Discussion CPU boost frequency and marketing

[Vattila]

Although Ryzen 3000 reviews, in general, have been favourable, there has been some disgruntlement — or puzzlement, at least — about the advertised boost frequency. Reviewers, such as derBauer (video), has pointed out that the maximum boost frequency is rarely, if ever, reached. He makes detailed tests showing the frequency behaviour of the tested chips during various workloads, and depicts the frequency as a curve over time. Sometimes he sees a few spikes hitting the stated max boost frequency, or very close to it, but in his commentary, he discounts those spikes since they happen before the test starts. His commentary makes the viewer expect that the CPU should reach boost frequency more often, i.e. during the workloads tested.

Is this fair understanding of maximum boost frequency? Should the maximum boost be a measure of core frequency under any prolonged workload, even if single-threaded? My understanding of maximum boost is that it is the frequency you should only expect of cold silicon, sitting there idle and then boosting up as soon as it is given something to do, e.g. as the user clicks a browser tab, after having left the previous tab idle for seconds while reading, and with nothing much happening in the background. Even after just milliseconds of work, a core's temperature rises, meaning it is no longer able to hold that maximum boost.

Is frequency at all important? Most enthusiasts know that it is not comparable across CPU brands nor generations, and is only useful as a relative measure of performance between SKUs in the same series. However, I think it still is to some degree incorrectly used as a performance measure between brands — knowingly or unknowingly. For that reason, it has significant marketing value. For example, the 5 GHz milestone means much to both vendors and customers, it seems, even though it in itself has no special significance.

If we are to have meaningful frequency specifications for CPUs, do we need better metrics? Interestingly, we recently saw AMD introduce the concept of "Game Clock" for the Radeon RX 5700 series. Do we need something similar for CPUs? For example, consider Typical Single-Thread Boost (ST Boost) and Typical Multi-Thread Boost (MT Boost), measured for typical workloads such as Cinebench Single-Threaded and Multi-Threaded. Then we would have 4 frequencies on the box, which for Ryzen 3900X — according to derBauer's testing — would be:

• 4.6 GHz Max Boost (rare momentary spikes from cold and idle conditions)
• 4.5 GHz ST Boost (under Cinebench Single-Threaded workload)
• 4.1 GHz MT Boost (under Cinebench Multi-Threaded workload)
• 3.8 GHz Base (worst-case MT workload in hot and power-constrained conditions)

PS. The discussion about the built-in overclocking feature, Precision Boost Overdrive, is another matter. Here the AMD marketing probably invited to disappointment (referring to potential +200 MHz overclock), since as it turns out the Ryzen SKUs do not seem to have much overhead left beyond the warranted specifications. This is in some sense very good, as derBauer and others have pointed out. It means the customer gets more or less the full potential performance out of the box. However, it leaves overclockers disappointed, since they expected to be able to extract more.

 

[IEC]

The AMD video about PBO with Robert Hallock showing PBO frequencies should definitely be called out for being misleading.

Most egregious example from that video reproduced below:


This video gave the false expectation to many that clocks above 4.5GHz would be feasible and readily achievable, when in practice you'll never see that high a frequency on any AVX loads or MT tasks (barring golden sample manual OC like some lucky redditors). Only on the lightest ST workloads do you see transient 4.5GHz+ boosts. The highest clock I've ever recorded was 4591 at idle, so theoretically I'm hitting my advertised boost (-9MHz), but I never see it in real workloads at stock.

Will that change with 3950X and/or a maturing 7nm process? Probably not by much. Unless they make changes to the design to allow for higher fMax I doubt 4.5GHz all-core OCs (esp AVX loads) will be anywhere near common anytime soon.

That said, my optimized per-chiplet OC 3900X almost doubles the MT performance of my max OC 8700K at approximately the same power usage (7316 CB20 MT vs 3706 CB20 MT). It tears through MT workflows and puts even my Threadrippers to shame. Impressive chip despite the rough edges.

 

[nicalandia]

my optimized per-chiplet OC 3900X almost doubles the MT performance of my max OC 8700K at approximately the same power usage (7316 CB20 MT vs 3706 CB20 MT).

What was the score of "un-optimized" with your settings? I've read that stock its about 7,000 so an additional 5% in performance increasement is like an Intel Tick.... And a huge 25% increasement from Zen+ 12 core 2920X and 38% over 12 core 1920X... Impressive

 

[Zucker2k]

Should the maximum boost be a measure of core frequency under any prolonged workload, even if single-threaded? My understanding of maximum boost is that it is the frequency you should only expect of cold silicon, sitting there idle and then boosting up as soon as it is given something to do, e.g. as the user clicks a browser tab, after having left the previous tab idle for seconds while reading, and with nothing much happening in the background. Even after just milliseconds of work, a core's temperature rises, meaning it is no longer able to hold that maximum boost.

So you mean to say that the single core frequency boost displayed on the box is not meant or guaranteed for actual work? How does Intel do it then? Do we have to redefine the meaning of "boost" because AMD is struggling to hit advertised boost clocks in this iteration of the zen architecture? We had no such struggles with the definition with the zen+ iteration. The OP is a bit of a stretch.

In my opinion, AMD is fighting a hard frequency limit brought on by a combination of power and temps. If you look at the power to performance curve of TSMC's 7nm process, the 8 core chips, for example, should ideally be operating around 65w - 75w. However, AMD has pushed them all the way to the fmax limit with additional power to the extent the silicon needs ultra high end cooling to breach advertised clocks on all cores - something that was much more easily attainable on Ryzen 2000..

Ryzen 3000, is a beast in its own right, unfortunately, AMD had another beast in the 9900k to contend with. In their quest to unseat the 9900k, they outdid themselves with the factory overclocking/boosting algorithms. You can imagine how the reception would have gone had AMD's full process-shrinked chips lost across the board to a skylake derivative on equal cores/clocks (they trade blows). Without pushing the limits on clocks, it'll have been a huge embarrassment for them, being a full node up and all. Isn't it somewhat counter-intuitive to go down a full node and still maintain the same power budgets? Where are the 45w hexacores and octocores?

 

[Vattila]

This video gave the false expectation to many that clocks above 4.5GHz would be feasible and readily achievable

Yeah. AMD Senior Technical Marketing Manager Robert Hallock — which seems to me, in all his presentations and demeanour, to be a genuine PC-enthusiast trying to do clear and fair marketing and engage with the community — is now under criticism for his promotion and claims about Precision Boost Overdrive (which currently seems to do very little, if anything at all), and in particular the reference to possible gains of another 200 MHz over and beyond the 4.6 GHz frequency advertised for Ryzen 3900X (see his PBO explanation video).

As more reviews, experience and evidence emerges for the Ryzen 3000 boosting behaviour, it seems to corroborate my intuition that AMD did not hit the frequency targets for Zen 2. I think Lisa Su hinted about it when she announced the launch schedule earlier this year ("we need to get frequencies where we want them", or something to that effect). That schedule was a bit later than most expected, me included.

Also, the rough edges seen in AGESA versions, PBO ineffectiveness and odd boost behaviour, in particular, inconsistent idle behaviour in Windows showing high power usage (due to aggressive boost behaviour, it seems, causing "observer's syndrome" in monitoring tools), all point to AMD having struggled with frequency and optimisation of boost behaviour before launch.

Maybe the AMD engineers put too much priority on power-efficiency and targeted a sweet spot a little too low on the power-vs-frequency curve, hoping and expecting they would reach the frequencies necessary on the high-end to challenge Intel single-thread performance dominance. However, it now seems the SKUs have trouble even reaching advertised boost frequencies — to any meaningful degree, at least.

Hopefully, Zen 2 is doing better than expected on the power-efficiency part of the curve to compensate. In the end, Ryzen 3000 is still impressively competitive (especially as 10nm Ice Lake-S is a no-show), and Zen 2's service in EPYC 2 is after all the top priority, where power-efficiency at the sweet spot on the frequency curve is more important.

Still, I hope AMD and TSMC have more they can do to refine the Zen 2 implementation and the 7nm process to gain some frequency. If so, perhaps we will have a refresh of Ryzen 3000 before Zen 3 arrives. It will be interesting to see how Ryzen 3950 fares when it launches in September, and whether the silicon is any better.

 

[Vattila]

So you mean to say that the single core frequency boost displayed on the box is not meant or guaranteed for actual work? How does Intel do it then? Do we have to redefine the meaning of "boost" because AMD is struggling to hit advertised boost clocks in this iteration of the zen architecture?

Momentary boosts in frequency do have value, even if they are never seen in prolonged workloads, not even single-threaded workloads. This boost behaviour from cold and idle conditions improve responsiveness and energy consumption (race-to-idle, i.e. complete the task at max frequency and then return to idle and cool conditions as soon as possible).

But I agree with you, that it is unfortunate marketing, if customer expectations are not met, e.g. due to established industry assumptions. As you point out, PC enthusiasts have expectations set by experience with other brands and processor architectures. To address this I suggested a more comprehensive marketing scheme for boost frequencies in my OP. I do think such an approach would be more meaningful to the customer.

What would be nonsense is to cap the maximum boost frequency down to the Typical Single-Thread Boost frequency. And not capping it, but just refraining from specifying it on the box, even if rare, is obviously a missed marketing opportunity, knowing how much value customers put on frequency, even if misguided.

 

[IntelUser2000]

So you mean to say that the single core frequency boost displayed on the box is not meant or guaranteed for actual work? How does Intel do it then?

Although Ryzen 3000 reviews, in general, have been favourable, there has been some disgruntlement — or puzzlement, at least — about the advertised boost frequency.

This is just an unfortunate consequence of clocks going beyond the point of sanity.

Even Intel should have had trouble beyond 4.5GHz had 14nm retired back in 2016. The thing is though it did not, so they had the same uarch and process for 3 additional years to refine.

You remember the "5GHz overclock" claims for Intel chips. 4790K claimed that. 6700K claimed that, and they did it again with 7700K. Outside of labs and super golden samples that was not possible. Only with Coffeelake they finally reached that limit. I don't think a hypothetical Icelake-S clocking less than 14nm CFL has to do much with 10nm transistor performing less as much as Intel selling you factory overclocked chips starting with Kabylake.

I actually find it impressive that AMD managed to reach that high in single generation and the PBO can clock just as high as manual overclocking is able to.

In the end though, they are just eating up overclocking headroom and selling them as "stock".

 

[lopri]

In some stability testing I saw a 3700X not even maintaining its base frequency (3.6 GHz).

 

[IntelUser2000]

In some stability testing I saw a 3700X not even maintaining its base frequency (3.6 GHz).

Maybe that's due to thermals. Tell us the link?

 

[lopri]

Of course thermals played a role. But I was not running it in a sauna. Unfortunately I do not have screenshots and cannot reproduce it because my 3700X is dead.

Here is another member's (GhostPirateLeChuck) 3900X not hitting the base clock while running a BOINC project.

https://forums.anandtech.com/thread...at-density-of-tsmc-7nm.2567790/#post-39877318

 

[Kenmitch]

Of course thermals played a role. But I was not running it in a sauna. Unfortunately I do not have screenshots and cannot reproduce it because my 3700X is dead.

Did you kill it or was it already a dead cpu crunching? < play on dead man walking.

Are you still within the return window? Unless you did something crazy and know you killed it an exchange isn't unreasonable.

 

[Kenmitch]

As for as the topic goes....My 3600 speaks for itself.

The data in the graph was grabbed just before the single core test ended. My 3600 is running at stock settings with PBO turned off.

I initiated the single core test and Core #2 instantly locked on. It boosted to the advertised maximum and maintained it all the way to the end of the test.



I'm perfectly happy and satisfied with my 3600 purchase.

 

[IntelUser2000]

https://forums.anandtech.com/thread...at-density-of-tsmc-7nm.2567790/#post-39877318

That's a beefy HSF. Man, 66C with that heatsink and fan.

We're really pushing the limits.

 

[tamz_msc]

This is what AMD claims in its official reviewers' guide:

The fact is that 4.6 GHz, for those who actually get it on one thread, happens only under some extremely specific circumstances, like running an infinite loop that does nothing, or sometimes while idling in Windows, and is sustained only for a momentary duration of time. It rarely happens under any task that would constitute a real workload, like CPUZ benchmark or Cinebench 1T, which are considered relatively light workloads.

The fact is that nowhere is it stated that this behavior is intentional. If we take the reviewers' guide to be the true behavior of the boost algorithm, then it is clear that such behavior isn't borne out of the tests people are conducting in the real world. This is why it is such a big deal.

Also, people are habituated to how Turbo Boost works on Intel. As long as the load isn't AVX, a 9900K will happily boost to 5 GHz on one thread unless thermally constrained. AMD doesn't differentiate their boost algorithm well enough, which leads to people believing that the boost frequency printed on the box is a simple 1T boost just like it is on Intel.

 

[IntelUser2000]

It's not that big of a deal. If it can't do 4.6GHz most of the time, it can do 4.5GHz.

Windows jumps threads around and even on Intel chips its easy to knock it off the 1T frequency.

Also, people are habituated to how Turbo Boost works on Intel. As long as the load isn't AVX, a 9900K will happily boost to 5 GHz on one thread unless thermally constrained.

That's cause the 9900K can boost to 5GHz with 2 cores as well. So not subject to the Windows bouncing.

 

[Atari2600]

Here the AMD marketing probably invited to disappointment (referring to potential +200 MHz overclock), since as it turns out the Ryzen SKUs do not seem to have much overhead left beyond the warranted specifications. This is in some sense very good, as derBauer and others have pointed out. It means the customer gets more or less the full potential performance out of the box. However, it leaves overclockers disappointed, since they expected to be able to extract more.

So 99.999% of the market get a better CPU than a version without the algorithms, while 0.001% of the market are unhappy that they got exactly what they paid for.

 

[beginner99]

I'm not happy with the situsation and hence mentally need to adjust. AMD clocks on the box are simply 200-300 mhz higher compared to intel. to compare, subtract them. When AMD said "boost up to 4.6" I would expect all-core OC to be relatively easily reach that, just like with intel. But it's not true. 4.4 is lucky and 4.3 probaly the norm.

In the end it turned out exactly as expected with intel clearly keeping ST lead due to clocks.

 

[Head1985]

I am also not happy.My 3700x never boost at 4.4ghz in games.Not even at 4.3ghz.Most of the time it sits at 4200-4250mhz Also temperatures are very high for 65w TDP cpu at only 4.2ghz.With mugen5 revb i am in games at 60-70°C.Other people reporting even 70-80°C.
PBO does nothing in games.I think cpu boost even lower in games with PBO on.

 

[Paul98]

I don't care what frequency the CPU runs at as long as the reviews show the real world performance.

Also I expect there may be some updates to this stuff moving forward. I don't think there would be much of an issue having a single thread running 50Mhz faster.

 

[Kenmitch]

I decided to make another single core run in CB R20 to see how my 3600 responds. I guess I wanted to see if the last run was a fluke.

It's following the same pattern as the last run as core #2 locked on and boosted to 4200MHz again and is holding steady.

Looks like core #2 is my favored core when it comes to single threaded loads....At least heavier ones.

 

[deathBOB]

Its disingenuous because it deviates from past experiences with boost. Before this chip generation boost was understood to mean that these are the clocks you will hit if you provide adequate (not extreme) cooling. Sure the fine print says "no guarantees" but consumers have come to expect boost to effectively reach target clocks in normal working conditions.

 

[Kenmitch]

So 99.999% of the market get a better CPU than a version without the algorithms, while 0.001% of the market are unhappy that they got exactly what they paid for.

My 3600 currently runs better when I just let it run as designed.

PBO might be worth looking into once uefi matures.

 

[Markfw]

So even though these are an insane value for what you get ($200 for a hex core that many are happy with) and they do exactly what they were designed to do, becuase it won't do the max boost when some want it to, they are unhappy.

Go buy Intel then or stop wining. geeezzzz

 

[tamz_msc]

So even though these are an insane value for what you get ($200 for a hex core that many are happy with) and they do exactly what they were designed to do, becuase it won't do the max boost when some want it to, they are unhappy.

Go buy Intel then or stop wining. geeezzzz

The 3900X is supposed to do 4.6 GHz on one thread, according to AMD, yet it doesn't do that for many people. That means that they're not doing what they're designed to do.

 

[Markfw]

The 3900X is supposed to do 4.6 GHz on one thread, according to AMD, yet it doesn't do that for many people. That means that they're not doing what they're designed to do.

It DOES do it, but by your own words " yet it doesn't do that for many people". That means it works, but if it does not do it for everyone, that does not mean its broken or defective.

As I said, if you don;t like the 3900x, go buy Intel.