Ryzen: Strictly technical

Page 78 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

Zucker2k

Golden Member
Feb 15, 2006
1,810
1,159
136
Per Gamer's Nexus data, going from 4.0 Ghz (78 amps at 1.162 volts) to 4.2 Ghz (115 amps at 1.381 volts) requires 75% more power for a 5% frequency increase. So yes, each extra 100Mhz past ~4 Ghz is that bad.

Also, 4.7 ghz is just completely not possible on Ryzen 2000 without sub-ambient cooling.
So basically, XFR2/Precision Boost are the real highlights of Ryzen 2000 (R2K). The silicon itself is a dud; a paltry 100mhz over Ryzen? Where did the 10% fmax modest predictions all the resident experts were touting not so long ago in their predictions vanish to? And all this magic results in a further 48% power penalty at stock :eek: And double that if you try touching the overclocking button - and that at a record shattering 4.2Ghz!!

I'm sorry, if Intel tried to pull this stunt the whole internet would be up in arms lol. At this point in time, if you're on Ryzen, you should be seriously petitioning AMD to introduce the alchemy on R2K to Ryzen. That should save you a pretty buck towards the purchase of better RAM or GPU.





Trolling is not allowed.


esquared
Anandtech Forum Director
 
Last edited by a moderator:

Gyronamics

Junior Member
Apr 22, 2018
5
1
36
Where did the 10% fmax modest predictions all the resident experts were touting not so long ago in their predictions vanish to?

...

I'm sorry, if Intel tried to pull this stunt the whole internet would be up in arms lol.

Are you mixing up what casual speculators said and what AMD said (and then delivered). That's fairly unreasonable...
 

Wall Street

Senior member
Mar 28, 2012
691
44
91
(42/40).(1381/1162)^2 = 1.483 , that is, 48.3%...

Although power should scale at the square of voltage and linearly with clock, for my 75% more power calculation I was using P = I V.

78 amps x 1.162 volts = 90 watts at 4 Ghz
115 amps at 1.381 volts = 159 watts at 4.2 Ghz
159 / 90 ~ +75%
 

Wall Street

Senior member
Mar 28, 2012
691
44
91
So basically, XFR2/Precision Boost are the real highlights of Ryzen 2000 (R2K). The silicon itself is a dud; a paltry 100mhz over Ryzen?

The silicon is not really a dud. The 2700x can do 4.0 Ghz at 90 watts while the 1700 that Gamer's Nexus tested used 144 watts. That is 37% less power on 2000 series vs. 1000 series.

Everyone here is looking at max clocks, but the best improvement is in parts like the vanilla 2700 which should probably still perform on par with an 1800x despite being at only 65 watts.

The new Ryzen silicon is much more efficient that Ryzen 1, however the 2700x is what you get when you try to throw 110% of that extra power efficiency back into clock speed.

FYI, Intel also has crazy exponential power scaling too, however there bad scaling is in the region of 4.5-5.0 ghz. The main problem AMD face is that people have it in their head that 5 ghz is an achievable number and this part simply won't ever hit that number (or even get within 10%).
 

Abwx

Lifer
Apr 2, 2011
10,854
3,298
136
Although power should scale at the square of voltage and linearly with clock, for my 75% more power calculation I was using P = I V.

78 amps x 1.162 volts = 90 watts at 4 Ghz
115 amps at 1.381 volts = 159 watts at 4.2 Ghz
159 / 90 ~ +75%

Power scale as a square of voltage and linearly with frequency, obviously the current they state for 4.2GHz is either wrong or made up, the CPU total capacitance doesnt change with voltage or frequency, so there s no mean that getting to 1.381V@4.2GHZ from 1.162V @4GHZwill increase the current to 115A.

Assuming that their number at 4GHZ is right the current should increase to 98A or so at 1.362V@4.2GHZ, or else their sample follow laws of physics of another universe.

Btw, using 1.162@78A@4GHZ as basis we can compute that the CPU impedance at this frequency is 1.162/78 = 0.0149 millihom.

This value doesnt change with voltage but decrease linearly with frequency, hence its value at 4.2GHz is 0.0149.(40/42) = 0.0142 millihom.

The current at this frequency should be 1.381/0.0142 = 97.25A, actually slightly more because of leakage but nowhere at the level they are stating.
 

Wall Street

Senior member
Mar 28, 2012
691
44
91
Btw, using 1.162@78A@4GHZ as basis we can compute that the CPU impedance at this frequency is 1.162/78 = 0.0149 millihom.

This value doesnt change with voltage but decrease linearly with frequency, hence its value at 4.2GHz is 0.0149.(40/42) = 0.0142 millihom.

The Ryzen die integrates linear voltage regulators, so some of the internal components are running at voltages we can't even see. Therefore, I don't think it can be treated as a simple LCR load.
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
The Ryzen die integrates linear voltage regulators, so some of the internal components are running at voltages we can't even see. Therefore, I don't think it can be treated as a simple LCR load.

dLDOs for core domain are only used in EPYC.
On AM4 platform have been disabled (placed into bypass) since B0 silicon.
 
  • Like
Reactions: Drazick

Abwx

Lifer
Apr 2, 2011
10,854
3,298
136
The Ryzen die integrates linear voltage regulators, so some of the internal components are running at voltages we can't even see. Therefore, I don't think it can be treated as a simple LCR load.

Ryzen voltage regulators are in serial with the CPU cicuitry and their dissipation is included in the package power.

That being said Hardware.fr got 156W@4.3GHz@1.341V under Prime 95 measured at the ATX connector, this amount to barely 140W at the CPU level, and to 104.4A@1.341V.

If they use the same voltage but at 4.2GHz the current would decrease to 99.4A, wich is close to the figure i quoted, at 1.381V this would increase to 107.5A.

https://www.hardware.fr/articles/974-7/overclocking-pratique.html

https://www.hardware.fr/medias/photos_news/00/55/IMG0055234.png
 
  • Like
Reactions: lightmanek

Space Tyrant

Member
Feb 14, 2017
149
115
116
So basically, XFR2/Precision Boost are the real highlights of Ryzen 2000 (R2K). The silicon itself is a dud; a paltry 100mhz over Ryzen? Where did the 10% fmax modest predictions all the resident experts were touting not so long ago in their predictions vanish to? And all this magic results in a further 48% power penalty at stock :eek: And double that if you try touching the overclocking button - and that at a record shattering 4.2Ghz!!

I'm sorry, if Intel tried to pull this stunt the whole internet would be up in arms lol. At this point in time, if you're on Ryzen, you should be seriously petitioning AMD to introduce the alchemy on R2K to Ryzen. That should save you a pretty buck towards the purchase of better RAM or GPU.
Not even close.

The real highlight in my experience has been the 11% increased clock speed at the same voltage as my Ryzen 1600. I ran that chip at 3.6GHz @ 1.15vcore. This 2700X is currently running all cores at 4.0GHz @ 1.125vcore.

Of course, if you want to talk extreme OC, Der8auer hit 6GHz. But, other than the entertainmment value, that doesn't matter to me.
 
  • Like
Reactions: KompuKare

B-Riz

Golden Member
Feb 15, 2011
1,482
612
136
2700X, everything on auto in BIOS for the CPU, Asus Prime X470-Pro 4008, Wraith Prism, Arctic Silver 5, CPUID HWMonitor 1.35, CPU-Z 1.84, Win 10 Pro 1709 updated, high perf power plan.

Am doing a prime95 small FFT test now on all cores, package power is at ~131W and cores at ~85W, temp 67*C, vcore jumps around 1.243 - 1.286, fan at ~2900 RPM.

Bus and multiplier move around in CPU-Z; 95.xx - 99.xx for bus and 38.5 - 39.25 multiplier.

Core clock is 3892 to 3919, so, above the 3.7 advertised.

Core voltages are 1.188 to 1.200.

I'm guessing AMD is advertising the core usage power and not the package usage power.

I am impressed with this.
 
  • Like
Reactions: lightmanek

Hitman928

Diamond Member
Apr 15, 2012
5,182
7,633
136
I'm guessing AMD is advertising the core usage power and not the package usage power.

The advertised TDP is basically without XFR. In other words, the all core turbo falls to the base frequency (3.7 GHz for the 2700x). With XFR and Precision Boost 2, you'll see higher frequencies and above TDP power. I think Stilt mentioned that by default the power limit is set to ~140W and the CPU can use up to that much assuming you keep it cool enough. I've seen reports that some motherboards also allow you to disable the upper power limit if you want to see if the chip will boost to even higher frequencies.
 
Last edited:

B-Riz

Golden Member
Feb 15, 2011
1,482
612
136
The advertised TDP is basically without XFR. In other words, the all core turbo falls to the base frequency (3.7 GHz or the 2700x). With XFR and Precision Boost 2, you'll see higher frequencies and above TDP power. I think Stilt mentioned that by default the power limit is set to ~140W and the CPU can use up to that much assuming you keep it cool enough. I've seen reports that some motherboards also allow you to disable the upper power limit if you want to see if the chip will boost to even higher frequencies.

Seems legit. :)

HWMonitor does show my max Package power at 139.99W.

Am curious if moving to the H110GTi cooler would make a huge difference or if the Mugen 2 would be just as good...
 

epsilon84

Golden Member
Aug 29, 2010
1,142
927
136
The silicon is not really a dud. The 2700x can do 4.0 Ghz at 90 watts while the 1700 that Gamer's Nexus tested used 144 watts. That is 37% less power on 2000 series vs. 1000 series.

Everyone here is looking at max clocks, but the best improvement is in parts like the vanilla 2700 which should probably still perform on par with an 1800x despite being at only 65 watts.

The new Ryzen silicon is much more efficient that Ryzen 1, however the 2700x is what you get when you try to throw 110% of that extra power efficiency back into clock speed.

FYI, Intel also has crazy exponential power scaling too, however there bad scaling is in the region of 4.5-5.0 ghz. The main problem AMD face is that people have it in their head that 5 ghz is an achievable number and this part simply won't ever hit that number (or even get within 10%).

I agree its not a dud, but its also not that much of an improvement over 1st gen Ryzen either. You can't really compare 4GHz for 1st and 2nd gen Ryzen and conclude its 37% more efficient, because to hit 4GHz on Ryzen 1000 you need a disproportionately high voltage, 4GHz on Ryzen 1000 is equivalent to 4.2GHz on Ryzen 2000, you need a lot of extra voltage to get that last couple hundred MHz.

I think its fair to say that, the 'efficiency curve' of Ryzen 2000 has increased to 4GHz, compared to 3.8GHz on Ryzen 1000. Push past these clocks for the respective chips and power consumption skyrockets.

As you said, Intel also suffers from this too, but with my 8700K for example, it scales pretty well up to 5.0GHz, and can be adequately cooled by my relatively humble CM Hyper 212, so I disagree somewhat that scaling is bad between 4.5 - 5.0GHz. The 'wall' for CFL seems to be between 5.0 - 5.1GHz, where from that point on, like Ryzen 2000 past 4.0 - 4.1GHz, you need to pump a lot of extra voltage for minimal gains.
 
May 11, 2008
19,303
1,130
126
SIDD (Static VDD Power Supply Current)
"SIDD(Temp(SIDD)) – feedback between static leakage and temperature can induce self heating." - Salishan Hi Speed Conference 2007 (April 26, 2007 - V & V @ AMD: Ensuring a Solid HPC Foundation and Data Confidence from Core to Cache and Beyond . . .)

https://i.imgur.com/A752J7G.png
It is related to static leakage.

High SIDD is better because of electron ballistics and Ti-states(XFR/Boost). Hotter the transistor the more it can be overclocked for a given voltage, etc. Complex stuff.
"The key is to counterbalance leakage power increase at higher temperatures with dynamic power reduction by the Ti-states." - Ti-states: processor power management in the temperature inversion region ( Taipei, Taiwan — October 15 - 19, 2016 @ MICRO-49(The 49th Annual IEEE/ACM International Symposium on Microarchitecture)




I wonder if ballistic conduction can happen in a cpu transistor. The paths the electrons travel are relatively short. But it can happen as written in this article :) :

Here is a nice article about a possible new technology for processes lower than 10nm.
This might be the reason why Intel is delaying their 10nm technology.
They want to come out with a similar solution perhaps...
https://www.allaboutcircuits.com/ne...s-could-be-the-next-leap-in-microelectronics/

The size of the latest transistors (as of 2017) is 10nm, which makes transistor features on the order of tens of atoms in size. Such small devices become harder to produce as making such feature sizes relies on more advanced equipment which comes at a cost.

But the problem with small transistors does not end there. When transistor gates become very small (atoms across), quantum effects–which are normally insignificant such as electron tunneling–become apparent and can have detrimental effects.

Take electron tunneling, for instance: If the gate becomes too thin and electrons can tunnel through then charge stored on the gate can be lost which requires the user to replace that lost charge. The result is a transistor that consumes more current which in turn results in more heat dissipation. Individual transistors (even if very leaky) have near immeasurable amounts of current loss and temperature rise but when several billion devices are placed on a single piece of silicon, the effect adds up and becomes a serious problem. But current does not just leak from the gate; current can tunnel from the source to the drain if the two are in close proximity which can impede a transistors ability to control current.



Enter Molybdenum Disulphide
So how can devices become smaller? There are many ideas being developed by many engineers and researchers alike, including the use of diamond, graphene, and even organic compounds.

Recently, a team from Stanford created a 1nm transistor using molybdenum disulfide but is difficult to produce. This work was further advanced by researchers at the IEEE International Electron Devices Meeting in December who created working complex circuits using realistic manufacturing techniques.

However, a group (also from Stanford) demonstrated how 10nm devices can readily be produced using molybdenum disulfide. But what makes the devices made by Stanford impressive is how they were produced and their exhibited characteristics which could make them much faster than silicon-based devices.



molydis.jpeg


MoS2 devices. Image courtesy of Stanford University.


Read More



According to the Stanford researchers, when molybdenum disulfide transistors are on the scale of 10nm, electrons moving between the drain and source begin to exhibit ballistic conduction, which is when electrons stop scattering as they move through the material. This scattering is what causes resistivity in materials—without it, a material essentially has 0 resistance.

Therefore, electrons in molybdenum disulfide transistors could pass through unaffected by the semiconducting material and hence could operate much faster compared to their silicon counterparts. Eric Pop (an EE at Stanford) estimates that one in five electrons undergo ballistic conduction in the 10nm devices. However, Eric Pop believes that, if the quality of the semiconductor material is improved and the transistor reduced in size, this ballistic conduction number will increase. This would result in higher-speed devices which can conduct more current without heating up as much as a silicon device.

The devices, however, are not just small and fast. They could be easily manufactured using similar techniques to silicon devices. Silicon chips are so successful due to their ability to be produced on massive scales using layered techniques whereby each layer is either added or subtracted from a single wafer of silicon (and each wafer can produce hundreds of devices).

Stanford researchers were able to produce the 10nm devices by taking a silicon wafer and growing molybdenum disulfide on top. Then, a gate was grown on top of the semiconductor by first depositing 20nm of aluminum and then allowing it to oxide (where it shrinks to 10nm).


https://en.wikipedia.org/wiki/Ballistic_conduction
 

IRobot23

Senior member
Jul 3, 2017
601
183
76
I have been reading about cores per CCX and how next gen will have 6 cores per CCX and so on. I don't know how would it take so much advantage of it (over 4C per CCX) and why even people want more cores per CCX?

If app knows Ryzen topology I don't see the problem about CCX. Only thing that comes on my mind is space and more power.... Probably some changes would need to be made andprety much some over time would be needed. Wouldn't it be better, if they do more focus on (lowering latency of) Infinity Fabric and ZEN2 core?

I see 4 core per CCX ideal for stacking and so on... each core has 8MB of LLC fast and low latency. Everything else DRAM.

Which program would even make use of it? Unless some games would completely mess with ryzen and would share data across CCXs.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I have been reading about cores per CCX and how next gen will have 6 cores per CCX and so on. I don't know how would it take so much advantage of it (over 4C per CCX) and why even people want more cores per CCX?

Seems more likely that it is still 4 cores per CCX. 4 of them however. Having more cores in a CCX would reduce the amount of cross-CCX traffic.
 

IRobot23

Senior member
Jul 3, 2017
601
183
76
Seems more likely that it is still 4 cores per CCX. 4 of them however. Having more cores in a CCX would reduce the amount of cross-CCX traffic.


As I said... why would you need cross-CCX traffic? You clearly have 8MB of L3 per core.

Intel even decreased size of LLC cache in Skylake X design.
 

mat9v

Member
Mar 17, 2017
25
0
66
As I said... why would you need cross-CCX traffic? You clearly have 8MB of L3 per core.

Intel even decreased size of LLC cache in Skylake X design.
And increased L2 cache sizes - in all, small decrease. You ask, why would I be left with cross-CCX traffic - because some programs (games) use more cores and while Windows is aware of CCX topology, when the game requests 6 threads and they can't be on 4 cores because of load balancing, it transfers 2 threads to another CCX. If the game is not CCX aware then the threads transferred may be the ones that require a lot of communication hence the latency of crossing IF causing problems. 6 cores per CCX would work much better in such cases and remember that there is no reason AMD could not add more L3 cache to compensate for more cores in CCX. In fact there were rumors that 7nm based ZEN will have 64MB of cache per module (and 256MB per CPU in biggest EPYC) so maybe there is something to that.
 

mat9v

Member
Mar 17, 2017
25
0
66
Thats the only issue... I mean R5 1500X has two cores per CCX and there is no difference between 4+0 vs 2+2 (or near zero)
There is no way to test such because bios setting between 2+2 and 4+0 are ignored. You could conceivably test it by forcing program to use specific cores but that would require R7 CPU to even try and the results would be tainted. You can't test 2+2 / 4+0 on 1500X anyway - bad example.
 
Status
Not open for further replies.