XFR Extended Frequency Range; Incomplete WIP?

There seems to be contradictory information from AMD. There is a chart that says XFR is for X binned chips, but shows the 1700 also getting some XFR Boost.
The fact that R5 chips are getting more XFR Boosts shows that this limit is adjustable in the code.
It also suggests there is a way that XFR could work on top of a manual overclock.

1 Like

I see no reason why XFR wouldn't be able to run on overclocked systems per se, but there are indubitably operational constraints. One of the major constraints in my opinion is how the differently spec'ed SKU's are implemented in practice.

AMD allows overclocking on all SKU's. They allow OEM's to explore the limits of their CPU's and GPU's. However, they always provide a safety limit of some sorts, which is fair enough. On the FX8350, they issues SKU limitations for some motherboards, on account of the power supply of those mobos not being beefy enough to deal with the rampant power needs of an overclocked FX8350, even of a stock FX8350 on mobos with a basic analogue VRM with a definite Ampère limitation. That is just a sound practice in my opinion.

I suspect that they want to limit high frequency scheduling above the safe operational spec of non-X SKU's, because a lot of these will be implemented in mobos that are just not up to the task. There are several inherent technical problems when doing high frequency scheduling and switching. It has gotten to a point where you need the chokes to not only regulate the power delivery for the CPU, but also to be exactly on spec to avoid unwanted side effects and at the same time keep the operational conditions within very fine boundaries. This costs money in R&D (and premium parts, but the impact on the BOM of a mobo just because of the premium chokes and FET's is rather limited given the scale).

I'm glad to see that AMD basically has made Ryzen so thoroughly well engineered that these CPU's are able, from day one, with still pretty much experimental UEFI of first gen mobos, to run at their spec full-time within the specified power envelope. Do not forget that these are 95W TDP SKU's, not 140W TDP ones. That is a difference of almost 50%!!! It is a far cry from what the FX8350 was capable of, and from what the competition is capable of. An FX8350 was drawing more than 180-200 W in the blink of an eye when overclocked. It needed both beefy VRM's and beefy cooling solutions, and FX8350's would die frequently in some of the most popular applications for that chip, namely rendering farm rigs. Pretty much everyone I know who overclocked his FX8350, has had a big scare while exploring how fast it would go at some time or other. So you can't push the envelope with Ryzen as much. Big deal, the performance for the power envelope is amazing! Imagine how crazy the engineering is to get it to work that consistently at spec in that power envelope! Can 2011-3 SKU's do that... they can run consistent, but in a power envelope that's at least 50% larger, in a power envelope that's at least 15% larger than the default power envelope of the FX8350, and everyone and their dog criticized the FX8350 for that power envelope, Intel first.

People also have to decide what they want from their CPU. They might want less cores that ramp up higher, and dedicate the entire power envelope to clock speed. OK, get a consumer SKU from Intel or an R5 then, pump it up to 120 W and beyond, and let it rip. I bet you that the R5 will be more consistent in performance than Skylake, which is a bitch for throttling, despite being the Nth gen of that particular technology.

There are criticisms I've seen against Ryzen that seem to the point, like immediately there were some people exploring how they could reduce the solder layer between the die and the heat spreader. I think that it is true that the present Ryzen chips are limited in performance because of manufacturing constraints. I believe that the next gen of Ryzen chips will have a precision soldered casing with a very thin solder layer, but I believe it was impossible to implement that for the first production batch. I believe that the actual technology allows for wider operational ranges, but the production toolchain simply does not. There is a reason why AMD soldered the heat spreaders, and didn't just do a cheaper gooey TIM-based pressure fit casing. I know of several people who are experimenting with resoldering the casing to reduce the solder layer to get better thermal conductivity of the casing. I think they hit the problem right on the head.

Second problem I think Ryzen is facing, is that mobo manufacturers are struggling with VRM specs. I think that with the Ryzen specs comes a whole new chapter in VRM design. Remember the problems Intel chips caused with PSU's because of the new power states? In comparison, that challenge was nothing. Imagine having to deal with the inductance of very high frequency switching extremely variable loads without any tolerance, or it's bye bye die lol. Imagine having to use a bunch of chokes and digital switching power regulation, but at the same time having to avoid power spikes due to increased resistance at high frequencies, etc... a radical advancement in the operational envelope of the CPU, requires a similar advancement in everything around it.

2 Likes

@Zoltan
Am I right to assume the higher XFR for R5 is due to the disabled cores making it consume less power and therefore less heat?
There seems to be a correlation between the artificially imposed 41XX OC ceiling and the max XFR limits. Do you think that an un- hobbled XFR will act like Nvidia's GPU Boost 3.0 and find the max boost possible organically allowed by the silicon?

I am not the one you addressed. But I think I can answer your points

Yes having less cores actives means less heat production and thus slightly higher headroom, however we need to weigh this against the fac that the R5's are in effect defective R7's so the silicon lottery isn't really on our side to begin with. There's also the hard current limit of the 14nm Process the die was made on which doesn't tolerate high voltages well. So at the end of it we are back in R7 territory with regards to clocks on the R5's.

Probably/Maybe. There isn't quite enough info out there to go beyond speculation but technically speaking it could totally be done, business decisions and other factors may limit the extent of that however.

Will the defective silicon be a factor if it is bypassed by being disabled?

Realistically it has an effect on the entirety of the chip. There's a reason some chips are stable at much lower voltages than others this is linked to the 'leakyness' and thermal characteristics of each die.

Simply speaking because the dies used for R5's where of lower quality and had two or more cores disabled/defective also means that the majority silicon of the chip would be lower grade since it is this that lead to some cores being defective in the first place.

Some recent benchmarks though have shown that the R5 line etc will too reach 4Ghz easily on air like the R 1700. Voltages vary between 1.42V and 1.36V.

Ofcourse this is not going to be the case for all chips. Some R5's may have been perfectly capable R7's but to make up production numbers there will be some 'golden' R5's that are technically fully functional R7's with disabled cores. This time though they will physically be disabled, so no Phenom X3 to X4 or X6 mods.

On a tangent from this: I somewhat expect AMD to later make purpose built Ryzen chips that only have say 4 or 6 cores in silicon rather than disabled ones as the manufacturing yields get higher. Or alternatively they could just start selling 8 core chips for dirt cheap too once there aren't enough R5's that can be churned out.

I suspect many of the 1700 benchmarks and OCs were perfectly functional chips downclocked and undervolted.
Do you anticipate a hack to unlock cores in the R5s such as was done with RX 460s?
EDIT: I just realized you answered this already.

Who knows. There are two thermal constraints that are indepedent of the design, and that's thermal density and thermal capacity. Thermal density-wise, what you say may be true, it's certainly true that less cores on the same die does mean less thermal density. Thermal capacity-wise, is another story. It's obvious that the thick layer of solder stores thermal energy, has thermal capacity, making the dissipation harder. It's highly probable that a tool chain optimisation will lead to thinner and more precise solder joints between the CPU and the heat spreader in future generations or production batches. How these two factors correlate is impossible to say, basically only Global Foundries knows.
The less power part in relation to heat production is not linear. Less power does not mean less heat. Heat is inefficient use of power. The relation is logarithmic.

The discussion of AM4 VRMs and their ability at such an early stage to keep Ryzen operating at their respective speeds is of interest to me. I've noticed there is quite a difference in build quality even within X370s. If I understand you correctly, AMD's coded OC safety limits were in anticipation of the varying quality of these boards?

I think that is highly probable, as they've used the same strategy in the past. For them it's a solution for the potential problems that can come with their "all SKU's unlocked" policy, together with the fact that they also allow for lesser spec OEM implementations. They can't upgrade the solder joint with a microcode update, if they don't incorporate a hard cap, they will create dissatisfaction as updated production batches come out of the same SKU's.

I plan on acquiring an R5 1600X in the next month and I was wondering if there is setting in the AsRock X370 UEFI that should remain enabled to allow XFR to operate above a manual overclock. One poster stated that XFR is disabled upon overclocking, but I suspect this was because he disabled some power saving settting(s). Have any of you experimented with this in order to find a way to leave XFR enabled?

I don't have an R5 and probably won't be getting one any time soon. With every UEFI update, there have been settings reset and altered on both the AsRock and Gigabyte mobos I have with R7's. The update frequency is pretty high, so you'll just have to try out what works on R5. I've noticed that cumulative settings sometimes cause a feature to stop working etc, there is some degree of experimentation involved.

Actually, I'm only about 99% sure about the 1600X. I'm thinking that it might be foolish not to go for the 1700 at that price if I could only be assured of the quality. My biggest fear is that at this late stage I might get an actual 1700 that won't even do 3800Mhz and not a 1800X in disguise like so many reviewers seem to have with their 4Ghz OCs.

all reviewers use 1.4+v to get their 4GHz... i can hit 4.2GHz on 1.47v... it runs hot at that point. Better to keep it at 3.8-3.9 at 1.35v

1 Like

I appreciate what you are telling me, I wasn't aware that many needed to push that much voltage.
I'm not comfortable with running such high voltage 24/7 even though I believe my cooling might be adequate.
I am thinking on selling off the AIO loop I currently run on this machine as I've suddenly become scared of leakage on the new system. I bought a new Air cooler to tryout on the new chip, but if necessary I'll get a larger one if my results are not acceptable.

there are better gains in performance from overclocking memory.

In my case the i can't get anywhere higher 2933MHz with decent timings.

you can "safely" run up to 1.5v or so, really up to 1.55v fine.
by "safely" i just mean if you have good cooling

you have to remember that voltage numbers are different on AMD chips than intel anyway.

1 Like

I'm getting around 78'C on 1.47v @4.2GHz on full load for around 20min run. (running on aio) it may have been going into ~90'C easily.

That's one of the compromises I had to make with my budget unfortunately.
I almost bought a B350 because the Taichi I lusted for was too much money for me, but suddenly I was informed that the Crossfire advertised was only at PCI-E 4x. Forced now to buy an X370, I called Newegg to complain about their bogus sale prices and was able to negotiate an AsRock Killer SLI/ac for $1 less than the B350 board I almost purchased. The compromise was that it only specs 2667 and not 3200, but I'll just have to suffer--- LOOL!

@anon5205053
This 4200Mhz is on a 1700 non-X?