AMD Updates Ryzen Product Pages to Elaborate on "Max Boost Clocks"

[cucker tarlson]

Zen2 does overclock like a turd but the IPC improvement from Zen/Zen+ neutralizes the clockspeed difference going from Intel's high clocked Skylake core design unless you move to the Sunny Cove cores re-design. I have a feeling Sunny Cove isn't expected to clock as high on 10nm+ as Skylake on 14+++. Zen2 clocks the same as Sunny Cove.

but isn't sunny cove gonna do the same what you just said ryzen 3000 did ?

 

[trog100]

when you have a chip with lots of cores/threads and heat being the ultimate limiter which it is you have two options.. run the chip at slower all cores working speeds or have a sliding clock scale which relates to how many threads are actually being used..

as for poor overlooks this is what happens when manufacturers are forced to run their chips at close to their real abilities as opposed to market segment speeds.. its as it should be..

trog

 

[TheLostSwede]

But the CB R20 scores and performance across some popular benchmarks is on point with the major reviews out there. I was checking the clock speeds in the newest Ryzen Master, they did not go higher then 4,05Ghz all core so its a mystery for me. This is pointing to a measurement error of the software.

Ryzen Master doesn't seem to be that accurate tbh. I see higher clocks in hwinfo, but not sure if that's accurate either...

when you have a chip with lots of cores/thread and heat being the ultimate limiter which it is you have two options.. run the chip at slower all cores working speeds or have a sliding scale clock scale which relates to how many threads are actually being used..

as for poor overlooks this is what happens when manufacturers are forced to run their chips at close to their real abilities as opposed to market segment speeds.. its as it should be..

trog

The problem is that the sliding scale is borked. Playing games, I never see as high core speeds as I do when running benchmarks and we're not talking BF V here. Fire up something like Apex Legends or Overwatch and you'll have a couple of cores boosting high and the rest running around 1-2.5GHz...
Sure, it might not be a well threaded game or even need the extra cores, but the boosting cores hit 4.2-4.3GHz only.
As such, something isn't right.

 

[biffzinker]

but isn't sunny cove gonna do the same what you just said ryzen 3000 did ?

Yes, I'm looking forward to the outcome.

 

[Vayra86]

The problem is that the sliding scale is borked. Playing games, I never see as high core speeds as I do when running benchmarks and we're not talking BF V here. Fire up something like Apex Legends or Overwatch and you'll have a couple of cores boosting high and the rest running around 1-2.5GHz...
Sure, it might not be a well threaded game or even need the extra cores, but the boosting cores hit 4.2-4.3GHz only.
As such, something isn't right.

Meh, I beg to differ. Its really just a problem with the definition of 'turbo' or 'boost'.

In fact, I think we can safely say only Nvidia has its GPU boost story in good order. As in: you always get an advertised clock, and most of the time, you get much higher clocks. Even if you nearly cook the GPU it will still run base clock or you'll be using it completely out of spec.

Perhaps what we need is an industry standard wrt boost clocks. I could imagine you'd set that at something like a total achieved frequency across all available threads divided by number of threads. And when you do it like that, suddenly AMD doesn't look all that bad. Another approach could be the total deviation from advertised clocks - again, if you'd put Intel versus AMD spec sheet and reality; AMD will be coming out much better.

Let's not go blind on that peak clock number and define something around thát. I really don't like Intel's Turbo and the spec sheet trickery they've deployed over the years (and how recent parts handily turn the 'headroom' into 'used room' and how CPUs royally boost beyond TDP on stock bioses). They're playing the game for marketing, AMD is just bad at it (once again... it never ends does it) and you're right they've left a gap here for a lawsuit. But AMD does deploy a much better type of boost.

I mean for this example here, what should AMD advertise, 4375 mhz on the box instead of 4.4? 25 mhz...

As for the 'need' to see what conditions are required to get the advertised boost clocks...maybe. Maybe not... isn't the actual defining factor for a CPU in the end performance? You cannot grasp CPU performance based on clocks alone, nor could you do it based on the advertised turbo/boost clock of a spec sheet, after all there are barely any use cases that only hit one core.

 

[bug]

The problem is that the sliding scale is borked. Playing games, I never see as high core speeds as I do when running benchmarks and we're not talking BF V here. Fire up something like Apex Legends or Overwatch and you'll have a couple of cores boosting high and the rest running around 1-2.5GHz...
Sure, it might not be a well threaded game or even need the extra cores, but the boosting cores hit 4.2-4.3GHz only.
As such, something isn't right.

4.2-4.3 is not far off 4.5. You're only supposed to get 4.5 if you have load on a single core, while the others are literally shut off.

 

[londiste]

No, we do not need to redefine turbo/boost, it has been a stable thing for years - boost clock is what a CPU will run at with single core load.

 

[Vayra86]

No, we do not need to redefine turbo/boost, it has been a stable thing for years - boost clock is what a CPU will run at with single core load.

Nuh-uh! With a single core load 'while remaining within stated power and temperature specs' - all of which are subject to change between CPUs and generations, nodes etc.

Its already shady as hell. And the only reason it was 'stable for years' is because Intel was the only one playing the game.

 

[londiste]

Nuh-uh! With a single core load 'while remaining within stated power and temperature specs' - all of which are subject to change between CPUs and generations, nodes etc.

Power and temperature specs are something I would never want a CPU to reach. Why would that be a problem?

Edit:
Definitions, specifications and limits are not the problem. Adherence to these is. For example:
- I can see from specs that my i5 8400 has a 100C temperature limit and power limit 65W - with short-term (8 seconds) extended 81W limit on top of that - where both work as expected.
- Similarly, my 3600X has a 95C temperature limit that works as expected but a 95W power limit that never seems to be reached, nevermind the extended 128W limit. At the same time, CPU is a bit short of its spec boost clock.

 

[Vayra86]

Power and temperature specs are something I would never want a CPU to reach. Why would that be a problem?

The problem lies in the fact that a much better boost mechanic is under pressure by the current definition, which makes it detrimental to innovation. We are looking at a topic where a 25 mhz deviation from advertised boost clock (on a 4400 mhz CPU...) is reason to believe this may turn into class action. Meanwhile there is a competitor that has safely lodged its business model within the current definition and is actually showing much greater deviation from 'advertised' turbo if you're looking beyond 1 core.

A 1-core turbo is also not very relevant, its a remnant from the days when a quad core was considered high end and many applications were single threaded. The number means effectively nothing today.

 

[londiste]

The max I have seen with load is 4300 (with a split second of 4319) for a CPU where spec says 4400. I have seen 4400 reported only at idle, again for a split second. Apparently, this is even not the worst result.
I fully expect AMD to figure out the boost in BIOS at one point but a month after release seems like it would be an appropriate time already.

Meanwhile there is a competitor that has safely lodged its business model within the current definition and is actually showing much greater deviation from 'advertised' turbo if you're looking beyond 1 core.

A 1-core turbo is also not very relevant, its a remnant from the days when a quad core was considered high end and many applications were single threaded. The number means effectively nothing today.

Remember all the flak Intel got for not publishing static frequency tables?

 

[Vayra86]

The max I have seen with load is 4300 (with a split second of 4319) for a CPU where spec says 4400. I have seen 4400 reported only at idle, again for a split second. Apparently, this is even not the worst result.
I fully expect AMD to figure out the boost in BIOS at one point but a month after release seems like it would be an appropriate time already.

Alright, show me the examples of Intel CPUs running 4.7 or whatever is on the sheet under anything other than idle (which is when they drop to idle clocks...). I've never managed to get them at stock in any use case, even just browsing already loads more than 1 core.

Another problem by the way is already apparent if you look at how TDP is defined and how it is used in practice.

Remember all the flak Intel got for not publishing static frequency tables?

Right! Well then... time to redefine things no? Get with the times. We're looking at 16 thread consumer CPUs nowadays.

 

[Vya Domus]

So this has pretty much turned into GPU-like dynamic clocks, not much of a surprise here seen it coming for years.

 

[londiste]

Alright, show me the examples of Intel CPUs running 4.7 or whatever is on the sheet under anything other than idle (which is when they drop to idle clocks...). I've never managed to get them at stock in any use case, even just browsing already loads more than 1 core.

Another problem by the way is already apparent if you look at how TDP is defined and how it is used in practice.

Which Intel CPUs? I cannot see a single one with 4.7 in the spec. Care to provide a link?
Yeah, Intel fucked up with TDP for K models... and AMD decided to follow suit.

 

[Vayra86]

Which Intel CPUs? I cannot see a single one with 4.7 in the spec. Care to provide a link?
Yeah, Intel fucked up with TDP for K models... and AMD decided to follow suit.

8700K, 9900K, even 7700K, the list is long. At stock none of them frequently run at peak turbo. There are always more threads at work, which kicks you down immediately. My 8700K would be doing 4.3 99% of the time lacking an OC. So in practice if you'd put it side by side with a stock Ryzen 3rd gen, Ryzen would still clock higher most of the time.

And Intel didn't fuck up at all, this is working as intended, make no mistake. Long term strategy for 'sustainable growth' is what they call this in business. There are no mistakes, there is only carefully screwing you over every time and getting away with it.

 

[londiste]

8700K, 9900K, even 7700K, the list is long. At stock none of them frequently run at peak turbo. There are always more threads at work, which kicks you down immediately. My 8700K would be doing 4.3 99% of the time lacking an OC.

Come on, you should know better than this.

Spec boost clock is basically for a single thread. Spec base clock is for all-core load. And all of the listed CPUs will reach that at stock.
8700K base 3.7 and boost 4.7
7700K base 4.2 and boost 4.5
9900K base 3.6 and boost 5.0

Yes, there are cases where this is not reached due to power limit, mostly due to full AVX2 load although at stock settings with stupid TDP configuration they will do even that just fine.
Same applies to Ryzen 1000 and 2000 processors. They can even boost a little above specced clocks, consistently.

Now with a Ryzen 3000 CPU there is a pretty good chance you will not reach spec boost clock in any circumstances.

 

[Vayra86]

Come on, you should know better than this.

Enlighten me.

 

[londiste]

You seem to be trying to redefine boost clock as base clock for no obvious reason.

And Intel didn't fuck up at all, this is working as intended, make no mistake. Long term strategy for 'sustainable growth' is what they call this in business. There are no mistakes, there is only carefully screwing you over every time and getting away with it.

Talking about customer POV here. I don't like the antics around TDP one bit.

 

[Vayra86]

Come on, you should know better than this.

Spec boost clock is basically for a single thread. Spec base clock is for all-core load. And all of the listed CPUs will reach that at stock.
8700K base 3.7 and boost 4.7
7700K base 4.2 and boost 4.5
9900K base 3.6 and boost 5.0

Yes, there are cases where this is not reached due to power limit, mostly due to full AVX2 load although at stock settings with stupid TDP configuration they will do even that just fine.
Same applies to Ryzen 1000 and 2000 processors. They can even boost a little above specced clocks, consistently.

Now with a Ryzen 3000 CPU there is a pretty good chance you will not reach spec boost clock in any circumstances.

Understood. But then we're not talking about the same thing exactly. Its clear that AMD's XFR is not as rigid in setting its turbo clock and much more fine-grained, and in doing so, it achieves better results than Intel's pretty rigid turbo clock based on # threads loaded.

Whether or not this is up to further tweaking / refinement of AMD's BIOS or AGESA remains to be seen. Ryzen 3 is a clear departure in terms of clocks, reaching quite a bit higher than before. And maintaining that across multi-threaded loads in a very good way. It would be a shame if the lacking definition of 'Turbo' or 'Boost' would stand in the way of that development.

But... that's what I've been repeating for a page now.

And no, we're not redefining boost as base. I'm saying, refine the definition to make it a more useful measure of performance. The intent is the same. Applying turbo as what it can be for a single thread is ridiculous, when nobody ever gets to a pure single threaded scenario anymore. Its a bit like stating top speed for cars, when in reality you know you'll never hit that in any condition other than a race track with a very long straight.

Let's put this very example through a hypothetical 'class action' and the end result of that: AMD hastily tweaks XFR so that 1 core boost finally hits 4.4 consistently, and the result is that higher thread count loads all lose 100-200mhz in return. Hurray?

 

[Vya Domus]

Agreed, XFR does a much better job at targeting the optimum frequency/power curves than the rather dumb Intel boost clocks that buckle under AVX loads and whatnot.

 

[londiste]

Base and boost are useful to define a range of frequencies CPU can reach under load and this is exactly what these have been used for. The actual frequencies reached between these has several factors, amount of loaded cores, power and temperature being the primary ones.

There would be no problem if AMD had put boost frequency 200MHz lower in specs and said it can reach higher than that. Which is technically what they did but with PBO but at frequencies that Ryzen 3000 CPUs are not always able to reach.

Actually, now that I think of this, could Ryzen 3000 boost spec be with PBO?

Agreed, XFR does a much better job at targeting the optimum frequency/power curves than the rather dumb Intel boost clock that buckle under AVX loads and whatnot.

Ryzen 3000 CPUs do get a frequency hit under AVX2 load, it is simply considerably smaller. I see 100-200MHz less frequency in CB20 or Prime95 with AVX2 compared to CB15 or Prime95 without AVX2. Whether the size of hit is due to architecture or smaller process node, not sure.

 

[Vayra86]

Base and boost are useful to define a range of frequencies CPU can reach under load and this is exactly what these have been used for. The actual frequencies reached between these has several factors, amount of loaded cores, power and temperature being the primary ones.

There would be no problem if AMD had put boost frequency 200MHz lower in specs and said it can reach higher than that. Which is technically what they did but with PBO but at frequencies that Ryzen 3000 CPUs are not always able to reach.

Actually, now that I think of this, could Ryzen 3000 boost spec be with PBO?

Agreed, they could state boost clocks are for example 4.3 Ghz.

And you know what'd happen next? We'd still complain they can't clock high enough to buy them, after all what changed compared to previous gen then?! When in fact they *nearly* clock to 4.4, like literally a hair away from it.

It would hurt sales, and in turn, it would push AMD to tweak XFR differently, and possibly, less favorably for end users. Perhaps they'd go the Intel Turbo route instead? Much easier and then they could even advertise with 4.5 or better...

So that is why I feel re(de)fining it is in order.

 

[londiste]

Its clear that AMD's XFR is not as rigid in setting its turbo clock and much more fine-grained, and in doing so, it achieves better results than Intel's pretty rigid turbo clock based on # threads loaded.

AMD and Intel boost are not that different in functionality. They are different in details - how fine-grained the frequency is, how fast they are and how wide is the range of frequencies eventually reached. AMD works with 25MHz increments while Intel works with 100MHz. AMD works in few ms, Intel works at a magnitude higher ms. Frequencies reached is basically the range between Base and Boost and the width of that is dictated by set limits.

AMD's acronyms and technologies are a bit of a mess. XFR/XFR2 (eXtended Frequency Range) is not the normal boost and is the part of the boost algorithm that boosts beyond boost clock given no limits are reached. Kind of similar in function to Intel's Boost 3.0 (which was an abject failure), but different because XFR works. PBO is to increase power limit on top of that.

Let's put this very example through a hypothetical 'class action' and the end result of that: AMD hastily tweaks XFR so that 1 core boost finally hits 4.4 consistently, and the result is that higher thread count loads all lose 100-200mhz in return. Hurray?

I do not see any reason why would getting one-core boost faster decrease all-core boost.

It would hurt sales, and in turn, it would push AMD to tweak XFR differently, and possibly, less favorably for end users. Perhaps they'd go the Intel Turbo route instead? Much easier and then they could even advertise with 4.5 or better...

What exactly do you think is different in Intel Turbo route?
You just said that AMD method is better and it is. AMD is going way beyond what Intel is doing with boost and successfully so.

 

[Vayra86]

I do not see any reason why would getting one-core boost faster decrease all-core boost.

What exactly do you think is different in Intel Turbo route?

My assumption (! which you are making me rethink right now) is that Intel has focused on getting that high turbo clock number because it sells, when in practice it really doesn't do much. Maybe I'm wrong in the idea that the two cannot exist together (high top end clock and optimal efficiency across all threads), because that IS based on Intel's turbo principle.

You correctly note that AMD is following the Intel route when it comes to headroom and TDP, and maybe I'm wrong and AMD simply has work to do.

 

[londiste]

My assumption (! which you are making me rethink right now) is that Intel has focused on getting that high turbo clock number because it sells, when in practice it really doesn't do much. Maybe I'm wrong in the idea that the two cannot exist together (high top end clock and optimal efficiency across all threads), because that IS based on Intel's turbo principle.

You correctly note that AMD is following the Intel route when it comes to headroom and TDP, and maybe I'm wrong and AMD simply has work to do.

Both are clearly trying to write as high as possible boost clock to the spec

Boost and base clock are not really tied to each other but are more of a result of how architecture and manufacturing process for a particular processor end up working in given limits - today this is primarily power limit.

One thing AMD does with Ryzen 3000 that is new and awesome is single-core boosts at high voltage. Ryzen 3000 CPUs dump a lot of voltage (up to around 1.5V) into single core to get to these 4.x GHz boost clocks. While the comments about misleading idle figures are accurate, these are actually correct load numbers. Monitoring tools say so, reviewers with external voltage monitors are reporting results in that range and indirectly power consumption figures (from both monitoring software as well as external monitoring) also support it. These frequencies are very clearly on an unfavorable upwards slope on the efficiency curve, further than Intel has so far been willing to go.