking the CPU but limit its power. That is also different.
Like i said, I undervolted my 6900xt and it runs cooler, uses less power but the performance stayed exactly the same.
Than you have a CPU which uses 170w and you cap it at 65w loosing lets say 30% of performance, since you just want to have a efficient CPU for which you pay a lot money. Now, you pay for performance or for the tweaking? I don't see a gain here but regression in performance instead by lowering power usage exponentially

ratirtGreat but that is not the point we've been arguing about. everyone knows about the efficiency and power a CPU requires to operate. We are talking about limiting a 170W cpu to 65w for efficiency but this will drop the performance very much and it may not be worth paying so much for performance which will not be there any longer.
The other thing is, we are not talking hear about tweaking the CPU but limit its power. That is also different.
Like i said, I undervolted my 6900xt and it runs cooler, uses less power but the performance stayed exactly the same.
Than you have a CPU which uses 170w and you cap it at 65w loosing lets say 30% of performance, since you just want to have a efficient CPU for which you pay a lot money. Now, you pay for performance or for the tweaking? I don't see a gain here but regression in performance instead by lowering power usage exponentially

Solid State BrainI am not sure what real-world consumer scenarios there are for using a 170W TDP CPU (with even higher peaks) at 65W besides circumstances with severe cooling or power constraints, or just efficiency for the sake of efficiency.

Drawing from personal experience, I was referring more to using it at a more reasonable 100-120W, which will likely yield most of the MT performance at a significantly lower heat output, making cooling manageable with ordinary coolers. Probably many people in recent years got spoiled by affordable commercial AIO water cooler kits, but even a high-end air cooler will have troubles efficiently dissipating 170-180W+ without getting itself noticed.

Depending on your priorities, it can be worth it to trade 10-15% MT performance in exchange for a lower power draw. Perhaps if one used the CPU in a continuously operating rendering cluster of some sort, even losing more (even 25-30%) would be acceptable as well as long as efficiency increased.

fevgatosNope, i was talking about limiting it to 105w, the same as the 5950x was.

Do you think people should have been runnimg their 5950x at 170 watts instead of stock? Cause if they didnt they would be losing perfromance.

Valantar... or we just accept that there exists a thing called a stock configuration, that this is an inherent trait of the product available to consumers, and that stock-to-stock comparisons and metrics are thus the most relevant baseline comparison for any performance measurement, including relative efficiency. Stock settings are a configurable variable inherent to the product just like the number of cores, clock speeds, cache, accelerators, ++++, so it's just another part of the product.

Solid State BrainAn issue with this is that over the years motherboard manufacturers have made it impossible to set a real "stock configuration" or "CPU manufacturer default" in the consumer space. Motherboards will come with unlocked limits by default, or even use enhanced frequency settings when possible, pumping up voltages to keep everything stable.

Another issue is that even if one CPU manufacturer enforced stricter motherboard defaults, the other could easily take advantage of this self-limitation by allowing its own partners to configure motherboard settings as they wish, for better "stock" benchmark scores in the reviews.

As I've mentioned elsewhere, it's on the reviewers to come up with improved and fair testing methodologies that don't simply use "motherboard defaults" (often inadequate) or unlocked limits, but it takes more time and research, i.e. money.

Valantar[...] And Intel with ADL decided to throw everything out the window and just set PL1=PL2=240W for their K SKUs, despite nominal 125W TDPs

Valantar[...] reviewers should follow chipmaker defaults for chip testing, as that's just a variable that needs to be controlled for.

Solid State BrainI think PL1=PL2=240W must have been just in the reviewers' guides provided by Intel, possibly in expectation that every motherboard can and will behave differently depending on default settings.

The publicly available ADL spec sheet recommends PL1=125W, PL2=241W and Tau=56s. These are not mandated, and there are and even suggestions of using lower PL limits and Tau time if the cooler/system cannot handle them.



Curiously, the true Intel hardware default (as per spec sheet notes) for PL2 is supposed to be PL1 * 1.25, with Tau=1s. That means that for a 125W TDP processor, PL2 would be 156W.

ValantarThis is all kind of going in circles though. If you can afford the higher core count CPU, want an extremely efficient CPU rather than peak possible performance, and have the time, knowledge and skills to tune it to a lower power level, by all means do so. That's pretty much a given - unless you go under whatever the tipping point for efficient operation for the CPU is, where it starts being truly starved for power, you'll gain efficiency as you drop in power.

It's like buying Intel's 35W T CPUs back in the day - you got the same core count as the non-T variants, but a much lower power limit and accompanying lower clocks (but higher efficiency) - but you also paid roughly the same as for the non-T SKUs.

The question is the relevance of stock settings, and the effect of their power levels. @fevgatos argues that they are irrelevant because anyone can tune their chips to be more efficient. This, IMO, is nonsense, as we know that the vast majority of users have neither the time, knowledge, skills or desire to do this tuning, meaning stock settings determine the behaviour - and thus efficiency - of the vast majority of CPUs actually used.

Of course, if you do have the time, knowledge, skills and desire to do this, you can gain a lot. My 5800X performs better at the 110W PPT I've got it running at than it does at its stock 138W PPT, for some weird reason. Modern boost algorithms and their interactions with thermals, current, and core hopping for ST tasks (and the lack of core hopping for nT tasks) makes the relationship between settings and performance more complicated. But you can't expect the average user to understand, let alone do anything about any of this. And that's what @fevgatos refuses to acknowledge.

I think the clock speed gains AMD are showing here are really impressive, but I still don't like the TDP jumps - but they're essentially forced into those by the competitive situation, seeing how Intel has already gone there (and beyond). Luckily, most consumer applications are not anywhere near continuous nT loads, and will thus not need all of this power. Which of course complicates the efficiency picture even further.

RandallFlaggWhen will the real AMD fans stand up and make a thread about how to get higher clocks, more power, better cooling, and better benchmarks?

This like watching a car enthusiast forum where everyone is talking about swapping in taller gears (less acceleration) and disabling cylinders on their car to get more MPG.

ratirtIf you measure with the 3 metrics you evaluate the processors value with the 3 metrics but if you follow our friend @fevgatos argument, that the CPU is efficient when you extremely limit the wattage of a CPU, lowering drastically its performance, the conclusion here is, there are no inefficient CPUs

R0H1TIs that supposed to be a joke? Did you miss the once in "half a millenia" drought in Europe, due to Global warming :rolleyes:

At this point we should just lump all the climate change deniers with their flat earther brethren, not saying you are a denier (since I don't know your position) but spending more power for very little to virtually no benefit these days makes zero sense!