Discussion Redefining the boost frequency of AMD processors - THE REAL BOOST

[Kocicak]

I bought a 12 core Ryzen processor with claimed boost frequency of 4600 MHz. I does reach this frequency with extremelly light load, but under sustained light load it appears, that only two cores run at around 4550 MHz. It feels like I got less than I payed for.

It may be usefull to use the boost frequency as a real meaningfull parameter. I propose this definition:

Boost frequency is the frequency at which at least third of the cores of the processor can run with sustained load.

The above mentioned 12 core processor would need to have four cores capable of sustained running at 4600 MHz to be sold as having the boost frequency 4600 MHz. My processor has "the real boost" just 4500 or even 4450 MHz.

It sound fair and reasonable, at least to me. What do you think?

 

[Atari2600]

Hence why they always say "boost up to"

 

[Topweasel]

Boost and especially any PBO driven number is only ever meant to be limited time thing. Even with Intel, unless you have a retail board, have time limits imposed on their Turbo 2 and up clocking. It's only because the people posting on the forums aare all running on retail boards, which have most of the turbo stuff set as unlimited (also with things like MCE enabled) that we tend to see different activity. All CPU's now are designed not around sustained clock speed but the idea of finish quickly. Hell thats all PCIe 4.0 M.2's work around. Do everything as absolutely as quick as possible at first trying to shave as much time off the job as possible and wind down from there. Hell as mentioned on this very forum about first gen Ryzen's X offset. It wasn't about anything other then playing with the fan curve so that XFR could run for a few seconds to minutes longer. With Zen+ and Zen 2 marketing slides AMD has been pretty clear how the turbo's work, people just ignored them because we are used to Intel's overclockability and therefore the chips ability to hold what are supposed to be temp clocks for as long as you have the cooling (which there isn't really a good enough cooling to keep 4600 on Zen 2) when you turn the protections off.

 

[Kocicak]

I believe that PBO is more about all core load and maximal power consumtion of the CPU, when running light loads this simply does not apply.

Say, this CPU can run 4 cores at 4600 MHz at 1,5 V and the cores would degrade after XX hours of working under these conditions.

OR this cpu can run 4 cores at 4400 MHz at 1,3V and the cores would degrade after XXXX hours, which means the CPU in normal usage scenario would last some expected normal lifetime of a consumer CPU.

I would advertise the CPU as having boost frequency of 4400 MHz and be done with it. Everybody would be happy.

Perhaps the best thing would be to advertise two boost frequencies:

• The first with sustained load of (at least) one third of the cores.
• The second with time limited load of (at least) one sixth of the cores.

The absolute maximum frequency which any of the cores can reach without doing any real work has no meaning at all.

 

[uzzi38]

Say, this CPU can run 4 cores at 4600 MHz at 1,5 V and the cores would degrade after XX hours of working under these conditions.

OR this cpu can run 4 cores at 4400 MHz at 1,3V and the cores would degrade after XXXX hours, which means the CPU in normal usage scenario would last some expected normal lifetime of a consumer CPU.

Voltage is not the only thing that affects degradation. You should really read The Stilt's explanation of voltages with Zen 2, but long story short in high current workloads (where all cores are boosting and doing so in a very strenuous situation like Blender) the voltage limit is around 1.325V. For low-current workloads (where only 1-2 cores are boosting in again, a strenuous situation) the voltage limit is instead around 1.47V. All of this depends on the piece of silicon itself.

Perhaps the best thing would be to advertise two boost frequencies:

• The first with sustained load of (at least) one third of the cores.
• The second with time limited load of (at least) one sixth of the cores.

The absolute maximum frequency which any of the cores can reach without doing any real work has no meaning at all.

While I certainly wouldn't mind better clarification on what determines boost, there's no reason to bother. The way 3rd Gen Ryzen acts is in-line with both Intel and AMD's definitions (and I mean actual definitions, not what we've experienced with previous CPUs) of boost since boost became a thing.

 

[Kocicak]

I have been thinking about it during the lunch and as a consumer I want to know these four frequencies of the processor:

• all core sustained load at rated TDP
• all core sustained load at 150% of rated TDP
• sustained load of at least one third of the cores
• time limited load of at least one sixth of the cores

Ryzen 3900X would be characterised with for example this set of frequencies: 3800/4100/4300/4500 and I would know what I am getting and would be guaranteed to get what I pay for.

I would also want the turbo switch on the computer case to switch between 100% and 150% of the power limit of the processor.
THE RETURN OF THE GOOD OLD TURBO SWITCH!!! Yea!

 

[Gideon]

• all core sustained load at rated TDP
• all core sustained load at 150% of rated TDP

Well, to be fair the first thing they need to do is redefining TDP (or use another metric entirely) as the definition is abusrdly circular in nature.

If AMD were trying to put out a useful metric for their CPUs, they'd just spec the PPT since that is the power drawn by the part and therefore the heat a cooler needs to be able to dissipate continuously. Of course, this'd make their 65 W TDP parts 88 W ones and 105 W TDP parts become 141 W ones, which looks worse in marketing.

I'm ok with it and IMO this is what they should do. If they are worried about comparisons to intel, think of a new acronyme (no need to use "TDP").

 

[DrMrLordX]

We already had a thread like this awhile back. This chip is not an Intel CPU. Sorry, no 4.6 GHz all-core overclocks (I was hoping for one as well, but my hopes are dashed). It's still really, really fast. AMD already covered their assets so there's no point in complaining about having the fastest "consumer" desktop CPU in the world (until the 3950x comes out).

If anything, I would be more upset about PBO being mostly worthless.

 

[Kocicak]

Perhaps it would be necessary first to define some "standard workload." I do not know enough about computing to know what operations a CPU can do, and what sorts of operations for example Cinebench does.

After this standard workload would be defined, power consumption of the processor (however called) would be simply defined as power consumption running standard workload at given frequency.

 

[DrMrLordX]

Perhaps it would be necessary first to define some "standard workload."

There can't be one. If you examine the voltage tolerances of the 7nm process as implemented in Matisse, you will begin to see why the boost behavior is so wonky. The amount of voltage you can "safely" feed the chip is entirely related to current draw and temperature. You can feed up to 1.47v handling background threads, while hitting TDC/EDC limits restricts you to about 1.325v or so. Actual FIT tables vary from chip to chip.

Anyway, your typical Matisse requires somewhere around 1.45-1.47v to hit 4.6 GHz reliably, assuming it can even go that high (e.g. it's a 3900x). So you can't really expect that speed under normal boost behavior when any more than one core is engaged, because engaging more cores by necessity increases current draw, causing the dynamic boost map to draw voltage limits down according to the FIT tables. You CAN manipulate this process somewhat by fiddling with voltage (and therefore, temperature) in ways that are invisible to the boost algorithms - by manipulating LLC settings or by applying voltage offsets. In general, default boost behavior likes LOWER voltage than expected, while PBO likes HIGHER. In general. You can also manipulate voltage in such a way as to improve MT boost behavior (default) while hurting ST boost behavior.

The other thing to consider is that temperature has an enormous effect on the voltage required to achieve stability on Matisse. Der8auer documented this effect shortly after launch. The boost algorithm can safely try out higher clocks without (necessarily) pumping more voltage if it stays cool during operation. That's one of the reasons to use negative voltage offsets and/or to encourage vdroop with low LLC settings. The boost algorithm is a bit volt-happy, so if you can get it to run cooler by just undervolting the thing without the algorithm knowing it's going on, it'll see lower-than-expected temps and attempt higher clocks. If it works you get more performance and if it doesn't you get clock stretching (ugh).

Anyway, going back to temperature . . . boost behavior is governed by hotspot temperatures, not temperatures measured at the edge of the die. Temp monitoring programs won't show this since AMD chose to report only the hottest temps on the CCDs to you (HWiNFO64 can report per-CCD temps, if you want them). That's a good thing since it prevents wingnuts from overriding or otherwise ignoring hotspot temp limits risking damage to their CPUs. Radeon VII taught those of us who bought it to respect the hotspot temps above all else. Tdie is meaningless in this brave new sub-10nm world. Those hotspots are small, and it's really hard to cool them. I want you to consider that my watercooling solution is complete overkill and can probably handle ~1kW heat load with a coolant temperature delta of 5C. And that's a conservative estimate. A 3900x in default behavior mode can't sustain more than 142W through the socket under any circumstance - it'll power throttle. If I were cooling nothing but a hot plate with uniform heat flux across its entire surface and a total heat flux of 142W, the block would probably keep it within 10C of ambient, if not lower. Instead, if I run Prime95 SmallFFTs which pushes that chip to 90% of 142W (127W - I think my voltage adjustments are messing with the boost algorithm's ability to max out the chip) and I let it run for a wee bit, it'll hit temps of 58C with an ambient temp of ~28C according to the motherboard case temp sensor. It appears as though adding more cooling capacity wouldn't help. The loop is now limited by my inability to get heat away from the specific hotspots on the chiplets quickly enough for the loop to do anything with it. The chiplets have to reach temps like 58C before they stop heating up. Oh, and even worse: according to HWiNFO64, nearly half of the 3900x's power consumption isn't even represented by the chiplets themselves. The rest of it is the I/O die. During this activity of Prime95 SmallFFTs @ ~3.9 GHz on all cores, CPU core current averages ~61a and CPU core power averages ~70W. So in actuality, a heat load of 70W generates temps of 58C with a massive ambient cooling solution. Holy crapoly. We are going to need something a lot better than indium solder and copper heatspreaders to get that chiplet heat spread out and into our cooling solutions. If we had that, cooling the chiplets would be trivial, and we could achieve reliably higher clockspeeds at voltages of 1.325v-1.36v which in turn would yield higher boost clocks in MT scenarios. It would also make overclocking a lot more fun.

 

[Kenmitch]

What do you think?

Let the dead horse rest in peace. I see no reason to keep beating it.

It feels like I got less than I payed for.

What are your intended uses for the 3900x?

It's most likely the fastest current offering for the AM4 platform no matter how you slice the pie.

 

[Pohemi]

Well, to be fair the first thing they need to do is redefining TDP (or use another metric entirely) as the definition is abusrdly circular in nature.

I'm ok with it and IMO this is what they should do. If they are worried about comparisons to intel, think of a new acronyme (no need to use "TDP").

I believe Intel undervalues the TDP on their processors as well by listing the optimal TDP at NON-boost clocks, not the maximum value at maximum boost. Someone correct me if I'm mistaken.

 

[UsandThem]

Let the dead horse rest in peace. I see no reason to keep beating it.

This +++++++

 

[Kocicak]

DrMrLordX - I undestand that there is variability between chips and that the temperature is important. The specs of a product are (or should be) about guaranteed performance, the variability should just cause random better performance than specified.

An example:

Here is a product Ryzen 3900X, 12C/24T, with SPC = 105W (standard power consumption) and these specified frequencies: 3800/4100/4300/4500 MHz.
The specifications are valid only when running standard workload (SWL) at CPU temperature below 85°C.
The cooling solution we provide will keep the processor at 75°C in 21°C ambient temperature while running SWL at SPC.

You as a consumer are guaranteed this:

1) While running 24 threads of SWL at SPC (sustained), procesor will run at 3800MHz OR QUICKER. Or when running 24 threads of SWL at 3800MHz, power consumption will be lower than SPC. When using provided cooler, the ambient temperature must be lower than 31°C for this to be true.

2) While running 24 threads of SWL at 150% SPC (sustained), procesor will run at 4100MHz OR QUICKER. Or when running 24 threads of SWL at 4100MHz, power consumption will be lower than 150% of SPC. I need to use my own cooler and keep CPU below 85°C for this to be true.

3) While running 8 threads of SWL (sustained), at least 4 cores of the procesor will run at 4300MHz OR QUICKER. I need to keep CPU below 85°C for this to be true.

4) While running 4 threads of SWL (time limited), at least 2 cores of the procesor will run at 4500MHz OR QUICKER. I need to keep CPU below 85°C for this to be true.

You know what sort of load is your worst case scenario, if it is lower than SWL and you can count on these specs, or higher. Naturally SWL should be as high as attainable, so that most users know that the specs are valid for them.

All the sample variability is hidden in "or quicker", "lower than", "at least".

You, as a consumer, know what you are getting and you are guaranteed to get this performance, if the CPU will stay under specified temperature.

You know what you are getting and you are guaranteed to get this performance - this is the point of what I am trying to do here.

 

[DrMrLordX]

@Kocicak

I don't know what else to tell you. AMD can't produce a CPU like that using TSMC 7nm HP. Not unless you want a 1C chip that can and will boost to 4.6 GHz all the time. It won't happen otherwise. And if you really want good boost performance from the chip, 85C is an awful CCD temperature. You seem to think a 3900x should run 3.8 GHz minimum in a 24T workload, but if you have the wrong AGESA version and the wrong LLC settings then heavy AVX2 workloads won't even run that speed. Furthermore, AMD will never make any guarantees about what the chip will do running "150% of SPC" or ~160W.

 

[Kocicak]

Every single number I wrote were just examples how proper CPU specification in my opinion should look like, in order to achieve, beside other things, that the boost frequency is a frequency at which the CPU can actually do some work and it is not just some number without any real meaning, which some CPUs dont ever reach and some do, but do not do much computing at.

 

[Hitman928]

I prefer the way AMD is doing boost now once it's properly understood. You get the very best possible performance given your cooling and silicon quality. I would stop worrying about frequency to such a minute degree. The performance is what matters in the end. Look at reviews and compare performance between CPUs and make your decision from there.

 

[UsandThem]

I prefer the way AMD is doing boost now once it's properly understood. You get the very best possible performance given your cooling and silicon quality. I would stop worrying about frequency to such a minute degree. The performance is what matters in the end. Look at reviews and compare performance between CPUs and make your decision from there.

But he already bought the 3900X, and states it usually only boosts to 4550 MHz. He feels that "It feels like I got less than I payed for".

It sounds like someone bought a $500 CPU, and instead of enjoying it, is focusing on things that were already discussed (and fixed with a BIOS update) before he even bought the CPU. Based on their posting history here, they were well aware of its boost speeds on various loads, yet they still gladly bought it a couple days ago.

Sorry my bold key got stuck on certain parts when writing that. I imagine it is very distracting to read.

 

[Kocicak]

Sorry my bold key got stuck on certain parts when writing that. I imagine it is very distracting to read.

When I saw your text I immediatelly remembered a pair of dalmatine dogs I saw in the park last week!

 

[bluechris]

I prefer the way AMD is doing boost now once it's properly understood. You get the very best possible performance given your cooling and silicon quality. I would stop worrying about frequency to such a minute degree. The performance is what matters in the end. Look at reviews and compare performance between CPUs and make your decision from there.

Half of me agree with you but the other half says this is wrong and i explain.
I put up a 3600 in a GB x570 Pro and this machine will be my esxi homelab server.
Since it was my first experience with ryzen i had read half the internet and from what i understood this chips needs cooling.
So i installed win10 for testing purposes and i tried hard 1st to get the best from the stock cooler.
No matter what i did, i got max 4150 single core and 3900 allcore with the cpu reaching 90c.
So after that and since i got a ton of water cool components and i was missing only a cpu block, i order an EK.
Long story short, with a D5 pump, a GTX360 rad in push pull configuration i did some tests and no matter what i did i only gained 50-75hz. My single core went to stable 4200mhz and allcore to 3975.
The above after 2 days of constand testing and trying and in all core tests the cpu temp never climbed above 64c.
So to summarize, my opinion is to get a good AIO or better a good Noctua heatsink or anything equivalent, enable PBO and call it a day.

There is no point trying to get something extra from this cpu's because they got so many automations inside them that if you try to go to push always something holds it and even if they are unlocked, i believe that are locked with their own way.

That's my 2 cents

 

[Thunder 57]

I bought a 12 core Ryzen processor with claimed boost frequency of 4600 MHz. I does reach this frequency with extremelly light load, but under sustained light load it appears, that only two cores run at around 4550 MHz. It feels like I got less than I payed for.

It may be usefull to use the boost frequency as a real meaningfull parameter. I propose this definition:

Boost frequency is the frequency at which at least third of the cores of the processor can run with sustained load.

The above mentioned 12 core processor would need to have four cores capable of sustained running at 4600 MHz to be sold as having the boost frequency 4600 MHz. My processor has "the real boost" just 4500 or even 4450 MHz.

It sound fair and reasonable, at least to me. What do you think?

Considering you wanted a "3600 x 2" for $249, I have a feeling you will always feel like you got less than what you paid for. Can't you just be happy with an awesome CPU? Or you could always buy Intel you know.

DrMrLordX - I undestand that there is variability between chips and that the temperature is important. The specs of a product are (or should be) about guaranteed performance, the variability should just cause random better performance than specified.

An example:

Here is a product Ryzen 3900X, 12C/24T, with SPC = 105W (standard power consumption) and these specified frequencies: 3800/4100/4300/4500 MHz.
The specifications are valid only when running standard workload (SWL) at CPU temperature below 85°C.
The cooling solution we provide will keep the processor at 75°C in 21°C ambient temperature while running SWL at SPC.

You as a consumer are guaranteed this:

1) While running 24 threads of SWL at SPC (sustained), procesor will run at 3800MHz OR QUICKER. Or when running 24 threads of SWL at 3800MHz, power consumption will be lower than SPC. When using provided cooler, the ambient temperature must be lower than 31°C for this to be true.

2) While running 24 threads of SWL at 150% SPC (sustained), procesor will run at 4100MHz OR QUICKER. Or when running 24 threads of SWL at 4100MHz, power consumption will be lower than 150% of SPC. I need to use my own cooler and keep CPU below 85°C for this to be true.

3) While running 8 threads of SWL (sustained), at least 4 cores of the procesor will run at 4300MHz OR QUICKER. I need to keep CPU below 85°C for this to be true.

4) While running 4 threads of SWL (time limited), at least 2 cores of the procesor will run at 4500MHz OR QUICKER. I need to keep CPU below 85°C for this to be true.

You know what sort of load is your worst case scenario, if it is lower than SWL and you can count on these specs, or higher. Naturally SWL should be as high as attainable, so that most users know that the specs are valid for them.

All the sample variability is hidden in "or quicker", "lower than", "at least".

You, as a consumer, know what you are getting and you are guaranteed to get this performance, if the CPU will stay under specified temperature.

You know what you are getting and you are guaranteed to get this performance - this is the point of what I am trying to do here.

Do you really think Joe Sixpack is going to care or want to go into that amount of detail, if it were even possible? I couldn't even get through it before I went "TLDR".

But he already bought the 3900X, and states it usually only boosts to 4550 MHz. He feels that "It feels like I got less than I payed for".

It sounds like someone bought a $500 CPU, and instead of enjoying it, is focusing on things that were already discussed (and fixed with a BIOS update) before he even bought the CPU. Based on their posting history here, they were well aware of its boost speeds on various loads, yet they still gladly bought it a couple days ago.

Sorry my bold key got stuck on certain parts when writing that. I imagine it is very distracting to read.

Bold key? Genius! Sure got a laugh out of that one. Lucky the "AnandTech Mean Moderator" hasn't read this, because, as was pointed out multiple times by multiple people, that bold thing is annoying. But at least you are using the same size font everywhere now, so that's a start.

@uzzi38, I can't help but hear Arnold Schwarzenegger say "Uzi 9mm" when I see your name.

 

[Guru]

I haven't see a person more ignorant and foolish than OP, who cares about what core boosts to what exact and specific frequency, who cares if its boosting core 1 or core 5 to a certain speed, for 5 seconds or 30 seconds. You have the overall performance of the performance and depending on system variables, OS variables, silicon variables it will perform +/- 2% faster/slower. There are literally 50+ reviews to read from and learn what kind of performance to expect.

You just bought your cpu and you are complaining, next time bother to read few reviews.

 

[ClockHound]

I bought a 12 core Ryzen processor with claimed boost frequency of 4600 MHz. I does reach this frequency with extremelly light load, but under sustained light load it appears, that only two cores run at around 4550 MHz. It feels like I got less than I payed for.

Perhaps you're trying to compensate for buyer's remorse and feel your e-peen has drooped about 1%. Disappointment is in the perception of the beholder.

If you don't have any tasks that utilize 24 threads, one option is to sell your 'operating as specified' 3900X for $100, have the Intel Influencers Fund cover the difference and give you a 9900KS space heater to fuel your unmitigated computing rants - 14% less ranting after mitigation.

 

[Ajay]

Let the dead horse rest in peace. I see no reason to keep beating it.

This +++++++

All these +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Damn, seriously, this thread is just plain nutz!

 

[DrMrLordX]

Every single number I wrote were just examples how proper CPU specification in my opinion should look like,

Yeah but it misses the point I'm trying to make, that you can't peg things down based on # of cores and clockspeed like that. What kind of cooling do you have? Does the code rely on AVX2 instructions? etc. The specification you already outlined is infuriatingly complicated - enough that openly advertising something like that would probably reduce sales. It still doesn't "cover all the bases". AMD doesn't want to get sued because the chip can't/won't hit 3.8 GHz on some PrimeGrid workload because a particular end-user's motherboard has default LLC options that make it boost to 3.77 GHz using the stock cooler. AMD opted to just tell us that it boosts "up to 4.6 GHz" and leave it at that. Which is what the CPU does (finally).