Ryzen, Skylake, and everything that's coming next. (MEGA discussion thread)

[tamz_msc]

How does GF 7nm compare to Intel 10nm?

Comparable in density. It's licensed from IBM so details are a bit scarce right now. GF claim that they are exceeding performance expectations on 7nm.

 

[maddie]

Comparable in density. It's licensed from IBM so details are a bit scarce right now. GF claim that they are exceeding performance expectations on 7nm.

Why licensed? I thought the whole IBM division was sold to and is now part of GloFlo?

 

[Tuna-Fish]

Why licensed? I thought the whole IBM division was sold to and is now part of GloFlo?

IBM sold manufacturing, kept research. Interestingly, since the later stages of process research require full-scale manufacturing equipment, IBM research rents time from GloFo fabs to do it.

 

[T1beriu]

It's licensed from IBM

Sorry, but that's not true. Neither IBM or GF have announced this. GF bought IBM's foundry business in 2014 and it's a part of a joint development alliance with IBM and Samsung that developed 7nm and are continuing this venture for 5nm.

GF Press Release on 7nm FinFET.

 

[raghu78]

Things looking very good for 7nm Zen 2 at 5+ Ghz in early 2019. GF 7LP slated for volume production in H2 2018. Features vs 14nm - >40% perf improvement , >2x transistor density increase. 30-45% cost reduction. 5 Ghz High performance version. 7nm+ with EUV in 2019 with products shipping by 2020. AMD and GF are looking to take the competition to the next level against Intel.

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...infet-process-and-fx-7-asic-platform.2508687/

 

[tamz_msc]

Sorry, but that's not true. Neither IBM or GF have announced this. GF bought IBM's foundry business in 2014 and it's a part of a joint development alliance with IBM and Samsung that developed 7nm and are continuing this venture for 5nm.

GF Press Release on 7nm FinFET.

According to SemiWiki, LP or leading performance is IBM terminology. Details are still missing except it being a research collaboration with IBM. Not clear if IBM is involved beyond that.

 

[WhoBeDaPlaya]

AMD marketing is finally getting a little smarter.
X399 > X299 from a naming perspective

 

[A///]

I might end up building a system in early 2018 or mid year. My system setup is about five years old. It isn't slow for what I do. It would be nice to get my hands on newer and more efficient tech.

Right now that 15 core Ryzen and 32 gb of RAM sound tasty.

AMD marketing is finally getting a little smarter.
X399 > X299 from a naming perspective

Someone commented on this on reddit a week ago. How they're trying to get under Intel's skin.

 

[CHADBOGA]

Someone commented on this on reddit a week ago. How they're trying to get under Intel's skin.

People are overthinking things and reading more into this than there is.

It would make more sense to use a numbering scheme to "get one up on Intel" for CPU's, than motherboards where no one really gives a fig.

 

[moonbogg]

That 6 core is going to be $400 isn't it. I might be willing to direct die cool a $300 or less CPU, but not more expensive than that. Its a warranty destroying risk that is being taken and I can't take the risk of having to pay twice for a more expensive CPU because the first one got broken from me trying to fix Intel's garbage TIM. Such a sad state Intel is in right now. I still can't believe they used paste on these CPU's.
I also can't believe they cut down their 8 core chip's lanes. I know why they did it. They are competing with the 1800X and compare the lanes to that chip, so they made it more closely match the 1800X specs. Problem is, 1800X is a mainstream CPU on a mainstream platform. If Intel wants to match the 1800X, they need an 8 core on their mainstream platform. But now they are stuck, so they had to gimp the 8 core chip which forces people to pay at least $1000 for a proper PCI lane count. Ridiculous. I can see AMD just running away with this one.
Also, X399 is a proper name. Its superior to Intel's X299 in every way. All Intel has going for it is more IPC, and that isn't as important for high core count chips. It matters, but much less than for a 6 or 8 core chip.

 

[ZGR]

I don't see it in such a dire state as you, Bogg. I'm glad Intel has moved the prices down an inch and to see the clocks increasing for core and ram.

Having never delidded a CPU before it does seem rather scary, but I'd be willing to do it. I guess I would buy the i5 LGA 2066 as an expensive guinea pig and the best i9 I wanted as well as a good delid kit.

It is great that overclocking has improved so much over the past generations to warrant the need for a delid. It may just be cheaper to buy a chip already delidded and binned for overclocking for those who don't want the hassle.

But if you don't delid and get a mediocre overclock then what? How will that compare to Broadwell-E? It almost seems justifiable to have to delid for those who are on X99 and want to upgrade.

 

[LTC8K6]

That 6 core is going to be $400 isn't it. I might be willing to direct die cool a $300 or less CPU, but not more expensive than that. Its a warranty destroying risk that is being taken and I can't take the risk of having to pay twice for a more expensive CPU because the first one got broken from me trying to fix Intel's garbage TIM. Such a sad state Intel is in right now. I still can't believe they used paste on these CPU's.
I also can't believe they cut down their 8 core chip's lanes. I know why they did it. They are competing with the 1800X and compare the lanes to that chip, so they made it more closely match the 1800X specs. Problem is, 1800X is a mainstream CPU on a mainstream platform. If Intel wants to match the 1800X, they need an 8 core on their mainstream platform. But now they are stuck, so they had to gimp the 8 core chip which forces people to pay at least $1000 for a proper PCI lane count. Ridiculous. I can see AMD just running away with this one.
Also, X399 is a proper name. Its superior to Intel's X299 in every way. All Intel has going for it is more IPC, and that isn't as important for high core count chips. It matters, but much less than for a 6 or 8 core chip.

I doubt that Intel set the number of PCIE lanes for these chips after RyZen was announced.

 

[krumme]

I don't see it in such a dire state as you, Bogg. I'm glad Intel has moved the prices down an inch and to see the clocks increasing for core and ram.

Having never delidded a CPU before it does seem rather scary, but I'd be willing to do it. I guess I would buy the i5 LGA 2066 as an expensive guinea pig and the best i9 I wanted as well as a good delid kit.

It is great that overclocking has improved so much over the past generations to warrant the need for a delid. It may just be cheaper to buy a chip already delidded and binned for overclocking for those who don't want the hassle.

But if you don't delid and get a mediocre overclock then what? How will that compare to Broadwell-E? It almost seems justifiable to have to delid for those who are on X99 and want to upgrade.

What if you delid and is so unfortunate you get a bad oc? On a 600 usd cpu. I dont think its the way to go.

Temps in the 90ties with watercooling or some hefty d15 and that even before some avx loading is a no go.

I would rather take it from the good side then. Dont delid. Oc to 4.1. Have low voltage and super fine efficiency. Bwe was a mess on 4.2 already. Then enjoy what you get instead of the last 10-13% perf. You get far lower price for 8c and improved efficiency vs prior gen. On a cpu that takes on all task very well. I think its a fine and safe package then.

 

[ZGR]

What if you delid and is so unfortunate you get a bad oc? On a 600 usd cpu. I dont think its the way to go.

Temps in the 90ties with watercooling or some hefty d15 and that even before some avx loading is a no go.

I would rather take it from the good side then. Dont delid. Oc to 4.1. Have low voltage and super fine efficiency. Bwe was a mess on 4.2 already. Then enjoy what you get instead of the last 10-13% perf. You get far lower price for 8c and improved efficiency vs prior gen. On a cpu that takes on all task very well. I think its a fine and safe package then.

For me at least, achieving 4.1 GHz would be a massive letdown; with Ryzen on the loose. If I spent that much on X299 I would sure as heck do it for the substantial clock advantage it has over the competition.

I don't value efficiency and performance per watt as much as others so it really comes down to our priorities.

 

[Bouowmx]

4.1 GHz is called "overclock", when maximum Turbo Boost (on any core) of Core i7-7820X and i9-7900X is 4.3 GHz?

 

[Ajay]

4.1 GHz is called "overclock", when maximum Turbo Boost (on any core) of Core i7-7820X and i9-7900X is 4.3 GHz?

In all likelihood, they are going to overclock better than that even without a delid. If it were me, I'd buy SKX from Silicon Lottery, so no chance of screwing it up (if they screw it up, it's on them). Hopefully, if the rumors are correct, we'll see reviews on the 19th and end this pointless back and forth.

 

[moonbogg]

For me at least, achieving 4.1 GHz would be a massive letdown; with Ryzen on the loose. If I spent that much on X299 I would sure as heck do it for the substantial clock advantage it has over the competition.

I don't value efficiency and performance per watt as much as others so it really comes down to our priorities.

Yeah, Skylake IPC vs Broadwell isn't much difference at all, certainly not worth upgrading to. Its all about OC performance for those like me who want 6 or 8 cores max. People looking for 10+ cores may have different priorities and expected workloads, so the core increase is plenty reason for them to upgrade. If I still had my 3930k then I'd just grab a SkylakeX 8 core and be done with it, but I don't have that chip anymore. I have something faster already, so the only thing that will make me happy enough is a high clock speed at a good temp and so far its not looking very good in that area.
And what about Coffee lake? That might be a good 6 core option to delid, but what will they do with the PCI lanes? I have an X16 GPU and a PCI based sound card. As soon as I plug the sound card into a mainstream board, my GPU will run at 8x speeds. That's ridiculous. Hopefully coffee lake has more lanes than the pathetic 16 Intel currently offers.

In any case, I owe AMD a huge round of applause. They are in it to make money and NOT to make me happy, but because of them we have insane stuff coming that we never thought we'd see. 16 core CPUs? That's crazy as hell. Who would have predicted that just a year ago when 8 core chips were the big dogs at over $1000? Whenever my time to upgrade again arrives, there will be some serious horsepower to be had thanks to all this.

 

[Bouowmx]

The expansions slots of a typical Intel Z chipset ATX motherboard listed below. Why CPU-connected slot 5 for a ×1 sound card? Why not a PCH-connected slot, like 1, 4, 6, or 7 to keep slot 2 at 16, if you're concerned?

1|-    1
2|---- 16 CPU (8 electrical if slot 5 is used and connected to CPU)
3|     0
4|-    1
5|---- 16 (8 electrical) CPU *
6|-    1
7|---- 16 (4 electrical)
* Only some Z chipset motherboards connect this slot to CPU; otherwise slot is 16 (4 electrical) to PCH.

Intel mainstream having 16 PCI Express lanes, I don't see how Coffee Lake will change that value.

 

[krumme]

What do you guys keep your cpu temp at?

I am a bit noise sensitive so my fan profile is lowest rpm at 60c but goes up steep to max at 70c.
I never get above that temp.

I am a fan of lean cpu and loved core duo, penryn and always found ib and bwe to be fine pieces. I think the 1700 is a fantastic piece of engineering with its perf/w and the thought of running a 7820x at stock would not frighten me.

Pretty excellent times if efficiency is what rocks your boat.

 

[moonbogg]

So threadripper supports ECC memory? SkylakeX doesn't, right? Isn't that kind of a big deal for workstations?

 

[imported_ats]

So threadripper supports ECC memory? SkylakeX doesn't, right? Isn't that kind of a big deal for workstations?

Unless threadripper has official support, it isn't. Actual workstations that care about ECC are going to buy hardware that has official support for ECC.

 

[tential]

Since I have no idea what thread we're allowed to post in that we can speak about both companies, I'll use this again.

https://www.reddit.com/r/Amd/comments/6lufpk/finally_my_brother_can_play_dota_with_me_side_by/

Ryzen 7 can play multiple setups. This is why I think it's stupid to compare Ryzen to an i7 quadcore.

Ryzen can do so much more...

 

[krumme]

Remember to read AT cpu guide for q2
http://www.anandtech.com/show/9793/best-cpus

 

[DrMrLordX]

Hi! Someone brought up the possibility of "phantom throttling" with Zen/Summit Ridge in one of the Skylake-X threads. Since I didn't want to pollute those threads with any more Summit Ridge posts, I decided to start poking around with performance, power draw numbers, and other things using various SIMD ISAs so that I could post results in a platform-agnostic thread like this one. y-cruncher to the rescue!

Current system specs: R7 1800x @ 4.0 GHz/1.40v vcore + Level 3 LLC, DDR4-3466 14-14-14-28 1T, um yeah there you have it.

This chip as it is currently configured is HOT. I had to run up a lot of voltages to get the RAM stable, and then I had to punch the vcore up to 1.4v with Level 3 LLC to get it stable in the most demanding benchmarks, like Prime95 Blend and y-cruncher (Yukina). I can run everything else with .05v less, but I insist on absolute stability. And I'm a bit crazy. But what do I care?

Anyway here are some of the current numbers yielded by y-cruncher:

Running the SSE 4.1 benchmark designed for Nehalem (Ushio), I was able to complete the 1 billion place Pi benchmark in ~128 seconds. I finally found the proper power readout for my CPU in HWiNFO64, which is CPU+SoC Power (SVI2 TFN). During this bench, I ran a whopping 250W! YOW! On the plus side my D15S is keeping it cool! With fans so loud that they might make your ears bleed. It's enough to bring a tear to your eye.

Running the "ADX"/x64 Zen benchmark designed for Summit Ridge (Yukina), SVI2 TFN never went over 220W, the chip ran cooler, and I completed the bench in ~98 seconds.

wat

Need moar dataz. Gonna try some of the other ISA extensions to see what happens.

edit: ran the Sandy Bridge binary that uses AVX (Airi). Power usage hit maybe 230W, was inconsistent, and split the difference temp-wise between the hot SSE4.1 build and the ADX build. Benchmark was completed in . . . .138 seconds? Lame. Gonna try the older stuff now.

edit edit: ran the A64 binary that uses SSE3 (Kasumi). Power usage sat around 220W but bounced around a lot. Temp-wise it ran maybe 1-2C cooler than the ADX build. Benchmark completion time: ~168 seconds.

I have decided not to test the binaries aimed at older systems since the SSE3 binary is the first one to show the expected trend of lower performance and lower power draw/less heat.

So that's . . . um . . . interesting!

 

[tamz_msc]

Hi! Someone brought up the possibility of "phantom throttling" with Zen/Summit Ridge in one of the Skylake-X threads. Since I didn't want to pollute those threads with any more Summit Ridge posts, I decided to start poking around with performance, power draw numbers, and other things using various SIMD ISAs so that I could post results in a platform-agnostic thread like this one. y-cruncher to the rescue!

Current system specs: R7 1800x @ 4.0 GHz/1.40v vcore + Level 3 LLC, DDR4-3466 14-14-14-28 1T, um yeah there you have it.

This chip as it is currently configured is HOT. I had to run up a lot of voltages to get the RAM stable, and then I had to punch the vcore up to 1.4v with Level 3 LLC to get it stable in the most demanding benchmarks, like Prime95 Blend and y-cruncher (Yukina). I can run everything else with .05v less, but I insist on absolute stability. And I'm a bit crazy. But what do I care?

Anyway here are some of the current numbers yielded by y-cruncher:

Running the SSE 4.1 benchmark designed for Nehalem (Ushio), I was able to complete the 1 billion place Pi benchmark in ~128 seconds. I finally found the proper power readout for my CPU in HWiNFO64, which is CPU+SoC Power (SVI2 TFN). During this bench, I ran a whopping 250W! YOW! On the plus side my D15S is keeping it cool! With fans so loud that they might make your ears bleed. It's enough to bring a tear to your eye.

Running the "ADX"/x64 Zen benchmark designed for Summit Ridge (Yukina), SVI2 TFN never went over 220W, the chip ran cooler, and I completed the bench in ~98 seconds.

wat

Need moar dataz. Gonna try some of the other ISA extensions to see what happens.

edit: ran the Sandy Bridge binary that uses AVX (Airi). Power usage hit maybe 230W, was inconsistent, and split the difference temp-wise between the hot SSE4.1 build and the ADX build. Benchmark was completed in . . . .138 seconds? Lame. Gonna try the older stuff now.

edit edit: ran the A64 binary that uses SSE3 (Kasumi). Power usage sat around 220W but bounced around a lot. Temp-wise it ran maybe 1-2C cooler than the ADX build. Benchmark completion time: ~168 seconds.

I have decided not to test the binaries aimed at older systems since the SSE3 binary is the first one to show the expected trend of lower performance and lower power draw/less heat.

So that's . . . um . . . interesting!

So you think that y-cruncher might be the best tool to stress-test Zen at the moment?