Discussion CPU boost frequency and marketing

[IEC]

Out of curiosity to see if I could reproduce these results, I ran the latest build of HWinfo64 in sensors-only mode (to attempt to avoid the "observer effect" from running it in default mode).

I'm not sure if the latest updates fixed clock detection or if I simply didn't notice it before, but the results I'm seeing with light desktop usage are actually transiently exceeding the advertised 4.6GHz boost now...


The only other thing I changed software-wise was installing the latest nVidia driver update. Doubt that changed anything, though.

 

[tamz_msc]

Out of curiosity to see if I could reproduce these results, I ran the latest build of HWinfo64 in sensors-only mode (to attempt to avoid the "observer effect" from running it in default mode).

I'm not sure if the latest updates fixed clock detection or if I simply didn't notice it before, but the results I'm seeing with light desktop usage are actually transiently exceeding the advertised 4.6GHz boost now...
View attachment 8792

The only other thing I changed software-wise was installing the latest nVidia driver update. Doubt that changed anything, though.

What happens when running CPUZ benchmark for example?

 

[Markfw]

What happens when running CPUZ benchmark for example?

Who cares ? we have at least 2 examples where its boosts as advertised.

 

[coercitiv]

So far — yet with a small sample size — over 60% of the poll voters want the stated boost frequency to be achievable in prolonged workloads (Cinebench Single-Thread, as suggested in the poll). So what should AMD do?

Do you reckon the stated boost frequency is achievable for prolonged periods of time for Intel CPUs with different 1C and 2C boost speeds?

I just ran CB 15 ST on a locked CFL CPU with 4.6Ghz 1C and 4.5Ghz 2C max turbo. The CPU boosted right away to 4.5Ghz but it took more than half the benchmark to register 4.6Ghz on a core for a brief period of time which didn't really affect average clocks for that core. It was also the only core which reached 4.6Ghz, even though the Win scheduler moves the workload around periodically. It took affinity locking for both CB15 and HWiNFO to obtain longer 4.6Ghz boost on 1 core, but even then the results were unreliable.

If anyone here thinks 1C boost speeds should be achievable for anything but opportunistic & brief workloads, then you're roughly a decade too late.

 

[IntelUser2000]

If anyone here thinks 1C boost speeds should be achievable for anything but opportunistic & brief workloads, then you're roughly a decade too late.

Yes.

I understand the concerns, but its really not a big deal. If you account for the 100MHz being off, then people are at best, complaining for 50MHz difference.

Other problems are much bigger. Like the one that VirtualLarry had. Or
GhostPirateLeChuck it wouldn't even stay at base frequencies properly. Or immature BIOS.

Zucker2K was saying on Cinebench and other benchmark you can reach the 1C frequency. So? I delved into that quite deeply with Lynnfield and Sandy Bridge. The 1C frequencies, even if the monitoring application says its reaching it, doesn't mean its doing it all the time. There are minor differences. 3GHz fixed frequency CPU is better than a 2.5GHz Base 3GHz Turbo CPU for the same reason.

 

[coercitiv]

Other problems are much bigger. Like the one that VirtualLarry had. Or
GhostPirateLeChuck it wouldn't even stay at base frequencies properly. Or immature BIOS.

As I said in another thread, call them out for immature BIOS at launch, maybe even for poor QC if hot or dead chips are a statistically significant problem, but reserve judgment on max boost clocks until firmware issues are sorted out. My personal speculation is motherboard makers may have influenced this issue as well, but we need to wait for facts. (Latest data seems to suggest Zen 2 changes VID based on the voltage it is being fed, the tests were done for undervolting, but the mechanic may work for overvolting too to a certain degree. If this is true, mobo makers can't really influence chip voltage while in Auto mode, as the chip will simply ask for less/more until it gets it's way.)

However, in my view expecting 1C boost clocks to be reliable in systems running multiple pieces of software at once is a stretch. Unless 1C boost is identical to 2C boost then all we're dealing with is a number that looks nice on the spec sheet.

 

[PotatoWithEarsOnSide]

Does it matter whether X performance on [insert your own benchmark here] was scored using Y or Z clocks?
Performance is what counts, not the clocks that are being recorded whilst achieving that performance.

 

[tamz_msc]

If anyone here thinks 1C boost speeds should be achievable for anything but opportunistic & brief workloads, then you're roughly a decade too late.

This is AMD's position on how boost clock works:

According to this definition, the 3900X should be doing 4.6 GHz on at least one core as long as the above limits aren't exceeded, depending on the type of workload. The boost algorithm is opportunistic, which means that it should sense any workload that is running for a nontrivial amount of time and put at least one core at 4.6 GHz if the parameters which are a factor in determining the boost clocks are lower than their threshold(limiting) values. Now Ryzen Master shows what these constraints are - along with temperature and Vcore(which is a characteristic of the silicon quality), the parameters are PPT, TDC and EDC. So far, so good.

The problem which arises is this - even if your are far from the PPT, TDC and EDC limits and cooling is sufficient(without getting into sub-ambient temperatures), the max boost is not reached during a typical workload because in order to do so the boosting algorithm tries to increase clocks by increasing Vcore as high as possible(because that is what AMD states is supposed to happen with low current loads). But doing so makes the CPU run up against its silicon characteristics, which is ultimately the deciding factor on whether the advertised boost clocks are achieved or not.

Thus it ultimately boils down to silicon lottery whether max boost is achieved or not, and getting those max boost clocks means an ungodly amount of voltage pumped into the best performing core(s) for a brief amount of time. It's unreasonable to expect that that's the way boost clocks are supposed to work, and that is precisely the reason why the likes of der8auer have criticized AMD on this issue.

 

[PotatoWithEarsOnSide]

To me, it seems as though the Ryzen CPUs aren't designed to be using just the one core at any given time, since you can pretty much guarantee that a single core isn't going to hit any of those limits (except VCore) on its own.
When you have multiple cores under load then the set limits do appear to perform a function.
In theory, what they have designed does actually work, but in practice you're always hitting the VCore limit on a single core well before any of the other limits ever come into play.

 

[coercitiv]

The problem which arises is this - even if your are far from the PPT, TDC and EDC limits and cooling is sufficient(without getting into sub-ambient temperatures), the max boost is not reached during a typical workload because in order to do so the boosting algorithm tries to increase clocks by increasing Vcore as high as possible(because that is what AMD states is supposed to happen with low current loads). But doing so makes the CPU run up against its silicon characteristics, which is ultimately the deciding factor on whether the advertised boost clocks are achieved or not.

So what you're trying to say in many words is you think AMD failed to correctly validate some chips for real world operation at max advertised clocks. And what I'm trying to say in just as many words is wait for mature firmware and then, under more controlled conditions, decide whether silicon quality was good enough or not.

 

[moinmoin]

So what you're trying to say in many words is you think AMD failed to correctly validate some chips for real world operation at max advertised clocks.

My impression so far is that due to binning in the worst cases only one core per CCX appears to be capable to reach the advertised max boost. You then have the possible double blow of both the new 2-1ms fine grained frequency management allowing to change speed faster than monitoring software may be able pick up and the Windows Scheduler being blissfully unaware of different core quality and as such regularly removing threads from good cores again.

And what I'm trying to say in just as many words is wait for mature firmware and then, under more controlled conditions, decide whether silicon quality was good enough or not.

There is certainly a learning effect. It will be interesting to see whether that will be more in the audience and third party software or in AMD's firmware and the board manufacturers' implementations.

 

[lopri]

Does it matter whether X performance on [insert your own benchmark here] was scored using Y or Z clocks?
Performance is what counts, not the clocks that are being recorded whilst achieving that performance.

Performance is very inconsistent. It varies run by a run in more than what I am used to seeing in other CPUs. AMD should at least market it differently.

 

[lopri]

My impression so far is that due to binning in the worst cases only one core per CCX appears to be capable to reach the advertised max boost. You then have the possible double blow of both the new 2-1ms fine grained frequency management allowing to change speed faster than monitoring software may be able to pick up and the Windows Scheduler being blissfully unaware of different core quality and as such regularly removing threads from good cores again.

Windows scheduler is a separate issue. You can lock affinity to one core and use that core only, it will still not boost to the max advertised frequency as you'd expect. It may be that new firmware can counteract the worst behavior, but in my view, this ultimately comes down to physics: The chips are pushed too hard out of the factory.

 

[lopri]

Apparently, Intel is following suit.

https://translate.google.com/transl...intel-i3-9350k-se-fait-un-passage-sous-froid/

Before giving you the first results, we recorded the temperatures during a Cinebench R15 as well as the actual Boost frequency. Strangely, it does not exceed 4.4 GHz while temperatures are relatively low. Hard to say if it is a concern for recognition at the bios of our motherboard or a processor bug. If you take a closer look, the parameters are set to have 1 core at 4.6, a heart at 4.5 and two cores at 4.4 GHz.

According to the author, the 9350K fails to reach the advertised Turbo Boost frequency (4.6 GHz) out of the box. The chip, however, manages to clock up to 5.35 GHz under water.

 

[Kenmitch]

Apparently, Intel is following suit.

https://translate.google.com/translate?sl=auto&tl=en&u=https://vonguru.fr/2019/07/20/test-le-processeur-intel-i3-9350k-se-fait-un-passage-sous-froid/



According to the author, the 9350K fails to reach the advertised Turbo Boost frequency (4.6 GHz) out of the box. The chip, however, manages to clock up to 5.35 GHz under water.

That settles it....I'm taking my rig out back, soaking it with gasoline and lighting it on fire! /s

The common denominator would be Windows....Hmm.

 

[Hitman928]

Apparently, Intel is following suit.

https://translate.google.com/translate?sl=auto&tl=en&u=https://vonguru.fr/2019/07/20/test-le-processeur-intel-i3-9350k-se-fait-un-passage-sous-froid/



According to the author, the 9350K fails to reach the advertised Turbo Boost frequency (4.6 GHz) out of the box. The chip, however, manages to clock up to 5.35 GHz under water.

I think, as others have mentioned, the main culprit is CPUs with a single core boost that is higher than a dual core boost combined with how Windows handles core loading over time. If you have a single core boost that really is just for when a single core is active, the time when this is true within a Windows environment is almost never. Add that to 8 core chips where maybe only 1 or 2 of the cores can even hit that speed and it will be even lest time spent at that boost frequency. I remember the original Ryzen 1700 having the same issue with it's 3.7 GHz single core boost. I think people didn't care as much back then because you didn't buy the 1700 at all for it's single core frequency but the issue was still there. Today, the same issue now applies to the highest AMD SKUs so people are noticing more.

With that said, does it really matter? I mean, what are people doing where they are really only using 1 core with no background processes using cpu time. In today's age, I would worry more about the 2 core boost frequency and beyond. If the 2 core boost level fails in a very lightly threaded load environment, I'd be more worried.

 

[fleshconsumed]

Additional 100Mhz to go from 4.5Ghz to 4.6Ghz amounts to 2.2% performance boost. While I think AMD may have stretched technical specification to the limits to arrive to 4.6GHz single core boost, I find it difficult to care about losing on 2.2% in single thread applications. AMD gave us very competitively priced 12c24t CPU at $500. I think it's good that people bring awareness to questionable single core boost numbers being advertised, it's good to have complete picture before making a purchase decision, but I also think this is largely making mountain out of a molehill.

 

[TheGiant]

Where are the 45w hexacores and octocores?

this is the million dollar question

I'm pretty sure that if the vast majority of 9900Ks were only hitting 4.8-4.9GHz single core despite Intel advertising the maximum turbo speed as 5.0Ghz, people would be flaming Intel to a crisp. I don't think it's entirely unreasonable to hold AMD to the same standard.

not many things are reasonable about r3000 startup these days (hello Adored)
remember those quotes- 16C@5GHz@maybe 125W, no no 105W AMD is gonna do it.....

as 3900X owner my take on this- a little more than the thread title says, TLDR - AMD is and egineering not business company

1. what I didnt expect and got it
2. what I did expect and didnt get it
3. what I did expect and got it
4. general marketing


AMD got this time excellent engineering feets, but as excellent is their engineering, bad is their marketing and support of moronic part of their fanbase

1. didnt expect and got it
   1. built in overclocking and tuning
      1. to tune my 6600k to 4,4GHz and undervolt it with RAM latency combination and BIOS setting it took several iterations of doing nothing but spending power on prime95 etc or enjoying a restart while working
      2. AMD pretty much does it for you- saves a lot of trouble and time and with better algorithm than just set fix clocks- performance may be better than with manual OC-not seen before
      3. bottom line
         1. temps/cooling- so far Intel box cooler =crap, AMD ok for default, with 3900X the default cooler is crap and I have pretty much constant ambient of 21C or less
         2. crap regulation
         3. inconsistent performance- not because temps or so, just works that way
         4. problem is, competition is on this or better (with OC)rlevel of performance for 1,5 years 8700K vs 3600, or 8 months (9900K vs 3700/3800X) without supersophisticated algoritm
2. did expect and didnt get it
   1. TBH, 3900X is 2019 version of 12C skylake
   2. I was looking for a replacement of 6600K with 10%+ 4T absolute performance, I didnt get it- it is a trade off, sometimes faster, sometimes not
      1. sometimes 3900X boosts to achieve it, sometimes dont- I dont know if it is a matter of regulation end to end system inconsistency (CPU, sensor, BIOS, system, regulation SW) or what
   3. TDP- a tool of nothing now
      1. my 6600K TDP 91W not undervolted @4,5GHz all auto handbrake wall power-121W, undervolted 4,4GHz (109W), (32GB ram, SSD, z170board, 2TB HDD, gtx760, seasonic 550W PSU from 2006) vs my 3900X default 105W TDP, (32GB RAM, m2 SSD, b450 board, 3TB HDD, RX5700XT and better PSU) 225-250W fluctuating power
      2. so for absolute power- I gained a little- 12C, linear more power-, not 12C, same or little more power like 140W wall power- I expected a mid range same 110W of 8C/16T icelake or so and more with more cores
      3. I didnt buy 9900K because of power, but if I did no mistake
      4. now I got more cores, but less absolute performance in june 2019, but I could get 8 more absolute powerful cores in october 2018- gained nothing with my workload desktop type
      5. I thought in 2019 with 7nm, I get 8C with performance of skylake @4,5GHz with the same wall power as that 6600K so for me this part is a major dissapointment as the state of PC computing, not just AMD, Intel the same
      6. as Icelake looks as major user innovation but the desktop part is non existent, there is nothing to compare
   4. I did expect a better launch
      1. I am not entering BIOS or doing any adjustments because the system works
      2. this is the first time in like 15 years I fear to do any changes because of f..up
3. did expect and got it
   1. price vs absolute performance- Intel can pack, move and hide in the corner with 9920X
   2. no weakness chip 3900x, same as 9900K, but why buying 9900K when i get more cores and same everything or just very little worse
   3. expected a 10% less performance vs 9900K@5GHz and it looks like it or the same with 4T as 6600K @4,5GHz- a match of expectations
4. overall marketing and my own experience
   1. the first year of my first company top lesson learned- the content doesnt sell and the result is not on the first place
      1. people can not see difference even when you create world record unless the have something to compare
      2. if 200km you travel now in 2 hours average for 10EUR and your new hyperloop does it in 1,5hours promised for 7EUR and does it for 1,53hours for 7 EUR it is a problem
      3. an engineer can not understand this, how a result of 30% better and 30% cheaper than the competition is not good enough- yes because people can not differentiate better result with worse unless it is visible, the word visible is important, not result, visible (sorry for repeating)
         1. so if you promised not 1,5 hours and 7 eur but 1,8 hours and 8,5 eur you get happy and customers suddenly get happy with worse result than the actual
         2. an engineer has a problem of understanding this
         3. as I started to promise a little worse, so convservative approach, I earn 10x that much money as before with worse result- my own fact proven by 10 years of owning 7 companies (ATM)
         4. intel is very experienced in this and their TDP is at base clock
            1. as I hated this before because no one just runs at base clock I start to like it
            2. gimme number at super stock, let me do what I can get with mobo/cooling combination, its up to me
   2. AMD made the basic visible things not ok- the Joe user can see basic 2 things- windows task manager frequency and monitoring tool temperature
      1. first is frequency- not reaching it
         1. first part 4,6GHz it reaches with specific workload- here https://www.amd.com/en/products/cpu/amd-ryzen-9-3900x you dont see it
         2. this is a real bull.t o- same as if Intel says max turbo is 5,3GHz but only if you are looking at open command line and observing the "_" , otherwise its just 5GHz- yes you can do it but not when the market is used to someting different
         3. I personally consider the discussion here about sometimes reaching it and sometimes posting screenshot a misstep and a devaluation of technical knowledge of ppl posting it, this is a misstep of AMD
      2. temperature is just not ok- Joe 95C limit, 92C actual=panic
         1. the whole regulation ecosystem doesnt work
         2. it is working by brute force of water cooling/better cooling
   3. summary
      1. AMD promoted this https://www.amd.com/en/products/cpu/amd-ryzen-9-3900x
      2. so ppl dont compare to 9900K + 8T for free, but 4,6GHz boost
      3. suddendly the content what they got is not important, its the promise
   4. what we got with the result
      1. 200 EUR 8700
      2. 330EUR 9900K
      3. 500EUR no competition
      4. autotuning
      5. less MB prices

Sorry for the TLDR post, I just remembered the long nights feeling about what went wrong

 

[gdansk]

I would prefer if they were more conservative in their advertised clock speeds but I understand there was some pressure to reach a high number.

Really, who isn't going to buy a CPU because it's advertised as 4.4GHz not 4.6GHz?

 

[Vattila]

I just realised how AMD dealt with this issue for the Ryzen 1000 and 2000 series: Extended Frequency Range (XFR).

I think I am right in saying that XFR is not an technology in itself. It just a marketing term and describes the difference between the upper bound of the frequency range used by the Precision Boost algorithm and the "Max boost" number printed on the box.

AMD probably foresaw that these opportunistic and brief spikes to maximum frequency happen too rarely under load as to raise questions. So they simply scaled the number on the box back by 50-100 MHz and called the difference XFR.

However, AMD seems to have dropped the XFR notion for Ryzen 3000. Now the "Max boost" number printed on the box is the actual upper limit of the stock operating range.

Why did AMD drop XFR? I can see two arguments:

1. To stop the confusion about the term. Many talked about XFR as if it was a technology implemented in silicon. Precision Boost is the name of the technology. XFR was just a marketing name for the headroom between the conservative Max Boost number printed on the box and the actual limit for stock operation.
2. To be able to print a larger number on the box, i.e. the actual limit for stock operation.

I truly believe AMD fell short of their frequency targets for Zen 2. To allow for XFR (and ensure the advertised Max Boost was seen in actual workloads) they would have had to print numbers on the boxes 50-100 MHz below the already somewhat disappointing frequencies they achieved. Although much of the uplift from Zen 2 comes from the increased IPC, this is not printed on the box, so from a marketing standpoint, AMD may have felt that low Max Boost frequencies would look too bad.

Would you have preferred that AMD had retained some XFR headroom between the number on the box and the actual maximum boost frequency, even if that had meant lower numbers printed on the box?

 

[Hitman928]

You could consider XFR the precursor to precision boost. XFR was a "dumb" algorithm comparatively. Precision boost allows for a much more graceful roll-off in frequency depending on core use. This was shown best comparing the Ryzen and Ryzen+ generations. For some workloads the Ryzen+ processors performed much better than expected because of a smarter boost algorithm. Zen2 seems to take it a step further to try and squeeze out any last ounce of potential out of the CPU. I think this allows for better performance overall in real workloads even if the advertised max boost frequency is rarely, if ever, seen.

 

[Vattila]

You could consider XFR the precursor to precision boost. XFR was a "dumb" algorithm comparatively.

I don't think this is accurate. As you can see in the 1800X slide included my last post, the algorithm was already coined "Precision Boost" for the 1000 series. XFR has always just been a name for the difference between the Max Boost number printed on the box and the absolute max frequency the processor could boost to under ideal conditions.

AMD Senior Technical Marketing Manager Robert Hallock, as I understand it, clearly describes XFR as a marketing term for the ability of the Precision Boost algorithm to gain extra boost.

That said, you are right in pointing out that version 1 of Precision Boost was rudimentary, while version 2 is much more dynamic.

 

[Hitman928]

I don't think this is accurate. As you can see in the 1800X slide included my last post, the algorithm was already coined "Precision Boost" for the 1000 series. XFR has always just been a name for the difference between the Max Boost number printed on the box and the absolute max frequency the processor could boost to under ideal conditions.

AMD Senior Technical Marketing Manager Robert Hallock, as I understand it, clearly describes XFR as a marketing term for the ability of the Precision Boost algorithm to gain extra boost.

That said, you are right in pointing out that version 1 of Precision Boost was rudimentary, while version 2 is much more dynamic.

Yes, you are correct, they both existed with Ryzen 1, I was mis-remembering that part. The rest of my post still stands though. The original XFR and how it worked with PB was not a "smart algorithm", it was pretty limited and rigid in it's scaling compared to Zen+ and Zen2. With Zen2, XFR has basically just been swallowed by PB so what you see with the latest chips in boosting already takes advantage of what XFR was meant to take advantage of in the first place.

 

[beginner99]

With that said, does it really matter? I mean, what are people doing where they are really only using 1 core with no background processes using cpu time. In today's age, I would worry more about the 2 core boost frequency and beyond. If the 2 core boost level fails in a very lightly threaded load environment, I'd be more worried.

By itself it's not a huge deal but see more below. Also AMD themselves state they do not have a single-core boost. according to them 4.6 could be reached on all cores given the right circumstances (temps, voltages,...)

Additional 100Mhz to go from 4.5Ghz to 4.6Ghz amounts to 2.2% performance boost. While I think AMD may have stretched technical specification to the limits to arrive to 4.6GHz single core boost, I find it difficult to care about losing on 2.2% in single thread applications.

This is true but I don't like it due to the added confusion with PBO were one of AMDs own said you could increase the boost limit by 200mhz which obviously is not the case. That was misleading. The chips are at their limits. I wouldn't care if I could enable PBO and then get 4.75 ghz or what or with manual OC. You can't even reach 4.6ghz on a manual OC which is telling how close to the limit these chips are operating. It's probably also some disappointment in there. If the sticker says 4.6ghz, I would expect it's possible to reach such an all-core manual OC with water cooling. But not even that is possible. As I sad previously we need to adjust mentally as AMDs "ghz" number" has to be interpreted differently as intels current one.


On a different note:
I wonder about the chips longevity as reaching 4.5-4.6 ghz seems to require enormous voltage. What if they degrade and only reach 4.4 ghz 3-4 years down the road?

 

[Markfw]

By itself it's not a huge deal but see more below. Also AMD themselves state they do not have a single-core boost. according to them 4.6 could be reached on all cores given the right circumstances (temps, voltages,...)



This is true but I don't like it due to the added confusion with PBO were one of AMDs own said you could increase the boost limit by 200mhz which obviously is not the case. That was misleading. The chips are at their limits. I wouldn't care if I could enable PBO and then get 4.75 ghz or what or with manual OC. You can't even reach 4.6ghz on a manual OC which is telling how close to the limit these chips are operating. It's probably also some disappointment in there. If the sticker says 4.6ghz, I would expect it's possible to reach such an all-core manual OC with water cooling. But not even that is possible. As I sad previously we need to adjust mentally as AMDs "ghz" number" has to be interpreted differently as intels current one.


On a different note:
I wonder about the chips longevity as reaching 4.5-4.6 ghz seems to require enormous voltage. What if they degrade and only reach 4.4 ghz 3-4 years down the road?

Long term ? I have 12 year old chips that are still running. Unless you overclock them using extreme voltage, they will be fine.

Can you be more picky ?