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

[bjt2]

You know what I just remembered? Bdw-E has a coldbug, overclocks at best mid 4 GHz on air / water, and 5-5.2 GHz on LN2. Haswell-E does well over 6 GHz on LN2.
As an extreme overclocker, it will be really interesting to see where Zen falls there for me.

Also:
https://edgeup.asus.com/2016/06/17/broadwell-e-overclocking-guide/6/

^^^This is exactly what I was worried about when I was explaining my theory on Zen's turbo clocks that got me huge backlash a few pages back, and this is also the reason you don't see Bdw-E with high turbo clocks.

If Zen gets 4 GHz+ Turbo and performs nearly as well across the board as AMD has shown so far, it'll be a smash hit. If not, and it gets between 3.7-4.0 GHz turbo like I had guessed, it will still be impressive imho.

I know a lot of people here are more concerned about Skylake / Kaby Lake and single thread performance, but even if the Ryzen CPUs can not compete there, this will be a very good processor for AMD in the HEDT space, which they are targeting. (I have no idea why people are claiming that they aren't targeting the HEDT space, it has been stated multiple times that they are.)

/drunkpost

Maybe this is the reason behind the launch of HEDT CPUs first. Maybe the process is not mature yet to have more than 4GHz clocks and so a 4 core would be a disappointment...

 

[cytg111]

I have a feeling ZEN+ will come later with 7nm.

Is AMD not on record stating that 14nm wil be a very long node? I read 5 years+ into that.

 

[KTE]

AMD CodeAnalyst

Haven't used it in years. I'll have to fire up an AMD machine to install it and see how much has changed.

You are basically running code profiling then running the benchmark, stopping the profiling when the render has finished. Having minimal background work. AFAIR the rest of the tab viewd are selectable later, including IPC.

As I've already said when the csv was posted, the branch/misprediction rates were very low even for Deneb. Meaning that particular render is not testing the main, crucial bits where AMD has lagged Intel. Or the type of workloads which traditionally cause delays, long latencies, stalls AKA low IPC.

Any older AMD chip being ran through the profiler would show you the IPC value for that render, and where exactly the chip is struggling (vs another chip).

Sent from HTC 10
(Opinions are own)

 

[realibrad]

It's not but the point is if people aren't willing to spend 700~800$ (let's say 700$ just for the sake of argument) for something that's fairly close to 6900K then what makes you think that 600$ will sway them the other way? Also I do believe that octa core is still pro/gamer territory, although hexa core is mainstream IMO just that Intel doesn't want to sell it as such & would rather waste the die space of an i7 x700K on IGP, so the people looking for high end should be willing to spend that little bit extra considering the cash grab that 6950x or even 6900K is atm.

Indeed a quad core without IGP for ~250$ is perfect at that price, a vast majority of gamers out there have an i7 x700K & it's really a waste of a die with IGP that's barely used, if at all.

What's have you seen stats that say a majority have an i7? What demo is that? As far as I know, most gamers do not us i7s.

 

[jpiniero]

Maybe this is the reason behind the launch of HEDT CPUs first.

It was more about getting something out ASAP. There's obviously more effort that needs to be done to get Raven Ridge completed with the IGP and such.

What's have you seen stats that say a majority have an i7? What demo is that? As far as I know, most gamers do not us i7s.

I haven't seen anything recent, but I can tell you that the i5 is Intel's best selling brand by far. Albeit that includes the entire market (desktops + laptops). Steam Survey says 48% for quad core and 45% for dual.

 

[sirmo]

Is AMD not on record stating that 14nm wil be a very long node? I read 5 years+ into that.

AMD did say they are skipping 10nm, which is understandable. But I think we could see some 7nm Samsung/GloFo products as early as 2019.

 

[jpiniero]

AMD did say they are skipping 10nm, which is understandable. But I think we could see some 7nm Samsung/GloFo products as early as 2019.

It's probably the same deal as 20 nm was... nVidia will skip it too. I think for both of them it's more like end of 2019-2020.

 

[realibrad]

It was more about getting something out ASAP. There's obviously more effort that needs to be done to get Raven Ridge completed with the IGP and such.



I haven't seen anything recent, but I can tell you that the i5 is Intel's best selling brand by far. Albeit that includes the entire market (desktops + laptops). Steam Survey says 48% for quad core and 45% for dual.

That is what I thought. i7s are far behind i5s and i3s in terms of market share.

 

[AtenRa]

Is AMD not on record stating that 14nm wil be a very long node? I read 5 years+ into that.

They will certainly need 7nm in order to sustain or even increase whatever Server market share they will have gained with 14nm ZEN when Intel will transition to 10nm sometime in 2019 for Desktop/Server. They could keep producing 14nm products even if they will introduce new higher perf and higher margin 7nm products, so 14nm could definitely be 4-5 years in use.

 

[jpiniero]

They will certainly need 7nm in order to sustain or even increase whatever Server market share they will have gained with 14nm ZEN when Intel will transition to 10nm sometime in 2019 for Desktop/Server. They could keep producing 14nm products even if they will introduce new higher perf and higher margin 7nm products, so 14nm could definitely be 4-5 years in use.

Desktop is a bit fuzzy but I think Cannonlake Server is probably more like 2H 2018, maybe a bit earlier if you are special. Intel's going to be producing 14 nm products (ie: Coffee Lake) after Cannonlake ships so it shouldn't be surprising if AMD does a similar thing.

 

[krumme]

FO4 metric is a measure of how slow (or fast) a circuit is.

Absolute FO4 metric is the delay of the reference test cirucit on a given process.

All is normalized to the reference circuit. So changing the process, you may only measure the delay of the test circuit and know moreless the delay of all other circuits.

The test circuit is an inverter that feed other 4 identical inverter. Then the delay of the propagation of a switch is measured.

The FO4 numbers that we see around are RELATIVE FO4. If I say FO4 17, i am saying that that circuit has a delay 17 times the reference circuit.

Now let's talk of FO4 applied to pipelined processor.

Each stage implements different logic and of each stage a total FO4 (input of the stage/output of the stage) can be calculated.

A good design requires that all stages have the same or similar FO4, because the max clock is limited by the stage with the highest FO4, because when it ceases to function, all the CPU does not work.

The relation is roughly that the max clock (excluding TDP limits) achievable is inversely proportional to FO4.

How we can estimate FO4? We can only do this relatively. But this is enough.

If i have an x86 design that have X stages, another x86 design with Y stages has roughly X/Y the FO4 of the first, because both are x86 CPUs and do the same work. This is especially true with similar architectures.

INTEL has an advantage on the process, but since Zen managed to equal or slightly beat the clock, we can conclude that has lower FO4.

More stages means more penality for branch mispredition. But as can be seen by a table posted some page ago, on blender the branches are 7% of the instructions and of these only 5% are mispredicted.

So even if we double the stages (and so we can increase the clock), the loss of IPC is minimal, because we double the time only on 5% of 7% of the instructions...

The only problem is that this increase the power consumption, so we must do a balance.

And anyway we can't go too low because there is an overhead of 2-3FO4 to break the stages, so under 10FO4 is not advisable to go...

Thanx. Great. Explains why tight integration between code, uarch, pipeline and process is important.

 

[inf64]

I think Zen comes at a right time for AMD. Intel is in a bit of a slump and we are getting process node improvements at a slower rate due to physical restrictions ( getting below 10nm is VERY hard).

AMD needed this kind of a chip to finally get back part of their server market share and improve their margins. Desktop division needed a core of this capability so they could make an APU that can finally go toe to toe with intel mainstream desktop chips that feature iGPU (i3-i7). It is paramount for AMD to keep improving this core, moving it to 7nm as soon as foundry capability is there. Zen+ looks like a solid 10-15% bump since I think they will go for full 2x 256bit SIMD capability per core which should require full 2x256bit L/S subsystem. Couple that kind of a "refresh" core with Polaris-like iGPU in 2018/19 and AMD will earn a crap load of money in all segments.

 

[jpiniero]

AMD needed this kind of a chip to finally get back part of their server market share and improve their margins. Desktop division needed a core of this capability so they could make an APU that can finally go toe to toe with intel mainstream desktop chips that feature iGPU (i3-i7). It is paramount for AMD to keep improving this core, moving it to 7nm as soon as foundry capability is there. Zen+ looks like a solid 10-15% bump since I think they will go for full 2x 256bit SIMD capability per core which should require full 2x256bit L/S subsystem. Couple that kind of a "refresh" core with Polaris-like iGPU in 2018/19 and AMD will earn a crap load of money in all segments.

It's not that easy really. PC sales are basically in decline and now even non-cloud server sales are declining. Building a competitive product is in a lot of ways the easy part.

 

[sirmo]

It's not that easy really. PC sales are basically in decline and now even non-cloud server sales are declining. Building a competitive product is in a lot of ways the easy part.

I don't think this applies to AMD. For one they have non existent share today in the datacenter. And their share in desktop is abysmal. The market even if it shrunk by say half.. would still be humongous. And AMD can only gain in it not go any lower. Besides gaming PCs are on the rise. They have steadily increased each quarter for the past few years.

AMD is a much much smaller company than Intel.. 5% in datacenter and desktop would double their revenues. And that's certainly achievable provided they have a competitive product which each day it's looking more and more like they have.

PC market is only in decline compared to its hay day.. but it's still an essential market that's not going anywhere. In my opinion tablets don't really replace the PC. I think even tablets have plateaued.

 

[Glo.]

I do not know if this has been posted yet.

Lisa Su have said that all versions of Zen right now have AT LEAST 3.4 GHz. Which means this:

SR3 - at least 3.4 GHz at 65W TDP.
SR5 - at least 3.4 GHz at 95W TDP.
SR7 - at least 3.4 GHz at 95W TDP.

So lets focus on the highest end. If there are two SKUs in SR7 range, it means that slower one will have at least 3.4 GHz. Where I am getting from here?
http://www.bitsandchips.it/english/...speculations-about-zen-after-our-april-s-fool

If 14nm LPP will be good, Zen base frequency (8 cores version, TDP 95W) will be 3.7-3.8 GHz. Turbo 4.1-4.2 GHz. If 14nm LPP will be very good (or if AMD will commercialize a limited edition, like the Athlon Slot A 1GHz during the good old times), Zen will have a base frequency of 4 GHz (Turbo 4.2-4.3 GHz)

IPC at Broadwell Level, with higher core clocks, and lower power consumption, and lower price. That would be complete turnaround. From being behind, to being in front of Intel!

 

[DrMrLordX]

I obviously can't speak for anybody else, but my own beef with moving up to the HEDT isn't so much CPU cost, but the cost of 2011v3 mainboards. That and the fact you need 4 DIMMs for optimal performance (can get away with just two). You can get a perfectly serviceable 1151 board for half of what a decent 2011v3 board will set you back.

Combining an 8C/16T CPU with a mainstream platform might work out very well for AMD.

Was planning to go Skylake-E eventually, but from what I've seen so far I am beginning to revise that notion.

I agree that the LGA2011 v3 motherboards are expensive. Some of them are not even that good! So caveat emptor.

With AM4 I think it may be another minefield ala AM3+ and FM2+. We'll have to be vigilant against poor VRM configurations and/or weak VRMs. Maybe not as bad as past generations but still . . . watch out!

 

[Insert_Nickname]

With AM4 I think it may be another minefield ala AM3+ and FM2+. We'll have to be vigilant against poor VRM configurations and/or weak VRMs. Maybe not as bad as past generations but still . . . watch out!

I'm afraid you may be right. But AM4 does have stricter and tighter VRM specifications then either AM3(+) or FM2(+) ever did. It may still work out, provided you don't expect a cheap AM1 analogue board to handle the top SKUs.

 

[witeken]

by the way--that is just cost per wafer paid to GoFlo, right? I don't imagine that $18/chip or certainly not $24-28/chip reflects the costs of R&D, right? If itemized out, I wonder how much AMD has spent up until this point on each Zen chip--I assume that the costs of FX are (at least slightly!) beyond profit by now, right?

There are so many assumption in that calculation that you have to take those with many grains of salt. The funniest assumption was that of equal yield of 32nm vs. 14nm. And don't forget you have to at least double the price because GloFo also makes a margin on the chips, unless that costs is already factored in in the wafer price, but that is not specified.

And indeed R&D is not included in those costs. For comparison, Intel spends maybe 50% more on R&D than on wafer production, so more of someone's money goes to the development of future products than for the actual production of the product, which is of course necessary for all tech companies to make their business sustainable.

 

[DrMrLordX]

I'm afraid you may be right. But AM4 does have stricter and tighter VRM specifications then either AM3(+) or FM2(+) ever did. It may still work out, provided you don't expect a cheap AM1 analogue board to handle the top SKUs.

Yeah, those B350 boards and lower won't handle the same kind of stress as X370 boards. Common socket is nice, but the only chipset that will be good for "everything" will be X370.

 

[itsmydamnation]

They will do the same they did with Piledriver or what Intel does with Kabylake. They will bring a new SKU with 5% higher IPC and perhaps 100-200MHz higher clocks in 2018 and then in 2019 ZEN+ will arrive at 7nm.

Piledriver was more then that , it was 400mhz to base and more like 7% IPC improvement so ending up with 10-15%. If AMD can do that 12 odd months on from Zen release then things will be interesting!!!

http://www.anandtech.com/bench/product/434?vs=697

Is AMD not on record stating that 14nm wil be a very long node? I read 5 years+ into that.

You will probably find thats for vega/polaris. AMD is only going to target Two GPU releases a year so like Pitcairn and Tahiti there is probably going to be one or two of 4 polaris/vega chips that lives for a long time.

GF 7nm is IBM's 7nm and is well on its way, AMD will want to be on that node as soon as its ready, remember there are rumors of 48 core Zen+ SKU's.

 

[unseenmorbidity]

Yeah, those B350 boards and lower won't handle the same kind of stress as X370 boards. Common socket is nice, but the only chipset that will be good for "everything" will be X370.

Where did you see information on the boards?!

I do not know if this has been posted yet.

Lisa Su have said that all versions of Zen right now have AT LEAST 3.4 GHz. Which means this:

SR3 - at least 3.4 GHz at 65W TDP.
SR5 - at least 3.4 GHz at 95W TDP.
SR7 - at least 3.4 GHz at 95W TDP.

So lets focus on the highest end. If there are two SKUs in SR7 range, it means that slower one will have at least 3.4 GHz. Where I am getting from here?
http://www.bitsandchips.it/english/...speculations-about-zen-after-our-april-s-fool


IPC at Broadwell Level, with higher core clocks, and lower power consumption, and lower price. That would be complete turnaround. From being behind, to being in front of Intel!

We don't know the IPC really. We just know Zen is faster in blender. The power consumption is actually roughly the same, as was demonstrated in the demo.

I am hyped too, but let's not go too crazy just yet.

 

[KTE]

Anyone analyzed AMDs render scene yet?

I am really too short of time, not able to sit at a comp (on the move) or I would delve into it

Sent from HTC 10
(Opinions are own)

 

[DrMrLordX]

Where did you see information on the boards?!

Nothing anyone else hasn't already seen. But B350 has been MIA since AMD announced it, and the smart money's on such boards having 4+2 phase VRM configurations in the majority of cases. Maybe we'll be treated to 8+2 configs here and there using that chipset, but I doubt it.

Regardless I'm just making assumptions.

 

[blublub]

On Realworldtech:
http://www.realworldtech.com/forum/?threadid=163466&curpostid=163575

Blender branch prediction MPKI (misprediction / kilo instr) is not that high. So this benchmark likely does not tell us a lot about branch prediction performance of Ryzen.

Ok, but what about handbrake ?

 

[AtenRa]

This is where our opinions diverge, I do believe that with an octa core ZEN AMD is targeting the HEDT line, for me personally the high end mainstream stops at six core chips. Should AMD introduce a ten core successor to the SR7, it'll just as well be aimed at the enthusiast segment & priced accordingly.

Yes but as I've said around 600$ is where IMO the enthusiast segment starts & I see them happily paying 1000$ for Titan XYZ all the time, should they find the SR7 to their liking a 100$ more (even if it's just the launch price) ought not to be a deal breaker for them.

That they will & so the (IGP less) quad core ZEN should be priced competitively, the APU however could command a premium since AMD is clearly superior in the graphics department.

Ok I will agree that AMD made ZEN to be part of both worlds and what do i mean with that,

Today the 8-Core Core i7 6900K Broadwell-E starts at $1049 but the 6-Core starts at $375 and the second 6-core SKU with higher clocks and more PCI-e (40) lanes at $530.

I believe AMD with a single die (8-core ZEN) will be able to compete in the mainstream market against Core i5 (starts at $180) and be able to sell an enthusiast SKU with 8-Cores and compete against up-to Intels 6-core Core i7 6850K ($530).

Because per core performance will be slower than Intels Kabylake (10-15% ??), they will have to give 2x Threads or 2x Cores more at the same price.

Example,
4C 8T RYZEN CPU to compete against Core i5 from 180$ to $230,
6C 12T RYZEN CPU against Core i7 mainstream up to $330
and then 8C 16T RYZEN to compete against 6-Core Intels HEDT CPUs up to $500.

We also have to see how many PCI-e lanes ZEN has and how much it can OC but the more important thing is to see the performance in GAMES. I mean both handbrake and blender are fine but most people here are Gamers and gaming performance is something we havent seen yet.

And as i have said before, a 65W TDP 4C8T and 6C 12T RYZEN OEM Gaming SFF system with RX 460/470 at lower prices than Intels Core i5/i7 Kabylake + dGPU will be the target.