Ryzen 7 1700 + B350 Overclocking Tidbits

[sirmonkey1985]

I think we'll probably see 4c parts based off of both Summit Ridge and Raven Ridge. Potentially even RR parts without graphics.

I doubt they will make a unique 4c mask without any graphics; at their volume they want to be able to use a wafer as much as possible even if individual dies aren't all fully utilized.

Going by the options on my board they can definitely harvest 2+2 and 4+0 parts. Its not impossible they could harvest 3+1 one parts as well (3+3 is an option for me, cant recall for sure what else and don't feel like bouncing right now).

the way the die is laid out it would be really easy for it to be converted to a 4c part since each 4 core cluster is on it's own half the die. it's not like they're making a 4c processor from scratch and their yields aren't bad enough to be worth butchering 8c parts for 4c parts.. the chips that fail to meet spec requirements for the 4c/8t will probably be converted to the 4c/4t chips. and agree with what snowdog said with the 6c chips but i'd bet the initial ones they use are probably ones that couldn't hit spec clock speeds completely stable with all 8 cores, maybe later on you'll start seeing perfectly working 8c chips with disabled cores..

 

[ecmaster76]

I'm not sure anyone posting here knows what there yields really are. But I'll bet a wooden nickel we don't see any quad cores other than cut Summit Ridge and Raven Ridge parts for a while. Intel has so many dies because they own their fabs and have unmatched volume.

When was the last time AMD had more than two CPU lines going at once? (not counting console/embedded parts and MCM parts)

 

[bwang]

I'm not sure anyone posting here knows what there yields really are. But I'll bet a wooden nickel we don't see any quad cores other than cut Summit Ridge and Raven Ridge parts for a while. Intel has so many dies because they own their fabs and have unmatched volume.

When was the last time AMD had more than two CPU lines going at once? (not counting console/embedded parts and MCM parts)

we've had two lines (APU and FX) for a while now
never had more than one high performance die though, not since Athlon 64 and Athlon FX coexisted.

 

[Snowdog]

I'm not sure anyone posting here knows what there yields really are. But I'll bet a wooden nickel we don't see any quad cores other than cut Summit Ridge and Raven Ridge parts for a while. Intel has so many dies because they own their fabs and have unmatched volume.

When was the last time AMD had more than two CPU lines going at once? (not counting console/embedded parts and MCM parts)

I think the big difference is 8 core FX parts were dirt cheap, and AMDs IPC was so low that the quad core FX parts were really undesirable.
Check the prices even when new:
http://www.anandtech.com/show/6396/the-vishera-review-amd-fx8350-fx8320-fx6300-and-fx4300-tested

The fastest 4 core FX, costs more than the lowest end 6 core. The really didn't want anyone buying quad cores. They were just on the spec sheet to convince you to buy a 6 or (preferably) 8 core.

OTOH 8C Ryzens are much more expensive and, and Quad Ryzens should actually be very decent and desirable (the volume part).

It's one thing to waste half an 8 core die for a quad cores when almost no one wants them and they sell in minuscule volume.

But it changes when they are quite decent and likely to be in higher demand.

AMD will literally be doubling cost on what should be a volume Quad Core part, if they don't create a unique die for it. That has the potential to more than double profits if they do create a unique die.

Example
8C die cost: $100 4C selling price $120: Profit 20 dollars.
4C die cost: $50 4C selling price $120: Profit 70 dollars. (>3X as much profit).

Using an 8C die for 4C product is a terrible idea unless you are not planning to sell many, and actively discourage sales.

 

[ZeroBarrier]

Example
8C die cost: $100 4C selling price $120: Profit 20 dollars.
4C die cost: $50 4C selling price $120: Profit 70 dollars. (>3X as much profit).

Using an 8C die for 4C product is a terrible idea unless you are not planning to sell many, and actively discourage sales.

You forget that what AMD is probably going to do is take the 8C/16T chips that have failed cores (2 cores failed, 4 cores failed) and sell them as 6C/12T and 4C/8T chips instead. You can bet your bottom dollar they are stockpiling those suckers right now.

Making a unique 4C die would be really inefficient of them; in your scenario they would just toss them to the trash and lose all cost associated with those failed chips.

 

[ecmaster76]

yep, and a straight, non-raven ridge 4c part would be inefficient in it's own way. What would they do with those that had defective cores? Sell cut 2c parts?

For years now they have had one "high performance" part and one APU and used the same die for all models in each of the two lines. I think they stopped making non-Xbox bobcat cpus around when they sold their fabs

 

[Snowdog]

You forget that what AMD is probably going to do is take the 8C/16T chips that have failed cores (2 cores failed, 4 cores failed) and sell them as 6C/12T and 4C/8T chips instead. You can bet your bottom dollar they are stockpiling those suckers right now.

Making a unique 4C die would be really inefficient of them; in your scenario they would just toss them to the trash and lose all cost associated with those failed chips.

You are drastically overestimating "failures". AMD likely won't have enough "failures" to make 6 core parts. Let alone four core parts.

There simply aren't enough "failed" parts to scavenge. The disabled parts on CPUs/GPUs are more of a marketing move, than a practical recovery of partially "failed" parts.

In reality most cut down parts are perfectly fine. In the past people have managed to re-enable cut down units on video cards, and when the formula gets out, it works for almost everyone who tries it.

You are also failing to account for them selling many more 4 core parts, than 8 core parts.

If you sell 4x as many 4 cores, that destroys your ASP for big die... AMD desperately needs to get ASP up.

Building 4 cores off the 8 core die, only makes sense, if your 4 core was undesirable as 4 core FX parts and you aren't going to sell many.

 

[ZeroBarrier]

You are drastically overestimating "failures". AMD likely won't have enough "failures" to make 6 core parts. Let alone four core parts.

There simply aren't enough "failed" parts to scavenge. The disabled parts on CPUs/GPUs are more of a marketing move, than a practical recovery of partially "failed" parts.

In reality most cut down parts are perfectly fine. In the past people have managed to re-enable cut down units on video cards, and when the formula gets out, it works for almost everyone who tries it.

You are also failing to account for them selling many more 4 core parts, than 8 core parts.

If you sell 4x as many 4 cores, that destroys your ASP for big die... AMD desperately needs to get ASP up.

Building 4 cores off the 8 core die, only makes sense, if your 4 core was undesirable as 4 core FX parts and you aren't going to sell many.

If you somehow think that they don't have a pretty sizable portion of failures on a new uarch chip, then you are sadly mistaken. The only one overestimating here is you; overestimating a brand spanking new 14nm uarch chip and thinking their failure rates will somehow be impressively low, as if they somehow hit a magic silicon goldpot.

 

[Dreamerbydesign]

So many arm chair experts in this thread and I'd be willing to bet not a single person is directly actively involved in working in this industry let alone AMD. That's what forums are for right?


I'd like to hear a response from AMD directly on their plans.

EDIT for spelling errors and auto correct

 

[ecmaster76]

So many arm chair experts in this thread and I'd be willing to bet not a single person is directly actively involved in working in this industry let alone AMD. That's what forums are for right?

But all the sudden we have 1500 cup yield and business/invententory/ logistics experts here.

I'd like to hear a response from AMD directly on their plans.

Its pretty simple - they have announced 6c and 4c parts. They are all Summit Ridge based otherwise we would have heard rumors about it by now. They have Raven Ridge, a 4C APU, out in a few months which we have also heard of. No other designs have been announced.

It could happen I suppose but lately we hear about a new design with plenty of lead time either through previews or leaks. Since we have not I have high confidence they aren't planning on releasing anything else for the time being.

 

[Snowdog]

If you somehow think that they don't have a pretty sizable portion of failures on a new uarch chip, then you are sadly mistaken. The only one overestimating here is you; overestimating a brand spanking new 14nm uarch chip and thinking their failure rates will somehow be impressively low, as if they somehow hit a magic silicon goldpot.

I challenge you to show any sane defect rate where it makes economic sense to produce a volume 4 core part, from an 8 core die.

In practice it will ALWAYS be more economical to make real 4 core dies.

The only reason to use the 8 core diet to make 4 core part, is if you expect the part to be so undesirable, that no one really wants it. If that is the case AMD is hosed. They can't recover from another unpopular CPU.

Fun with Yield Calculators:
http://isine.com/DieYieldCalculator.html

 

[ZeroBarrier]

I challenge you to show any sane defect rate where it makes economic sense to produce a volume 4 core part, from an 8 core die.

In practice it will ALWAYS be more economical to make real 4 core dies.

The only reason to use the 8 core diet to make 4 core part, is if you expect the part to be so undesirable, that no one really wants it. If that is the case AMD is hosed. They can't recover from another unpopular CPU.

Fun with Yield Calculators:
http://isine.com/DieYieldCalculator.html

This only makes sense to me if AMD sold chips in the same quantity as Intel, but they don't; they sell a mere fraction of that.

I can not concieve a way that AMD could afford to get 3 different chips manufactured in miniscule quantities when compared to Intel at pretty low margins (let's be honest, AMD doesn't expect to sell as many chips as Intel does; not even close. They won't be ordering in the same quantites as Intel produces, and lower quantities means higher cost per unit).

Admittedly, I have no knowledge of chip manufacturing, cost, logistics, etc; so I very well could be wrong. I'm only going by my limited knowledge and experience of mass produced products.

 

[Snowdog]

This only makes sense to me if AMD sold chips in the same quantity as Intel, but they don't; they sell a mere fraction of that.

I can not concieve a way that AMD could afford to get 3 different chips manufactured in miniscule quantities when compared to Intel at pretty low margins (let's be honest, AMD doesn't expect to sell as many chips as Intel does; not even close. They won't be ordering in the same quantites as Intel produces, and lower quantities means higher cost per unit).

And using a big 8c dies to sell 4c chip, will make margins much worse.

It doesn't matter that Intel is bigger. It matters if AMD will have enough volume to justify the layout. I argue that if they don't get enough volume to justify a layout, they aren't going to have enough volume to survive.

The overhead for lower volume is all up front. Creating a layout, doing multiple iteration to sort out the bugs. This is costs many tens of millions of dollars. You amortize that over all the chips you produce.

Though it should be MUCH easier to produce a half size ZEN, since much of the iteration work applies, and they can pretty much cut the die in half.

The question in my mind is if AMD is planning for success ( Thus Volume, thus a unique 4C layout), or if they are planning for low volume treading water (in which case using 8C die for 4c parts).

 

[razor1]

And using a big 8c dies to sell 4c chip, will make margins much worse.

It doesn't matter that Intel is bigger. It matters if AMD will have enough volume to justify the layout. I argue that if they don't get enough volume to justify a layout, they aren't going to have enough volume to survive.

The overhead for lower volume is all up front. Creating a layout, doing multiple iteration to sort out the bugs. This is costs many tens of millions of dollars. You amortize that over all the chips you produce.

Though it should be MUCH easier to produce a half size ZEN, since much of the iteration work applies, and they can pretty much cut the die in half.

The question in my mind is if AMD is planning for success ( Thus Volume, thus a unique 4C layout), or if they are planning for low volume treading water (in which case using 8C die for 4c parts).

The flip side of that, the margins even lower how much lower, we don't know the mark ups of the chips and what the current margins will be. For all we know right now, the 599 ryzen could be at 55% margins for AMD, and the cut down 4 core's are at 30% margins just like all of AMD current products. So end results they stay flat, is this why AMD has uncertainty in the future predictions of their financial calls? The uptake of their 8 core parts is what they were uncertain of, because of the problems we see now in early reviews.

But later on, the 4 core cut down parts might not be necessary as the APU version of the 4 core parts might be more attractive for most buyers.

Now if you want to amortize R&D costs there are two different things we need look at, fixed and variable costs, so with out breaking that down its hard to say what the effect of the cut down products will be for AMD's bottom line. They might be just fine, which is most likely what it will be.

 

[ecmaster76]

But later on, the 4 core cut down parts might not be necessary as the APU version of the 4 core parts might be more attractive for most buyers.

Yep, Raven Ridge is the long term main stream part. We might still see cut Summit Ridge parts after Raven launches. The advantage of the common socket is that they can use either die for the graphics-less parts depending on demand and what they have sitting around.

 

[noko]

You have to have the same number of cores active for each CCX (Compute Complex) keeping it symmetrical. http://www.eteknix.com/amd-ryzen-zen-6-core-cpus-unsurprisingly-exisit/

• So if one CCX has two bad cores and the other CCX is all good you have two choices - throw it away or make a four core cpu with 2 cores active per CCX. So you can still end up with a 4 core chip if only 2 cores are bad
• Next one would be 2 bad in one and 1 bad in the other CCX - that would also make the cpu only good for a 4 core processor
• Obvious one is 2 bad cores in each CCX = 4 core processor
• Only time you can make a good 6 core is if each CCX has one bad core or only one CCX has a bad core - meaning the other 7 are still good in that case

So enabling cores may be more tricky than before. As for one CCX having all bad cores and the other CCX still good? L3 would be be cut down but would be interesting because you could have some with 3 bad cores in one CCX and could still save the chip. We just have to see.

 

[bobzdar]

You have to have the same number of cores active for each CCX (Compute Complex) keeping it symmetrical. http://www.eteknix.com/amd-ryzen-zen-6-core-cpus-unsurprisingly-exisit/

• So if one CCX has two bad cores and the other CCX is all good you have two choices - throw it away or make a four core cpu with 2 cores active per CCX. So you can still end up with a 4 core chip if only 2 cores are bad
• Next one would be 2 bad in one and 1 bad in the other CCX - that would also make the cpu only good for a 4 core processor
• Obvious one is 2 bad cores in each CCX = 4 core processor
• Only time you can make a good 6 core is if each CCX has one bad core or only one CCX has a bad core - meaning the other 7 are still good in that case

So enabling cores may be more tricky than before. As for one CCX having all bad cores and the other CCX still good? L3 would be be cut down but would be interesting because you could have some with 3 bad cores in one CCX and could still save the chip. We just have to see.

Benchmarks show 4+0 faster than 2+2 (at least in win10) due to no swapping of threads between CCX's despite half the L3 cache. I'd imagine the 6 cores will be 8 cores that have 1 bad core on one (or both) of the CCX's, but not sure they'll make 4 cores that way as it'd be slower, at least with how things currently stand.

 

[pgaster]

Added some power figures for 38x on stock cooler, it was not stable on Prime95 Small FFT but the instability seemed to be entirely thermal, so the datapoint is worth something.

pgaster: I'd be interested in seeing if your copy can do 37x at 1.2V like mine, it seems to be a sweet spot (just look at those 38x power figures!)

I finally got some guts, changed the multiplier to 37, saved, and rebooted. It worked fine, and Ryzen Master said voltage was just under 1.2V. I think it was 1.1875V to be exact.
If the Ryzen Master software is really correct I'm a very happy camper. It showed the temps not going over 50 C, even during a cinebench benchmark. I'm only using the stock cooler.
Of course maybe that's not a benchmark that really generates much heat. Still, it felt snapper and did a bit better on a few things. Cinebench multi score went from 1399 stock to 1503 with the Ryzen Master running in the background. CPU-Z bench went from 16733 to 19094 for the multi benchmark. Single thread numbers were about the same, but I expected that since the stock turbo speed is also 3.7Ghz.
Anyway, good clocking for an inexpensive board. I'm happy.

I have Trident Z RAM that I can't get to run at it's XMP 3200 speed. I'm just leaving it at 2133 for now. I know that low speed isn't helping me any, but maybe after a few bios updates I can run the RAM faster. Or maybe I can try some more manual timings and find something that allows 2400 or 2666.

 

[noko]

Benchmarks show 4+0 faster than 2+2 (at least in win10) due to no swapping of threads between CCX's despite half the L3 cache. I'd imagine the 6 cores will be 8 cores that have 1 bad core on one (or both) of the CCX's, but not sure they'll make 4 cores that way as it'd be slower, at least with how things currently stand.

I was wondering about that, one CCX you don't have that slow communication between the cores but then with 2 CCX you have double the L3. So maybe there will be two skews for the 4 core ones, one with more L3 and 2 CCX and then the one with less L3 and CCX of course each may have different clock versions.

 

[chris89]

That's interesting about the thermals. Like the way I understand it is like the 1st gen i7's when we weren't ready to handle that kind of heat. Past a certain point stability went out the window along with temperatures.

So my question is what have you tried? I'm just interested since it appears everyone is just basically taking it up to maybe 105mhz bus and fiddling with the did and fid. I too ask why DID? haha no point

I guess logically just stick with DID 1 as if it wasn't even a part of the equation. If I was testing it, I'd focus on lower FID multipliers... By increasing the bus clock from 100Mhz to 200Mhz.

Have you tried DID of 1, FID of 5, and bus of 200Mhz? 200 x 5 is just 1Ghz... Just to establish 200Mhz bus as a possibility. Then note what Auto voltage determined for 200 x 5 compared to 100 x 10... Make sense?

Maybe 200 x 5 calls on less auto-voltage than 100 x 10... Meaning cooler possibly. Then just revolve around the 200Mhz bus multiplied by say 10+ or whatever for 3000Mhz+ and see the results.

As any of us know at least on Intel non-unlocked i7's, to achieve 4Ghz we needed 200Mhz BCLK. I think it'll be the same way. HAHA Cool

 

[faugusztin]

same here, I think it's a bug in CPU-Z; power consumption matched what was expected for the real voltage setpoint. not 1.55V.

CPU-Z often has bugs like this with new platforms - for example in 2014 when i got my new X99 system, it took some time until it was reading the correct voltage. It is probably reading some other voltage from the CPU.

 

[bwang]

That's interesting about the thermals. Like the way I understand it is like the 1st gen i7's when we weren't ready to handle that kind of heat. Past a certain point stability went out the window along with temperatures.

So my question is what have you tried? I'm just interested since it appears everyone is just basically taking it up to maybe 105mhz bus and fiddling with the did and fid. I too ask why DID? haha no point

I guess logically just stick with DID 1 as if it wasn't even a part of the equation. If I was testing it, I'd focus on lower FID multipliers... By increasing the bus clock from 100Mhz to 200Mhz.

Have you tried DID of 1, FID of 5, and bus of 200Mhz? 200 x 5 is just 1Ghz... Just to establish 200Mhz bus as a possibility. Then note what Auto voltage determined for 200 x 5 compared to 100 x 10... Make sense?

Maybe 200 x 5 calls on less auto-voltage than 100 x 10... Meaning cooler possibly. Then just revolve around the 200Mhz bus multiplied by say 10+ or whatever for 3000Mhz+ and see the results.

As any of us know at least on Intel non-unlocked i7's, to achieve 4Ghz we needed 200Mhz BCLK. I think it'll be the same way. HAHA Cool

The BCLK is tied directly to the PCIe clock on this platform, so cranking it up like crazy is out of the question.

 

[chris89]

The BCLK is tied directly to the PCIe clock on this platform, so cranking it up like crazy is out of the question.

The great thing about PCIe is it has it's own dedicated clock assigned. Which is 100.00 Mhz. If you don't see a PCIe slot frequency in the BIOS, then we can all rest assured it's sitting locked static at 100.00 Mhz.

In the past we could turn up the PCIe slot frequency but we would either see 0.1 more fps or simply no post or crash.

Luckily, not an issue on Ryzen. Unless you manually increase the PCIe Slot Frequency.

So basically to sum up, there is no HyperTransport or Northbridge link ratio on Ryzen. It's a dyanmic ever evolving value that will yield more or less performance depending on FID or BCLK. DID should be left at 1. As anything greater than 1 will either complicate the mathematical calculation of the total clock or allow in a way higher peak clock allowance on an otherwise not possible value via DID 1. That may not make sense, but only under the most extreme scenarios would greater than DID 1 be necessary.

To basically Soup Up the Infinity Pipeline, by Throwing on some high lift racing camshafts and twin turbos. Turn down the DID to 1, Turn down the FID to 1 or 5 to start. Just like a high Boosted Supra, low compression pistons provides continuous stability.

Start off DID x1, FID x1, BCLK 200 Mhz. Will it boot at 200 X 1? Most certainly. Then slowly raise FID to x2, boot, x3, boot, x4, boot. It will take time but at the end, some numbers you or anyone have not yet seen before from a single socket cpu platform will result.

 

[ecmaster76]

The great thing about PCIe is it has it's own dedicated clock assigned. Which is 100.00 Mhz. If you don't see a PCIe slot frequency in the BIOS, then we can all rest assured it's sitting locked static at 100.00 Mhz.

In the past we could turn up the PCIe slot frequency but we would either see 0.1 more fps or simply no post or crash.

No I think bwang was pretty much correct.

PCIe on Zen is tied to the base clock directly. Supposedly you can get bclock up substantially by dropping PCIe to gen2 or gen1 modes but its still trickier than multiplier adjustment

 

[chris89]

Thank you for the reply. By the way sorry for the Ultra TMI if it is so. I have read by only one source which wasn't AMD, static BCLK & PCIe are 1:1. Has anyone even tried setting DID x1, FID x1, and increasing the BCLK 1-5Mhz at a time until you find the max? If it keeps on handling it above 120Mhz, I think we would know PCIe can't really operate past 120Mhz. We could assume static locked PCIe slot freq ratio like all modern platforms.

I think that no one has tried that, they are starting off at something like 40 multi and then trying to raise BCLK, which isn't going to work. You have to lower Multi to Raise BCLK. Which is why DID x1 is nice in this respect. Even 133 x 30 is 4Ghz... That's way better than 100 x 40... That's 10 less multi hops, which is 33.33% less meaning a potential 33.33% reduction in required voltage to yield such values.

I had to collect some more info which is highly sparingly, only one person reported stock voltage value they apparently thought the R7 1700 needed for what 3500Mhz? Value 1.18v... is that accurate? What's the peak voltage value considering XFR is my question? Many variables to account for, not at all straight froward just the way I like it haha.

If this is correct let's just say if we set stock DID to x1, leave FID to auto and leave BCLK at 100. Then find the voltage at which it finds works at x1 DID, auto FID (Conclude all values of BIOS/ and many software sources then take all the values and average them all together). Then we take that value and calculate overclock voltage values by math. Pretty easy. If it thinks its doing 3500mhz on 1.175v (avg), then 1175 / 3500 = 0.3357%. We want 4,000Mhz for multiply by 0.3357 = 1342.8mv... This is best way to properly calculate voltage/ clock. With the stock values to guide your math on calculating overclocking voltages.

This is quite a interesting anomaly if so, haven't witnessed such a debacle since Athlon 64 days. Even then we had a BIOS option to manual lock down PCIe to 100.00 Mhz. Only on like the worst motherboards was the PCIe slot Freq not adjustable.

What piece of software could even relay the PCIe Slot frequency anyway? I can't find anything...

 

[TheLAWNoob]

I got the same setup. BIOS seems tp be broken. I set 1.3V 3.8Ghz with 2933 RAM, in Ryzen master I got 1.2V 3.8Ghz with 2133 RAM.

Testing with P95 and seems stable. I'll stick to 3.8Ghz and the Wraith until my AM4 bracket for my H105 comes.

 

[chris89]

Nice dude, that's awesome you guy's are lucky to be able to mess around with this. I really like testing new stuff, and conclude my own personal opinion.

Sorry for the too much information if it is, I know many would not even read it all. It isn't so much rambling if you can take something from it and test yourself. To find the actual values.

I received some more details, luckily gamersnexus threw out some numbers to work with (Thumbs Up).

So basically I suppose we could say stock at load voltage on all threads at lets just say 3200Mhz on all threads is a round about 1.068v (gamersnexus source). I'm only happy we have some numbers to go off of, which is nice.

Okay so if we could safely assume 3200Mhz on all 8 threads can do just fine with Peak voltage of 1.068v. Also including the dynamic understandable effects of spread spectrum being disabled. Which causes the voltage and freq to jump annoyingly all over.

We can say 1068 divided by 3200 = 0.33375 (beautiful number by the way). That feels like a percentage value we can trust. Let's forget about XFR like it wasn't even there and disable it, to just focus on the percentage value we just found.

Mathematically let's just say if DID is x1, BCLK is left at 100.00 Mhz. At this point I'm sure we will reach a slope in the too high of ratio to bclk territory. 3200Mhz divided by 100 = 32. 32 divided by 100 is 0.32.

That's the percentage ratio between FID & BCLK of 0.32% at 100 x 32 = 3200. So I think we should now just think about up in the air at which ratio will it dislike? I'm going to say issues will arise near the 0.50% mark. So let's say 0.44% FID to BCLK is max ideal.

So knowing this we can either raise FID to 0.44 which is basically simple it's 4.4Ghz which is 100 x 44. Issues will arise in this territory. So let's try the 2nd approach by approaching 0.44% on BCLK. Meaning 144Mhz BCLK multiplied by 22.22 = 3200Mhz. 22.22 divided by 144 = 0.154305%.

So with this math we actually yielded a major difference from the stock 0.33375% of 1.068v at 3200Mhz. I explained my math above haha. I could explain over and over but just yield a huge headache. haha

Okay so 144 x 22.22 = 3200Mhz same as stock of 100 x 32 = 3200Mhz. The ratio went from (32 divided by 100 = 0.32%), now to (22.22 divided by 144 = 0.154305%)... That's great and exactly what we needed.

So now with 144 BCLK, it would call on apparently some unknown voltage maybe less than stock. This is where experimentation comes to play haha. Guesstimation all day is pointless without testing.

Basically what I'm getting at here is there are different paths to overclocking success other than the one known clear one. Meaning straight forward 100 x whatever = some clock and voltage.

For instance for me to achieve 1,250mhz game stable on the r9 380x. It wasn't possible on the standard straight forward approach. I had to first find max allowable voltage of 1.449v no higher. Though 1.449v was not high enough to make 1,250mhz stable. The reason for this is stock voltage was 1040Mhz x 1.175 (17.5%) = basically 1.225v (rounded up many decimals). So knowing that 1250 x 1.175 = 1468.75mv... Obviously 20mv too high and I had my eyes set on Game-Stable 1,250mhz. So I had to find another approach to achieve stability without going past 1.449v. That meant an adjustment of the VDDCI memory voltage ratio to core Voltage. I found that running the VDDCI at 1.233v actually allowed now 1,250Mhz stable. That's core voltage 1.449v divided by 1.175 = 1.233v. Which by stock the VDDCi was 1175mv... I know TMI majorly but maybe it will help someone achieve a clock which otherwise was not possible.

 

[lolfail9001]

I think that no one has tried that, they are starting off at something like 40 multi and then trying to raise BCLK, which isn't going to work.

XOCers did keep multi low for their BCLK OC runs. Guess who told us you have to drop PCI-E from v3 to v2 to v1 as BCLK goes up?

Pretty easy. If it thinks its doing 3500mhz on 1.175v (avg), then 1175 / 3500 = 0.3357%. We want 4,000Mhz for multiply by 0.3357 = 1342.8mv... This is best way to properly calculate voltage/ clock. With the stock values to guide your math on calculating overclocking voltages.

Assumes linear voltage/freq scaling that is anything but true in real life and especially not true on Ryzen for frequencies above 3.3Ghz.

 

[chris89]

That's cool thank you for the reply. I appreciate it. I'm just interested just as much as anyone else. I haven't tested it so I don't know for sure of anything. One thing is for sure that idk if there is a Spread Spectrum option anywhere to be seen? That would make it actually run 3,000Mhz at 100.00Mhz BCLK x 30. Rather than 99.80 x 30 = 2994. Spread Spectrum is said to increase stability. However I find disabled is the way to go. Only on the biggest big end would spread spectrum might pull in extra numbers by accident or chance really.

That's too bad if the PCIe slot freq is dynamic and tied to the BCLK, like back in the Pentium 4 days and Athlon 64. I hope they release a BIOS with the option to set PCIe slot freq to static 100.00Mhz for stability.

I remember this on the ASUS A8N-SLI, had that extra molex on the board like what?!? haha and the PCIe Freq adjustment in the BIOS... Which showed 1 more fps up to 15Mhz faster and extremely unstable. Set 100.00Mhz and totally stable way up on the BCLK/ Multi. Could even do 300Mhz BCLK from 200Mhz BCLK with 100.00Mhz PCIe Freq Multi. Though performance was so far out there, that the HT & NB Link couldn't go low enough to go well beyond 3Ghz at the time. That was scorching speed.

 

[pgaster]

Wow, I just had something crazy happen and wanted to share.

I was frustrated with some Trident Z RAM rated for 3200 that I couldn't get to run past 2400. It's a 16GB kit, 2 x 8GB. I stopped by my local Micro Center and noticed some "open box" G.Skill 3000 ram for $79.99 so I took a risk and bought it. It a 16GB kit, 2 x 8GB. Full normal price is $99.99 and Micro Center has lots of it in stock.

Just popped it in my cheap Asus Prime B350-Plus, went into the bios and picked DOCP. It set the RAM for 2933 and I sort of chuckled thinking for sure it would lock up, but I hit save and tried it anyway. I was shocked when it booted to windows no problem. This ram is single sided, so probably single rank, and has stickers, no heat spreaders. I kid you not. It's cheap in every sense of the word, but taking a risk paid off. I'm posting with it running now. No issues at all.

Here is the link: http://www.microcenter.com/product/474187/16GB_8GB_x_2_DDR4-3000_PC4-24000_Desktop_Memory_Kit

Microcenter's site has the part number wrong, they missed the dashes. Here is the link to the G.Skill page that has correct part number and more info: http://www.gskill.com/en/product/f4-3000c16d-16gisb

 

[TheLAWNoob]

Is the memory speed detected properly?

I tried setting my 3000Mhz RAM to 3000 and 2933, both showed up as 2133 in windows. I set it to 2400 and won't even boot. Voltage is 1.35V

 

[pgaster]

Yes it shows correct speed in cpu-z.

 

[chris89]

I hope this might help someone, I hear among the testers. For running 3200Mhz, needs 16-16-16-32 1.5v.

 

[TheLAWNoob]

Update: I updated the BIOS on my ASUS B350 Plus Prime and it can now do 2933Mhz for my RAM rated for 3000Mhz.

Cinebench R15 went up about 1% going from 2133 to 2933. Now it scores 1677 at 3.79Ghz (the base clock is slightly below 100)


Here's 3D Mark Time Spy, with RX 480 reference 8GB running at stock clocks undervolted.

 

[kirbyrj]

I'm running this combo on the Asus Prime B350-Plus. Just got my bracket for the Scythe Fuma in, so I'm trying to OC on the latest 0606 bios. Here are some observations:

1. Looks like there is still no set voltage. Only offset, and it's limited to a .2V offset. (Not a terribly big deal to me as I don't want to go much higher than that anyway). SOC voltage is the same. .2V offset max
2. If you manually set the CPU ratio, there is no cool n' quiet or whatever they call it to drop the multi and voltage in low usage situations (at least I can't figure it out anyway). Kind of disappointing. Tried under balanced and high power settings in Windows.
3. XMP / DOCP or whatever they want to call it will correctly read my RAM, but still won't let me use the settings. Attempting to use the profile results in me having to do a bios reset by shorting the pins. When I try manually setting timings, it will lock up and require a bios reset if I set it to anything over 2666 mhz.
4. Seems to top out at 3.8Ghz or so with the RAM at 2133mhz.

 

[mlcarson]

I've got a Gigabyte GA-AB350-Gaming 3 + Ryzen 1700 + DDR4-3200 Flare X. F5 Bios. No direct VCORE setting -- only DVID (Dynamic Vcore offset) offset. Default voltage appears to be 1.1875V
This is what I've found on my board/chip so far:
3.750GHZ - default voltage
3.800Ghz - 1.2475V (.060)
3.900Ghz - 1.3015V (.114)
3.925Ghz - 1.3375V (.150)
Stability testing is primarily using OCCT without error for 1 hr. The only changes have been the Core ratio and the DVID. I've not touched the Dynamic Vcore SOC. Cooling is via Air on an NH-D15.


[edited] - default voltage is actually 1.1875V rather than 1.20V - adjusted accordingly
New results:
3.900GHZ - 1.3255V (.138)
3.925GHZ - 1.3555V (.180)
3.950GHZ - 1.3915V (.204)
3.975GHZ - 1.4215V (.234)
4.000GHZ - 1.4755V (.288)
This was after checking for WHEA errors and doing additional Cinebench and Realbench testing.
I continued adding voltage to the 4.000GHZ clock until the WHEA errors went away but didn't feel comfortable doing long-term testing at that voltage.

So, what's the current consensus on where the voltage is worth the minor clock speed improvement? Some people have said 1.35V is good for normal usage. Others have said 1.45V is fine with good cooling. For me, it looks like a choice between a nice comfortable 3.925GHZ @ 1.3555V or a more aggressive [email protected]. The difference appears to be about 5C at full load/stress testing based on gigabyte SIV Smart Fan 5 Adv CPU sensor:
3925: 23C idle, 53C load
3975: 24C idle, 58C load

 

[Nobu]

chris89 some motherboards (Asrock Taichi, all or most gigabyte aorus gaming k* boards, I think asus and msi have one each) have a separate clock generator for pcie and bclk. The ones without, you're limited by how forgiving the other chipsets are about the pcie clock.

On my ga-f2a88x-up4 fm2+ board with no separate clock generator, I could get up to 107 stable, up to ~130 without onboard networking, audio, and (non-soc) sata (so I had to boot from USB, and couldn't access the drives that were connected to the 3rd party sata controller)
Edit: oh, and my GPU wouldn't work either, so I needed to use the APU's graphics.

 

[Sir-Flip]

• Ryzen has two distinct bumps in efficiency. I'm wary of saying where they are based on a single sample, but it looks like the first is in the neighborhood of 35x and the second, around 38x. Pushing for that last two bins (or eight, if you want to be picky) causes an insane increase in power consumption on my sample; the sample to sample cutoff may vary but I feel safe in saying that you will lose a whole lot of performance/watt right around 3.8+/- GHz.

That's totally true from my experience with my r7 1700x. I managed to get 3.81 Ghz with 112Mhz BSCL and 34x multiplier on stock voltage (1.350V)rock stable without any fine tuning or tweaking and also without much higher power consuption. But as soon as I increased the BSCL to 113Mhz, it took 1.40V Vcore to get it stable. The highest OC I got stable seems to be 3.94 with BSCL of 116Mhz, but it required 1.5V vcore with my chip, and I don't like it that high. That's why I only ran cinebench 4 times and occt for 5 minutes before shutting the system down. I personally went with 3.81 Ghz for the long run, althought I really want that 4Ghz, I don't think my chip is going to make it stable with somewhat sustainable Volts.
I'm running my 1700x on an Asus Crosshair VI with a nuctua NH-D15 and a Corsair HX1000i PSU with windows 10 Pro on an 240g ssd and everything else on a 3tb 7200rpm 64mb buffer HDD and a GTX 750 ti for video output until I can get my hands on some Vega GPU for gaming.
Really nice conversations by the way

 

[mlcarson]

I'm wondering how much the VRM Mosfet is affecting overclocking on the less expensive B350 boards. I know on the GA-AB350-Gaming 3 that the VRM gets very hot when you start overclocking at 3.9+ Ghz and adding voltage and suspect that it's causing instability. The temps will climb into the 90's even with good air flow when the CPU is at max load for an extended time. I think a larger heatsink or some type of directed cooling would be in order unless these things can handle 100C+ without causing instability. My CPU temps haven't been a problem at all regardless of the amount of voltage and multiplier chosen but the VRM temps seem to get out of control.

 

[CAD4466HK]

I'm wondering how much the VRM Mosfet is affecting overclocking on the less expensive B350 boards. I know on the GA-AB350-Gaming 3 that the VRM gets very hot when you start overclocking at 3.9+ Ghz and adding voltage and suspect that it's causing instability. The temps will climb into the 90's even with good air flow when the CPU is at max load for an extended time. I think a larger heatsink or some type of directed cooling would be in order unless these things can handle 100C+ without causing instability. My CPU temps haven't been a problem at all regardless of the amount of voltage and multiplier chosen but the VRM temps seem to get out of control.

You might want to consider putting some sinks/bigger sinks and a 40mm fan on your VRM's.

https://hardforum.com/threads/am4-b350-x370-vrm-table.1928342/#post-1042945506