Navi/Ryzen 3000 launch Megathread

[Drak3]

1 minute ago, S w a t s o n said:

https://www.anandtech.com/show/14605/the-and-ryzen-3700x-3900x-review-raising-the-bar/19

Quote

Looking at the total power consumption of the new 3700X, the chip is very much seemingly hitting and maintaining the 88W PPT limitations of the default settings, and we’re measuring 90W peak consumption across the package.

Seems like that limit isn't iron clad.

 

Nor does it counter the point of these chips still taking 1.4V to hit 4.4GHz.

[S w a t s o n]

Just now, Drak3 said:

Seems like that limit isn't iron clad.

 

Nor does it counter the point of these chips still taking 1.4V to hit 4.4GHz.

It cant be removed as far as reviews go, I want it to be disabled or raised of course.

As for voltages.

1 hour ago, S w a t s o n said:

Resistivity goes up as the node goes down. Basically the "wires" in the CPU are so thin that the resistance goes up massively. This lead to TSMC and Intel starting ot use cobalt and other material changes in 7/10nm. So it does make some sense that a high voltage is required

https://fuse.wikichip.org/news/525/iedm-2017-isscc-2018-intels-10nm-switching-to-cobalt-interconnects/2/

Cobalt as a raw material is actually more resistive than copper but the barrier layers and other factors mean it's a better choice when implemented into a CPU. As the nodes shrink further resistivity will continue to go up.

 

[cj09beira]

1 hour ago, Drak3 said:

Side note, having done the number crunching of AMD's IPC figures awhile back, matching Intel is a tad disappointing, IMO.

do you mean matching single core performance? because ipc is certainly higher, 

amd now should start thinking of ways to improve clocks too thats their last major disadvantage 

 

1 hour ago, Drak3 said:

Seems like that limit isn't iron clad.

Nor does it counter the point of these chips still taking 1.4V to hit 4.4GHz.

i am hoping they will soon tell us what are the safe voltages for zen 2, because i have a feeling 1.4 might be too much

[leadeater]

13 minutes ago, Drak3 said:

At some point I would like someone to post an actual, reputible source saying that the hardware is actually limited to 140W.

The package power limit very much exists, Anandtech isn't the only place giving that information out. PBO can hit higher because it's allowed to under the EDC specification, 140A at 1.35V is 189W. However I'm not sure what voltage EDC is reference to or if that 140A value changes if the voltage is changed. Delivered by the VRM my also not be the correct wording as 90% of 189W is 170W which is suspiciously close to what the reviews show. Edit: 140A may be the input current rather than output current which is the way I read that.

[S w a t s o n]

6 minutes ago, leadeater said:

The package power limit very much exists, Anandtech isn't the only place giving that information out. PBO can hit higher because it's allowed to under the EDC specification, 140A at 1.35V is 189W. However I'm not sure what voltage EDC is reference to or if that 140A value changes with the voltage is changed. Delivered by the VRM my also not be the correct wording as 90% of 189W is 170W which is suspiciously close to what the reviews show.

We should check the specific mobo used and calculate the VRM efficiency, it could be that they are not at the 90% sweet spot because they are too low on the curve as well. It's very common to have a VRM in the 70%'s at low usage.

170W on an Aorus Master or Extreme actually would almost certainly be at "low" efficiency. Boards can compensate by turning off phases but then the x570 flagship boards with 12-14 VCore phases would just turn off like 50-70% of the phases at max load under the current power limits. AKA the entire VRM NEVER turns on. Which they wont do. I think???

[thorhammerz]

 

What's the fuss about? The chips are evidently capable of eating 170W under load, even at 1.45v core voltage.

 

Debauer was nice enough to provide a graph at 5:55 for those interested in frequency/power/temperature scaling. All-core boost frequencies beyond the advertised XFR boost is probably not going to be possible without sub-zero cooling and/or golden-samples.

[S w a t s o n]

1 minute ago, thorhammerz said:

What's the fuss about? The chips are evidently capable of eating 170W under load, even at 1.45v core voltage.

 

Debauer was nice enough to provide a graph at 5:55 for those interested in frequency/power/temperature scaling.

170W measured before efficiency losses. Read my last post for why that matters. On a high end board the VRM is probably not even over 80% efficient due to massive overkill. The sweet spot for a VRM is usually like 50%+ usage

[leadeater]

1 minute ago, S w a t s o n said:

It's very common ot have a VRM in the 70%'s at low usage

I don't think so, efficiency peaks very quickly so it would have to be a very large VRM design and very low power draw. 140W/170W wouldn't meet that low power draw condition though. "Delivered by the motherboard VRM" is a little non specific, to me that saying output current but that can equally mean what the VRM itself is handling i.e. input current.

 

Either way EDC seems like the likely reason for seeing higher power figures than 140W if they are taking max values or short term average since EDC is a short therm allowance. I'm thinking the order they apply are EDC then PPT then TDC however any single one can apply singularly.

 

Older Z97 example

[descendency]

 

[S w a t s o n]

1 hour ago, leadeater said:

I don't think so, efficiency peaks very quickly so it would have to be a very large VRM design and very low power draw. 140W/170W wouldn't meet that low power draw condition though. "Delivered by the motherboard VRM" is a little non specific, to me that saying output current but that can equally mean what the VRM itself is handling i.e. input current.

 

Either way EDC seems like the likely reason for seeing higher power figures than 140W if they are taking max values or short term average since EDC is a short therm allowance. I'm thinking the order they apply are EDC then PPT then TDC however any single one can apply singularly.

 

Older Z97 example

12-14 phases of 70A is a very large VRM for 170W of power. My 580 sits at just around 70% when idle at like ~45W on a VRM that is like 1/3rd as powerful. 170W on a VRM that can do 500W+ is pretty bad for efficiency and I think that's exactly the discrepancy here. I disgaree that it would not meet the lower power draw condition, just do the math on the x570 boards.

[thorhammerz]

1 minute ago, S w a t s o n said:

170W measured before efficiency losses. Read my last post for why that matters. On a high end board the VRM is probably not even over 80% efficient due to massive overkill. The sweet spot for a VRM is usually like 50%+ usage

 

Shall we go back and re-assess those VRM inefficiency losses to every other CPU review? Compare apples to apples, oranges to oranges .

[Drak3]

4 minutes ago, cj09beira said:

do you mean matching single core performance? because ipc is certainly higher,

Exactly. While it demonstrates that AMD's IPC numbers are good, we're seeing 1.4v at 4.3GHz, which is the number I used for my preliminary number crunching. Which is boring at this point.

 

7 minutes ago, cj09beira said:

i am hoping they will soon tell us what are the safe voltages for zen 2, because i have a feeling 1.4 might be too much

https://fuse.wikichip.org/news/2408/tsmc-7nm-hd-and-hp-cells-2nd-gen-7nm-and-the-snapdragon-855-dtco/

TSMC's 7nm uses cobalt contacts, but I can't find anything regarding a total replacement of copper with cobalt.

 

As it stands, I don't see any reason for Zen 2 to claw back higher safe voltages.

[S w a t s o n]

Just now, thorhammerz said:

 

Shall we go back and re-assess those VRM inefficiency losses to every other CPU review? Compare apples to apples, oranges to oranges .

Oh sure, I am not comparing to other CPUs. Only regarding if the power limit of 140W is being exceeded. To compare to other CPU's using the 170W number is 100% correct

[cj09beira]

8 minutes ago, descendency said:

 

amd really needs to get rid of stock coolers, just give us aib cards, never mind that one case where blower is better 

7 minutes ago, S w a t s o n said:

12-14 phases of 70A is a very large VRM for 170W of power. My 580 sits at just around 70% when idle at like ~45W on a VRM that is like 1/3rd as powerful. 170W on a VRM that can do 500W+ is pretty bad for efficiency and I think that's exactly the discrepancy here. I disgaree that it would not meet the lower power draw condition, just do the math on the x570 boards.

keep in mind though that while true the boards at auto will turn phases off to reach better efficiency

6 minutes ago, Drak3 said:

Exactly. While it demonstrates that AMD's IPC numbers are good, we're seeing 1.4v at 4.3GHz, which is the number I used for my preliminary number crunching. Which is boring at this point.

 

https://fuse.wikichip.org/news/2408/tsmc-7nm-hd-and-hp-cells-2nd-gen-7nm-and-the-snapdragon-855-dtco/

TSMC's 7nm uses cobalt contacts, but I can't find anything regarding a total replacement of copper with cobalt.

 

As it stands, I don't see any reason for Zen 2 to claw back higher safe voltages.

i meant that the max safe voltage might be lower than 1.4v

[Drak3]

Just now, cj09beira said:

i actually meant that the max safe voltage might be lower than 1.4v

Eh, the move to Cobalt might mitigate the die shrink enough that 1.4v is still safe. We'll see though.

[S w a t s o n]

3 minutes ago, cj09beira said:

keep in mind though that while true the boards at auto will turn phases off to reach better efficiency

 

20 minutes ago, S w a t s o n said:

Boards can compensate by turning off phases but then the x570 flagship boards with 12-14 VCore phases would just turn off like 50-70% of the phases at max load under the current power limits. AKA the entire VRM NEVER turns on. Which they wont do. I think???

Five 70A phases can do 350A max (but they will be uncoolable) or like 140A EDC limit (and be a normal VRM)

[cj09beira]

1 minute ago, S w a t s o n said:

 

 

i meant at iddle and low load, i think they turn everything on when at "high" loads

though some of the new vrms are crazy powerful but this to me is more of a hey look at me i have the bigger D** than anything else, and it has the advantage to be ready in case that zen 3 has even more cores

[leadeater]

12 minutes ago, S w a t s o n said:

12-14 phases of 70A is a very large VRM for 170W of power. My 580 sits at just around 70% when idle at like ~45W on a VRM that is like 1/3rd as powerful. 170W on a VRM that can do 500W+ is pretty bad for efficiency and I think that's exactly the discrepsnacy here.

Well yea but given how fast the efficiency hits 90%+ 170W isn't going to be 70%, I don't think motherboard designers are that short sighted (lol 80+ PSU specs when the system itself is that inefficient). 140A input just makes more sense since it lines up with an expected 90% VRM efficiency almost perfectly for the 170W we have been seeing.

 

Thinking about it if you have a current clamp on the EPS you are looking at input current to the VRM, vcore for that 3900X was 1.34 from memory (if not 1.35) which means 127.1A which is lower than EDC 140A anyway. So yea no idea.

[descendency]

1 minute ago, cj09beira said:

amd really needs to get rid of stock coolers, just give us aib cards, never mind that one case where blower is better 

They've got a few things they need to fix, including the launch drivers that come with new cards. These are never 'good' but AMD manages to make them worse than they need to be. They're good cards marred by a bad cooler, bad drivers, and being a bit late to the game. 

[S w a t s o n]

48 minutes ago, cj09beira said:

i meant at iddle and low load, i think they turn everything on when at "high" loads

though some of the new vrms are crazy powerful but this to me is more of a hey look at me i have the bigger D** than anything else, and it has the advantage to be ready in case that zen 3 has even more cores

Of course but I'm saying that at "high loads" the efficiency could easily be around 80% at best with all the phases turned on just to switch 170W worth of power. Each phase turned on causes switching losses

 

47 minutes ago, leadeater said:

Well yea but given how fast the efficiency hits 90%+ 170W isn't going to be 70%, I don't think motherboard designers are that short sighted (lol 80+ PSU specs when the system itself is that inefficient). 140A input just makes more sense since it lines up with an expected 90% VRM efficiency almost perfectly for the 170W we have been seeing.

 

Thinking about it if you have a current clamp on the EPS you are looking at input current to the VRM, vcore for that 3900X was 1.34 from memory (if not 1.35) which means 127.1A which is lower than EDC 140A anyway. So yea no idea.

It varies a lot from powerstage to powerstage. I see the one you linked is a 40A power stage that hits max efficiency at only 25% load which seems low. Even if the 70A stage has the same curve (it wont), that means at 25% each VRM can push 18A or 216A total on a 12 phase. Much more than EDC of 140A. AFAIK 50% is the typical sweet spot but I'm not buildzoid. It will matter which mobo they are using exactly. Some mobos have 40A some have 50A some 60 and some 70A. I know a lot of people were using aorus masters with 70A stages.

So basically 140A does not line up with any expected sweet spot for flaship VRM's that I know unless they are a lower one. They arent even close

[cj09beira]

1 minute ago, descendency said:

They've got a few things they need to fix, including the launch drivers that come with new cards. These are never 'good' but AMD manages to make them worse than they need to be. They're good cards marred by a bad cooler, bad drivers, and being a bit late to the game. 

i dont get why they have that problem, they seem to have those problems but then fix then in 1-2 days which is just weird, if its such minor things how are they not being caught, they need more R&D thats for sure 

1 minute ago, S w a t s o n said:

Of course but I'm saying that at "high loads" the efficiency could easily be around 80% at best with all the phases turned on just to switch 170W worth of power. Each phase turned on causes switching losses

from what i have seen from ahoc the difference in heat loss isn't large in any of them so i wouldn't worry about it, though its funny that the most efficient setting is disabling a chunk of it 

[S w a t s o n]

11 minutes ago, leadeater said:

 

Assuming 50% load on each VRM it would only take 4 phases. So they are running 3x the phases at 17% load (12 A) on a 12 phase to reach 140A

[leadeater]

13 minutes ago, S w a t s o n said:

Of course but I'm saying that at "high loads" the efficiency could easily be around 80% at best with all the phases turned on just to switch 170W worth of power. Each phase turned on causes switching losses

128A divided by 14 = 9.14A which is well in to 90% efficiency on that older Z97 graph I posted. Even at 140W (103.7A) that's 7.4A per phase across 14. These newer boards are using smart power stages too rather than the more simplistic designs. I'd have to go watch a bunch of GN/ActuallyHardcoreOverclocking VRM analysis videos to go look at the actual product component specs and up-verse myself in the specifics again but I really don't see any motherboard manufacture releasing a product on to the market with such atrocious efficiencies at standard operating levels unless it was an XOC only product, like a Kingpin GPU.

 

45W at 1.2V is 37.5A and on a 8 phase VRM that's 4.69A per phase which would explain the really bad efficiency at idle, my 290X is just as bad. Bigger VRM would only make that worse. On that though, since CPU power draw at idle is so low VRM efficiency must be bloody terrible. 

[S w a t s o n]

33 minutes ago, leadeater said:

128A divided by 14 = 9.14A which is well in to 90% efficiency on that older Z97 graph I posted. Even at 140W (103.7A) that's 7.4A per phase across 14. These newer boards are using smart power stages too rather than the more simplistic designs. I'd have to go watch a bunch of GN/ActuallyHardcoreOverclocking VRM analysis videos to go look at the actual product component specs and up-verse myself in the specific again but I really don't see any motherboard manufacture releasing a product on to the market with such atrocious efficiencies at standard operating levels unless it was an XOC only product, like a Kingpin GPU.

 

45W at 1.2V is 37.5A and on a 8 phase VRM that's 4.69A per phase which would explain the really bad efficiency at idle, my 290X is just as bad. Bigger VRM would only make that worse. On that though, since CPU power draw at idle is so low VRM efficiency must be bloody terrible. 

The efficiency of the 40A stage is not the same as the 70A stage. 9.14A on a 40A stage is not the same % load on the VRM as 9.14A on a 70A stage. the 70A stage will be much much less efficient, see my above post. A 12 phase 70A VRM is running at like 17% load per phase at 140A, not at all within the sweet spot.

 

25 minutes ago, leadeater said:

Yea I get that however unless you are right on that very low edge the efficiency comes on very quickly. You're right 70A per phase is very high and at 140W means only 10% utilization on a 14 phase design which would be easily around that efficiency cliff. That itself could be a reason not to buy a massive VRM board if only going with a lower SKU, wonder how that effects reviews too (total system draw comparisons)

My point is also that I feel the stage you showed is very special or something else is going on. The peak comes on way too early. I'm going to revisit some buildzoid videos but 25% seems WAYYYYY too early for peak efficiency. It shoudnt come on that quickly.

[leadeater]

1 minute ago, S w a t s o n said:

The efficiency of the 40A stage is not the same as the 70A stage. 9.14A on a 40A stage is not the same % load on the VRM as 9.14A on a 70A stage. the 70A stage will be much much less efficient, see my above post. A 12 phase 70A VRM is running at like 17% load per phase, not at all within the sweet spot.

Yea I get that however unless you are right on that very low edge the efficiency comes on very quickly. You're right 70A per phase is very high and at 140W means only 10% utilization on a 14 phase design which would be easily around that efficiency cliff. That itself could be a reason not to buy a massive VRM board if only going with a lower SKU, wonder how that effects reviews too (total system draw comparisons)