[ 3 / biz / cgl / ck / diy / fa / g / ic / jp / lit / sci / tg / vr / vt ] [ index / top / reports / report a bug ] [ 4plebs / archived.moe / rbt ]

Due to resource constraints, /g/ and /tg/ will no longer be archived or available. Other archivers continue to archive these boards.Become a Patron!

/g/ - Technology


View post   

[ Toggle deleted replies ]
>> No.75126147 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]


>> No.72054181 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

Heres voltage scaling for Zen 1 for comparison.

>> No.71770240 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

Zen1 voltage curve for comparison

>> No.70583598 [View]
File: 55 KB, 1509x905, Zen_power-efficiency.png [View same] [iqdb] [saucenao] [google] [report]


This is Zen1. Zen2 is a full die shrink over Zen1. If you clocked it at the same clock of the peak efficiency points; you can take the Zen1 power value and half it.

>> No.68389746 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

u wot m7

The voltage scaling of Zen is an artifact of the 14LPP process being design for low power ARM chips. It was never designed for high performance parts, it doesn't have the metal stack for it. The 12nm refresh only marginally improved this with a new transistor library and track. Its still coping with its heritage.
The voltage scaling Zen displays is not part of its architecture. Power delivery is extremely closely regulated, desktop Ryzen has finer granularity on power deliver than an intel Xeon, TheStilt has a whole write up on it. Its just working around nuances of the process.

And AMD is using a HP variant of TSMC's 7nm process. TSMC outlined their development of it in 2015 and 2016.

>> No.68119425 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

A 60% efficiency increase isn't out of the question. A straight 60% power reduction would be possible if Zen were ported to 7nm depending on where on the power curve you're making the comparison.
The stock 1800X drew 110w under a full load FPU torture test. Power consumption there was high because its clocks were just outside of the sweet spot for the 14LPP process.

Yes. You should do it.

>> No.67279318 [View]
File: 55 KB, 1509x905, zen1 voltage curve.png [View same] [iqdb] [saucenao] [google] [report]

yeah, I know
14LPP runs with good efficiency until it gets to 3.3GHz, after that the clock/voltage curve starts getting fucked up

>> No.67187289 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

Architecture plays an enormous role, but we already know some basics about the Zen core design, we know its built to clock in the range of 3ghz to 5ghz. The 14nm LPP parts are only limited by high voltage needed to push above 4ghz, and thats an artifact of the node they're on.

We can't say what the power curve will look like for Zen2, but we can make some reasonable estimates compared to Zen1.
Drive voltage to hit 3.5ghz with Zen2 parts will probably be around .8, scaling from mid 3ghz to mid 4ghz should be fairly flat, with increasingly higher vcores only being needed as it approaches 5ghz.

At launch a decent 1800X sample was pulling about 10w per core at 3.6ghz.
7nm could probe that down to 4w per core. We could have full 1800X performance at about 40w, not to mention the increased IPC.

>> No.67114021 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


>threadripper is straight up double ryzen at same clocks
>shocked it uses more power

You astound me anon. Also: zen is extremely power effecient - particularly when clocked around 3-3.3ghz.

>> No.67040914 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


>> No.66493307 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


Zen is a stupidly effecient design on the power draw front for x86.

>> No.65611194 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

>4ghz at 1.162v
Holy fucking shit. Now thats enormous.
Thats like a 500mhz gain compared to a first gen Ryzen sample.
Energy efficiency there would be insane.

>> No.64996174 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


>> No.64879206 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

If they managed to actually raise fmax that high it would be a huge feat. Zen's current handicap is the voltage scaling.

>> No.64205056 [View]
File: 55 KB, 1509x905, 8Rch6JF.png [View same] [iqdb] [saucenao] [google] [report]

I guess I wasn't clear about what I was referring to.
We all know that Zen's clock-to-voltage-input does not scale linearly once you hit 3.8-3.9 GHz. Once you hit that mark, you will need to increase the core voltage substantially just to sustain higher clocks, and the voltage increase only gets worse per 100MHz increment.

Take a look at this graph, Zen or the 14nm process is the most optimized up to 3.5GHz. Once you get to the desktop Zen's clockspeed range, the rate becomes linear again, but at a higher voltage increase per 100MHz. And then when 3.8-3.9GHz comes around, you have another sharp increase in input voltage needed. Yes, you can hit 5.3GHz on Ryzen, but at absurdly high input voltage. Intel's superior process and optimized architecture allows them to hit the same 5.0GHz range at a much lower input voltage.

That behavior indication that either Zen or the 14nm process was optimized for a much lower clock rate range. It could be that Zen can scale as well as Skylake can in the upper range, but I doubt it. Put Zen on the old 28nm or 32nm process and I doubt it could hold the same clocks as Piledriver could, because the Construction cores were specifically designed to work at higher clock ranges. Zen clearly wasn't. 12nm isn't going to change that behavior, apart from adjusting the ranges of the clock-to-voltage behavior. It won't be any more than a 300MHz increase at the upper range, since 12nm is largely based off of the 14nm LPP (it's not even a 12nm process, just another marketing scheme). Anyone who says otherwise has been drinking the AMD Koolaid for too long.

Proof or GTFO

>> No.64054418 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


Zen gets crazy effecient when you pull the clocks down, but I am unsure if it would still draw too much power in an absolute sense. Consider how power effecient desktop zen is and thats clocked well into diminishing return territory.

>> No.63973594 [View]
File: 55 KB, 1509x905, 8Rch6JF.png [View same] [iqdb] [saucenao] [google] [report]

Pick a point from the pic related graph, dial in the settings and boot. Done

>> No.63684982 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


>> No.63642796 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


>> No.63511788 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


>> No.63412743 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


Technically the 1700 is the exact same silicon as a 1800x. The binning aspect of it is basically because the 1800x are leaky motherfuckers - the 1700 chips are the good silicon based on Zen's fmax.

>> No.63218889 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


Basically less I/O and (most likely) worse vrms. I/O isn't too relevant as Zen is a true system on chip and brings 90% of its I/O with itself so it comes down to vrms and if you are considering A3XX boards you aren't overclocking anything.

IIRC AM4 spec from AMD mandates that any board should be capable of supporting the 1800x without exploding. Given the previously mentioned overclocking related issues as long as you aren't slamming 8c to over 3.6ghz you should be just fine.

There is a reason why the 1800x has base clocks of 3.6ghz and not higher - from that point on you NEED good vrms to handle the load.

>> No.62984980 [View]
File: 55 KB, 1509x905, zen fmax.png [View same] [iqdb] [saucenao] [google] [report]


I know man, flipping burgers is hard work but hey, you could read up on how shit actually works during your 20 minute lunchbreak to further your understanding of the technology we all love and adore!


>they are incredibly power efficient

Not after 3.5ghz they aren't.

>> No.62914646 [View]
File: 55 KB, 1509x905, fmax-vmin.png [View same] [iqdb] [saucenao] [google] [report]

Doesn't change what the nominal vcore range for that frequency is. It isn't high voltage by any stretch.

View posts [+24] [+48] [+96]