AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

[MajinCry]

If this comparison refers to the top Ryzen SKU, it would be very disappointing. If AMD needs a 8C/16T chip to catch up to a 4C/8T chip from 2011, they would still be many years behind.

Unless it's single thread. I've been posting this graph around a bunch, but it's pretty relevant.

In other words, a highly clocked Sandybridge can keep up with Skylake. Oh my.

 

[beginner99]

Some dude over at hardocp forum claim he got some small insider info that it takes a highly overclocked i7 2600k to keep up with ryzen and thats its IPC is good. There are some other shortcomings that apparently will be adressed in zen+.

Also Motherboards have bugs right now but will be fixed with bios updates.

What shortcomings? On the other hand this is nothing new. Wait a couple of month after release for the early adopters to find all the issues and then buy when they are fixed. Basically the go-to way nowadays also with software (games).

 

[LTC8K6]

Unless it's single thread. I've been posting this graph around a bunch, but it's pretty relevant.

In other words, a highly clocked Sandybridge can keep up with Skylake. Oh my.

Well, the article that goes with the picture says otherwise:

From a clock-to-clock performance perspective, Skylake gives an average ~25% better performance in CPU based benchmarks, and when running both generations of processors at their stock speeds that increase jumps up to 37%. In specific tests, it is even higher. When you scale up to a 4.5 GHz Skylake against a 4.7 GHz Sandy Bridge, the 4% frequency difference is only a tiny portion of that. There are other added benefits, such as the move to a DDR4 memory topology that has denser memory modules, as well as PCIe storage and even PCIe 3.0 graphics connectivity.

SB would apparently have to clock 25% higher than SL to match SL.
And they didn't even use fast DDR4 with Skylake.

 

[krumme]

Well, the article that goes with the picture says otherwise:

SB would apparently have to clock 25% higher than SL to match SL.
And they didn't even use fast DDR4 with Skylake.

Naa Go read again. The 25% was a friendly interpretation of their own numbers using ddr4 2133 vs ddr3 1833 with the sb. Actually it was perfect 23%. But i guess the ~25% is the discount favor you give to have a good relationship instead of just writing 23% flat out. Doesnt looks so good.
Keeping it on specs 2667 wouldnt hardly/or just reach those 25%.
A sb 5ghz matches aprox a 4ghz skl for perf st. i think we agree. But so what? 2600k wasnt running anywhere near 5ghz out the gate. Those comparison of oc vs non oc have imo always been stupid and seems always to have some agenda.
Like comparing an oc skl i3 vs an old non oc 2600k as a prime example of such crap (or vice versa). Everyone including reviewers seems to do this discipline. Ohh look this processor is so fast. There should be laws against such skewed comparisons.

(edit: fast memory gives better results especially the latency numbers. But its not only skl benefiting from it. The data we have shows the similar progress for sb as skl. The idea that skl is special here is unfounded - but as its especially at stock freq a good deal faster it needs faster memory to get the most of it)

 

[KTE]

Lol, total non sense and an aknowledgement that you dont even know what you are saying...

Leakage is proportional to voltage threshold, the lowest the threashold the higher the leaking current, that is, the device is closer to full conduction (when switched off) and that s why it will switch faster once driven by the full rail voltage.

From 0V up to the Vth the curent in the device increase quasi exponentialy (actually a polynomial of degree 4-5), the exact leakage is characterised by this subthreshold slope wich is measured in mV/decade, that is, the voltage delta that induce a current 10x bigger, the steeper the slope the lower the leakage knowing that there s no possible miracle in physics.

At Vth and above the transistor conduction (in function of the drive voltage) has converged to a square law out of the previous exponential law..

Strawman.

None of this is new to me, Mr Transconductance. Because you may as well talk about the science of how the Stars of Taurus align with your fanaticism.

You simply miss the point that Vt isn't changed on the same chips unless a company becomes desperate. But you again, have no idea.

Please do keep reading your textbooks.

Sent from HTC 10
(Opinions are own)

Insulting other members is not allowed.
Markfw
Anandtech Moderator

 

[IllogicalGlory]

Well, the article that goes with the picture says otherwise:

SB would apparently have to clock 25% higher than SL to match SL.
And they didn't even use fast DDR4 with Skylake.

The report in question is about a comparison, referring to a "highly-clocked" 2600K (clock speed could be anything) against some Ryzen chip (clock speed could be anything) in some test (could be one where Skylake is way ahead, could be one that doesn't care much, could be single-thread, could be multi-thread). Doesn't seem like there's anything worth discussing. The only third-party benchmarks we've seen so far that are approaching reliable are the original CanardPC ES benchmarks as far as I can tell.

 

[Atari2600]

You simply miss the point that Vt isn't changed on the same chips unless a company becomes desperate. But you again, have no idea.

Please do keep reading your textbooks.

Sent from HTC 10
(Opinions are worthless)

1. If you'd bother to read the posts, you'd see that Vt varies within a chip - each transistor having different Vt values. The longer your gate length, the more sensitive delay is to the variance in Vt of each transistor within a chip.

Thus, the longer your gate length on a design, the greater the chance of a bad spot really f**king over your delay speeds or conversely a bad spot really f**king over your power consumption - depending on what your goal is.

2. You said:

Vt has nothing to do with high/low leakage chips.

That opinion is analogous to saying the earth is flat.

 

[blublub]

I still hope that the naming scheme is not correct and will change but the more pressing thing is the all core turbo

Over at SA people state that the SKU with 3.6/4.0 GHz clock the 4.0ghz will be single core turbo only.

Is there any indication that this might actually be true ? That would suck big time because what's the the 3.6ghz then - base clock or all core turbo? - and in the case if the latter what the base clock then?

I had the impression from the horizon event and CES that 3.6 us base clock and 4.0 all core turbo with single core turbo not disclosed yet

Anyone with more insight on this matter?

 

[bjt2]

I read somewhere (maybe on CPC hardware twitter, I don't remember) that 4Ghz is maximum frequency boost on all core and that mean frequency boost on all core is between 3.65 and 3.8, depending on silicon quality and load and that AMD is deciding the exact value to advertise it...

 

[Atari2600]

Over at SA people state that the SKU with 3.6/4.0 GHz clock the 4.0ghz will be single core turbo only.

Eh?

Not the case.

 

[bjt2]

Look carefully this picture... Nothing worth to note?

RVdd=0.9V. Ryzen can have Vdd as low as 0.9V.

Now look at this image:

14nm FF has +80% performance than 28nm at 0.9V (red curve: 9T. 10.5T not shown, but presumabily better. We don't know if Ryzen is made with 9T or 10.5T, but certainly not 7.5T)
But what is the performance of the 28nm at 0.9V?
The FX9800P has a default Vcore of 1.05V for 2.7GHz.
Let's assume that at 0.9V it can do 2.2GHz (i have found older APU with a P-state of 0.925V and 2.3Ghz): +80% is 3.96GHz.
Call me crazy.

 

[bjt2]

Moreover, if the APU draw 15W@2.7GHz and 1.05V, not accounting the reduction in leakage, at 2.2GHz and 0.9V should draw less than 10W: 15*2.2/2.5*(.9/1.05)^2=9.6980W
Now from the graph 4 XV cores at 3.96GHz on the 14nm, with 0.9V vcore should draw 9.6980*2.4/1.5=15.5W. If 4 XV core draw like 2 Zen core, then 8 zen core should draw 62W.
Probabily since the threads are double and there are other things in the SoC... In 95W we should make it.

The only problem is to be sure of APU clock at .9V. I can't find online undervolt reviews...

 

[Abwx]

The FX9800P has a default Vcore of 1.05V for 2.7GHz.
Let's assume that at 0.9V it can do 2.2GHz: +80% is 3.96GHz.

If 2.7GHz@1.050V then it cant be more than 1.98GHz@0.9V, this yield 3.56GHz, still an impressive number, though..

The only problem is to be sure of APU clock at .9V. I can't find online undervolt reviews...

This wouldnt be much indicative due to the relatively high VTh of the used process, this shift the frequency/voltage curve exageratly at the lowest voltages.

Not so much with finfets because they have lower VTh as they do not need as high values than planar
to keep leakage very low..

 

[CentroX]

The report in question is about a comparison, referring to a "highly-clocked" 2600K (clock speed could be anything) against some Ryzen chip (clock speed could be anything) in some test (could be one where Skylake is way ahead, could be one that doesn't care much, could be single-thread, could be multi-thread). Doesn't seem like there's anything worth discussing. The only third-party benchmarks we've seen so far that are approaching reliable are the original CanardPC ES benchmarks as far as I can tell.

Highly clocked makes me think it is 5 GHz. And if ryzen was 4ghz, then it is Impressive IPC from ryzen.

 

[bjt2]

If 2.7GHz@1.050V then it cant be more than 1.98GHz@0.9V, this yield 3.56GHz, still an impressive number, though..



This wouldnt be much indicative due to the relatively high VTh of the used process, this shift the fequency/voltage curve exageratly at the lowest voltages.

Not so much with finfets because they have lower VTh as they do not need as high values than planar
to keep leakage very low..

I have found an older APU with a Pstate of 2.3GHz@0.925V
EDIT: and the CPUZ screen is 1.044V for 2.7GHz. I rounded a bit...

 

[krumme]

Well as bjt point out here
https://forums.anandtech.com/index.php?posts/38729698
The weak info we have actually points more to lower fo4 than excavator and towards 10.5t than 9t libraries.
Do we have more info about 10.5t and zen?
The 14lpp 9t uplift vs 28nm is certainly impressive on paper as it is.
Now when the sh.t hits the fan things always change, but certainly if its not realized with zen then zen plus tock will go a long stretch to realize it.
I wouldnt dare say its crazy any more

 

[krumme]

I would say its good arguments bjt. The impossible could be true and we could be looking at base freq in the sku leaks then or what?

 

[LTC8K6]

With SL at 4.6, SB would need to hit ~5.6 to keep up. That's 23%.

So I don't think a highly clocked SB can keep up with a highly clocked SL or KL.

I have no idea where RZ would fit in there, but it seems unlikely that it would be at SB level.

 

[bjt2]

I would say its good arguments bjt. The impossible could be true and we could be looking at base freq in the sku leaks then or what?

I hope, but on all forums the doom and gloom make me be a little conservative... Anyway now i am more convinced of the possibility of a 4GHz base at 95W. The 8c@65W can be 3.2-3.3Ghz base or even 3.4... If we stick at Abwx clocks, 3.56GHz with 62W consumption... The 65W can be even 3.6GHz...

 

[bjt2]

If 2.7GHz@1.050V then it cant be more than 1.98GHz@0.9V, this yield 3.56GHz, still an impressive number, though..



This wouldnt be much indicative due to the relatively high VTh of the used process, this shift the frequency/voltage curve exageratly at the lowest voltages.

Not so much with finfets because they have lower VTh as they do not need as high values than planar
to keep leakage very low..

1.044V for 2.7GHz does not seem an high voltage for me... The 15W APU are only a selection, not a new architecture... The LVT, RVT and HVT transistor mix should be the same of the higher power parts...

 

[LTC8K6]

I hope, but on all forums the doom and gloom make me be a little conservative... Anyway now i am more convinced of the possibility of a 4GHz base at 95W. The 8c@65W can be 3.2-3.3Ghz base or even 3.4... If we stick at Abwx clocks, 3.56GHz with 62W consumption... The 65W can be even 3.6GHz...

But CPC is sure that there is no 65W 8C chip...

 

[cytg111]

Strawman.

None of this is new to me, Mr Transconductance. Because you may as well talk about the science of how the Stars of Taurus align with your fanaticism.

You simply miss the point that Vt isn't changed on the same chips unless a company becomes desperate. But you again, have no idea.

Please do keep reading your textbooks.

Sent from HTC 10
(Opinions are own)

Let me understand this right, you guys have found a detail in a detail about Vt in a chip to become mortal enemies over?

 

[Glo.]

Look carefully this picture... Nothing worth to note?

RVdd=0.9V. Ryzen can have Vdd as low as 0.9V.

Now look at this image:

14nm FF has +80% performance than 28nm at 0.9V (red curve: 9T. 10.5T not shown, but presumabily better. We don't know if Ryzen is made with 9T or 10.5T, but certainly not 7.5T)
But what is the performance of the 28nm at 0.9V?
The FX9800P has a default Vcore of 1.05V for 2.7GHz.
Let's assume that at 0.9V it can do 2.2GHz (i have found older APU with a P-state of 0.925V and 2.3Ghz): +80% is 3.96GHz.
Call me crazy.

Well done analysis. It is interesting.

 

[bjt2]

But CPC is sure that there is no 65W 8C chip...

He has not seen it. But if that reseller have actual informations, then it should exist.

 

[coercitiv]

I hope, but on all forums the doom and gloom make me be a little conservative...

What doom and gloom? All I hear about Zen is solid clocks for an 8c part (3.8Ghz+ for MT loads) plus accurate and flexible power management. Except for power consumption at high clocks (4Ghz+) which many prefer to approach in a more cautious manner, Ryzen looks like a great CPU. Not ok, not good enough, but great. In marketing speech that would translate to freaking spectacular.

Even if the top 3.6/4Ghz part ends up being a ~120W TDP part like CPC Hardware suggests, the only fault Zen appears to have is obeying the laws of physics. At this point all I'm interested to find out about Zen is pricing and performance consistency. Frequency and power are no longer a curiosity, except maybe in relation to max oc numbers.