Quantcast
[ 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 ]
File: 208 KB, 2256x1267, AMD-40-IPC-Zen-Zen-.jpg [View same] [iqdb] [saucenao] [google] [report]
62298939 No.62298939 [Reply] [Original] [archived.moe] [rbt]

What we know:

More IPC
Higher clocks
More cores.
7nm LP (IBM node Glofo now has after being gifted IBM's fab business, and engineers, also targets 5GHz at iso power)


Optimal uber awesme cool Intel is literally fucking finished situation:
5.0GHz+
6 core CCX
IPC up by 10-20%
Memory controller latency down by >20ns
Memory controller support for >4000MHz
768KB-1MB L2 cache
3-4MB L3 per core
64KB L1D Cache
64KB cache lanes
IF decoupled from DRAM
1+ AGU
192 Integer registers
256bit AVX2 pipes
30%+ SMT Yield
32 PCIe lanes
Integrated MAC


Not in dreamland:

4.8GHz
6 core CCX
IPC up by 10%
Memory controller latency down by 10%
Memory controller support for >4000MHz
Caches stay same. 32B/512KB/2MB per core.
IF stays the same but lower latency and increased bandwidth.
1+ AGU
More Integer registers
256bit AVX2 pipes
SMT unchanged
28 PCIe lanes


Shit god damn it AMD you had one job(absolute worst that can happen)
3x 4 core CCX
4.5GHz
IPC up by 10%
I/O stays the same
IF only slightly improved
Still 128bit AVX2 pipes
Memory latency increases due to extra CCX

https://en.wikichip.org/wiki/amd/microarchitectures/zen
https://en.wikichip.org/wiki/amd/microarchitectures/zen_2


Discuss, what would you like to see from the 7nm Zen core?

>> No.62298954

AYYMD HOUSEFIRES 180W

>> No.62298967

This shouldn't need mention but please do report shitposters.

>> No.62299616

bump

>> No.62299687

>>62298939
Stop living on predictions and live the present.

>> No.62299785

>>62298939

Here is what should happen.

They are targeting 5.0 at normal use, while the current targets 3.0

we will likely see 8 core laptops with 4ghz never throttle, we may see amd go conservitive with its cpus to make them look even better when compared to intel, imagine amd putting out a 4ghz 8 core at 20-40 watts, but can be "oc'ed" to 5.5 on even the shit binned dies.

the infinity fabric will likely be the same as threadripper and epyc where instead of half ram its 1:1 possibly faster but not decoupled.

The 6 core ccx sounds like it will happen because I believe there is a 48core cpu in the pipeline, and the only way that happens is with 6 core ccx, or I misunderstood the release and they have 4 12 core cpus

Ipc wont get a boost architecturally, or it will be minor when compared to everything else that boosts it.

amd I believe is putting in a 256bit avx pipeline and that was a plan for zen2 so they can keep up.

Outside of that nothing else is known.

>> No.62299806

>coffee lake prediction
>i9 fiasco but for less money.

>> No.62300232

>only 10% more IPC
>when AMD is 30% slower than Intel

LEL

>> No.62300246

>>62300232
>>when AMD is 30% slower than Intel
Bullshit

>> No.62300252
File: 744 KB, 750x1268, 1499866130589.png [View same] [iqdb] [saucenao] [google] [report]
62300252

>>62300232
Hello Brian.

>> No.62300339

>>62298939
>7nm
That means late 2018 at the earliest but with typical AMD delays it's going to be more mid 2019.

I would love to see a clockless x86 CPU but that will likely never happen because muh GHZ.

>> No.62300357

>>62300339
Acktually, 7nm LP was moved a quarter earlier.
So it's Q4 2018 for mobile parts at worst.

>> No.62300404

>>62300339
>clockless
Are you retarded, there's not a single integrated circuit that doesn't run at a fixed frequency in one way or another.

>> No.62300426
File: 952 KB, 1920x2160, 10 cores uses more power and still loses lol.png [View same] [iqdb] [saucenao] [google] [report]
62300426

>>62298954
Kill yourself, housefires spammer.

>> No.62300431

>>62300404
https://en.wikipedia.org/wiki/Asynchronous_circuit#Asynchronous_CPU

>> No.62300456

Raven Ridge fucking where?

>> No.62300458

>>62300426
>12 cores eating less power than 6 cores
What in the fuck.

>> No.62300480

>>62300456
Q4 2018 for the chosen few.
ThinkPads A275 and A475 are Q4 2018.
What matters is how fast can pajeets fix Vega.

>> No.62300693

>>62298939
if being decoupled not going to happen this is suicide
if will probably get wider enough to be used on inter caches
reaching 5.0ghz casually wont happen they arent targeting to meet every damn node properties
avx2 meme dont have enough traction to even care about it so much

also we have to understand that zen2 needs to work on the currect am4 platform so such huge changes wont happen

>> No.62300746

I think if they went with 6 cores per CCX it would be more of a cost saving measure, to reduce the amount of processor dies needed per chip. The current 2xCCX configuration uses a lot of processor dies but not all processors need that much. If they went with 6 cores they could reduce the amount of MCM chips to just the 8+ core models.

>> No.62300775

>>62300746
>The current 2xCCX configuration uses a lot of processor dies
What?
One Zeppelin containes two CCX, and yields on those are 80%+.

>> No.62300871

>>62300775
I got mixed up and forgot the CCXs aren't really separate dies in MCM they're just not directly connected to one another.

I was thinking rather than using two CCXs to cover 4 core and 6 core models they could just use one to cover those packages and then only go to two when they need more cores.

>> No.62300894

>>62300871
Why would they waste wafers on that?
One 12 core die for all.

>> No.62300978

>>62300232
AMD actually has higher IPC than Intel. It's the clock speed and unoptimized software that keeps them under.

>> No.62300986

>>62300894
Well my thinking was that the current dies are 2xCCX with 8 cores so 1xCCX with 6 cores would be smaller (in theory) and then there's the improved 7nm process which would reduce it further.

I don't really know if any of that actually makes sense from AMD's perspective though.

>> No.62301003

>>62298939
>yfw GloFo manages to do 6.1ghz chips

>> No.62301017

>>62300357
That's too optimistic. No way AMD will have Zen 2 out that fast.

>> No.62301024

>>62300480
However long it takes them to do a complete respin

>> No.62301049

>>62300480
Lies. They are supposed to roll out Q4-2017
Hell, Asus even showed a "sneak peek" at their new RR laptop and in this day and age, product launch doesn't take that long. (Given that the architecture is already available)

>> No.62301184

>>62301024
Vega needs no respins.
It needs software.

>> No.62301202

>>62301017
Totally, rabbi!

>> No.62301204

I'm expecting absolute worst that can happen. It is AMD after all

>> No.62301232

>>62301204
Sorry it's AMD that has competent management right now.

>> No.62301278

>>62301232
tell that to vega

>> No.62301289
File: 98 KB, 361x453, UMA_MACACO_STRIKE_AGAIN.png [View same] [iqdb] [saucenao] [google] [report]
62301289

>>62298954
BRAAZILIAN SHITPOSTING 180W

>> No.62301311

>>62301202
p-pls no rajeesh

m-my sheckles

>> No.62301316

>>62301278
There's nothing wrong with Vega (besides AMD giving ~0 fucks abou t gaymen market).

>> No.62301346

>>62298939
5.0ghz is happening - it's on GloFlo's roadmap for 7LP.
14LPP (Ryzen Process) was listed as 3.0ghz (which also corresponds with Ryzen's peak performance/watt)

>> No.62301404
File: 152 KB, 1484x1092, frustrated.png [View same] [iqdb] [saucenao] [google] [report]
62301404

>>62301003
>mfw anudda shoa

>> No.62301419

>>62298939
>32 PCIe lanes
>28 PCIe lanes
Not possible. AM4 is set in stone until 2020.

>> No.62301519

>>62298939
i think zen 2 is not the next ryzen.
Zen 2 will have 7nm, 6 core ccx, 4.5-5ghz. at least 10% IPC per clock.
however after i core 8000 , amd will answer with ryzen 2000,
>14nm+ nodes, prolly reaching 4,3ghz without oc.
>small refinements, keeping the power consumption at identical levels.
>small architectural tweaks, giving a 3-5% ipc per clock
they will launch a bit after/ at the same time with raven ridge. Raven ridge will be 4c8t with igpus, mostly for laptops and oems.

Zen 2 will be ryzen 3000.

>> No.62301526

Should I wait for Zen2 or just get a Ryzen 5 1500x? I have a Pentium G3470 right now.

I'm still slightly butthurt how AMD can't match intel in single threaded benchmarks.

>> No.62301619

>>62301419
Zepplin is already 32lanes at a die level. There was a road map somewhere saying that there will be half a snowy owl SOC at some point which should have the full 32lanes, just not on gamer shit. I'm interested they have raven ridge at 24lanes for am4, or 16lanes like Bristol ridge or keep 32 which seems useless for RR's markets unless they're making a server branded version.

>> No.62301766

Clock speeds aren't solely a product of process node characteristics, you're also limited by the segment of the pipeline with the highest latency. Optimizing the Zen architecture alone could probably also yield decent clock increases.

AMD doesn't need to add the extra clock domains and buffers necessary to decouple IF from the memory controller, when they can just double IF bandwidth and cut latency by half. Most of the latency is transmission delay, anyways.

>> No.62301808

>>62301184
We don't know that yet anon.
There might need to be a new stepping for some of the shit to work.
It might also be software.

Regardless, you shouldn't buy Vega on promises of possible later performance - Vega 64 is meh.
Buy Vega56 if you're looking at a 1070 otherwise and you can buy one.

>> No.62301822

>>62301808
>We don't know that yet anon.
We know exactly that.
https://forum.beyond3d.com/posts/1997699/

>> No.62301823

>>62301619
Doesn't matter what it has at a die level, the socket can't expose more than 24lanes.
The socket is set in stone until 2020.

If you need more lanes, buy TR.

>> No.62301843

>>62298939
>that graph scaling
It's OK when AMD does it.

>> No.62301905

>>62300339
Anon, x86 is built around instruction latency, which you guessed it: is the the most simple 2 clock cycle ADD

>> No.62302272
File: 958 KB, 1000x1000, cover.jpg [View same] [iqdb] [saucenao] [google] [report]
62302272

TFW 1080p 120hz Vidya player so can't enjoy awesome multithreaded efficient cpu from AMD.

>> No.62302430

>>62298939

Probably will still have shit memory support.

Zen never again, should have bought an I5.

>> No.62303702

>>62302430
>implying a year later and memory situation won't improve
>implying it hasn't improved already
>implying anyone would buy an i5 over ryzen

>> No.62303853 [DELETED] 

>>62298939
Discuss, what would you like to see from the 7nm Zen core?
I honestly don't care what the top end is going to be. I'm still using a processor from 2008 and a midrange GPU from 2011. I'm more interested in what will be possible at lower power and what could be achievable by an SBC using a processor manufactured on the new 7nm process that Zen 2 is going to use. I would like a desktop in a passively cooled pizzabox form factor thank you very much.

>> No.62303875

>>62298939
>Discuss, what would you like to see from the 7nm Zen core?
I honestly don't care what the top end is going to be. I'm still using a processor from 2008 and a midrange GPU from 2011. I'm more interested in what will be possible at lower power and what could be achievable by an SBC using a processor manufactured on the new 7nm process that Zen 2 is going to use. I would like a desktop in a passively cooled pizzabox form factor thank you very much.

>> No.62304271

>>62303875
American size pizzabox or normal one?

>> No.62304451

>>62304271
The comfy one:
https://en.wikipedia.org/wiki/Macintosh_LC

>> No.62304547

>>62301519
There was an article in Forbes 2 days ago where they spoke to the team behind Threadripper and they admit Zen1 is pretty much a worst case, I can see them really pulling out some decent improvements over the next few iterations

search for amd-ryzen-threadripper-the-fascinating-story-behind-the-processor-that-beat-intel

fucking spam filter

>> No.62304567

>>62301519
I'm with this, zen+ is coming first. we will see the xx50 and xx70 ryzens on 14nm. 4.5ghz.

7nm will be 5+

>> No.62304585

>>62304567
new wraith max cooler included.

>> No.62304600

>>62304567
Zen+ was just a place holder name. Zen+ is Zen2.
Lisa said in an investor call that Zen+ and Zen++ were renamed to Zen2 and Zen3.

>> No.62304716

>>62304600
There needs to be a filler product in Q1 2018

>> No.62304757

>>62304716
APUs.
They have no nodes to jump onto anyway.

>> No.62304788

>>62304600
>>62304567
zen 2 is 7nm. global foundries will be able to ship chips in 2019.
2018 will need a new product. Even if they are small, incremental upgrades.
Eg instead of r5 2600 we might get r5 1650 with 3.8-4.2 ghz.


>>62304757
APUs wont be enough to combat intels 8000 series. Intel increased core count, so they took away ryzens advantage. Amd will react with increased clocks more probably.

>> No.62304796

These threads read like holistic paid ad articles.

>> No.62304806

>>62304716
>>62304788
They'll release higher binned Ryzen SKUs if anything at all. They'll be relying on APUs for their new product.
There is no new arch aside from Zen2, and that will be late 2018.

>> No.62304816

>>62304788
Ryzen already combats 8000 series.
No one would unironically buy these housefires.
Heck only new i3s sound decent.

>> No.62304851

https://semiaccurate.com/2017/09/06/intel-foundry-customer-bails/

RiP Intel

>> No.62304880

>>62304851
Well that's expected.
Intel's fabs are woefully behind the schedule on 10nm.

>> No.62304887

>>62304806
that might happen. Higher binned cpus will help. Zen 2 is far away, q4 2018-q1 2019 id guess.
Amd cant rely on ryzen 1000 series.
If they manage to reach 4.2ghz like thread ripper does, and ship every chip on the line up with 0.2 ghz increase, they will be competitive.

>>62304816
I dont really believe this. i5 will be 6 cores, and if they keep the same clocks as the 7000s ones, then they will be better than r5 in almost everything. Only extremely multithreaded stuff will be faster on ryzen.
And ryzen 1400 will be next to useless, if there is an i5 similarly priced.

>> No.62304930

>>62302430
Oh no, best affordable ram on either platform is 3200mhz>. all decent ryzen boards are running at least 2933 with xmp on hynix and micron ic.
>9% difference. might be able to tighten timings and lower the latency negating the difference in clocks anyways.
Why is /g/ always up in arms about single digit performance differences?
>should have bought an I5.
buying 7th gen and 200 chipset at all when 300 series only 8th gen is right around the corner. No wonder there was no H210 boards, kaby is already eol.

>> No.62305350

>>62300246
>Bullshit
Dude probably just didn't know what IPC actually is.

>> No.62305429

>>62298939
>768KB-1MB L2 cache
Probably not a good idea, the latencies would go up. It seems Zen's current cache subsystem is probably pretty good, so now's time to work on other parts.

They did mention that more things than they planned worked on first try (hence the IPC uplift target beat), so who knows what actually remains a bottleneck now.

I mean, bottleneck that can be removed/widened to extract more IPC effectively.

IMHO, something between your #3 and #2 scenario is likely, more towards #3, because the changes likely won't huge on paper, still could be +10% IPC. +5-8% would also be good though.

>> No.62306293

>>62298939
That 40% was basically just more cores.

Zen+ will be 10% better clock speed at best.

>> No.62306538

>>62306293
That's simply too dumb/out there to work as bait for anyone IMO.

>> No.62306647

>>62304887
>intel releasing a 6 core i5 for < $200

You also have to remember that the i5's STILL lack SMT, the only advantage the core 7k series had was clock speed, the 8k series is proving this even more.
Almost worst case scenario is that the second iteration has slightly higher overclockability, better IF, and some improvements to IPC here and there because its an entirely new arch.

They already win at power consumption, which is gonna hit intel in the mobile market hard, they already tied in IPC for most things (we've seen its bad at a few specific work loads but at its price point TR beats everything from intel until you get to about twice its price) and I'm assuming that by now AMD have been made aware of the strengths and weaknesses of the released Ryzens, and Threadrippers, made small adjustments for EPYC, and are carrying all those over, plus more to Zen+ on the refined manufacturing process.

>> No.62306793

What I would like to get:

4.7~ ghz ceiling.
More I/O (Most likely not gonna happen on AM4 sockets)
5-10% IPC
Refined memory controller for officially supported 3.2ghz RAM
Refined IF for lower latency and more bandwidth
4 or 6 core ccx don't matter to me
Open source PSP (Might happen with China being a big customer, AMD might just give them the backdoor on the other hand)

I'm not expecting too much honestly, AMD is strapped for cash and manpower, so if they can get higher clocks and less latency then that's ok with me.

>> No.62308052
File: 117 KB, 750x933, IMG_0771.jpg [View same] [iqdb] [saucenao] [google] [report]
62308052

>>62298939
>32 PCIe lanes
Oh
Give
Me
More
Senpai

>> No.62308148

>>62298939
>Memory controller support for >4000MHz
lol imagine how much this would cost hahaha

>> No.62308154
File: 33 KB, 218x210, 1385757966247.png [View same] [iqdb] [saucenao] [google] [report]
62308154

>>62301003
>6.1 ghz
>Over 6.000.000 kilohertz

Oy Vey

>> No.62308164

>>62302272
my 1600x sits on 100fps+ all the time

i am still gpu bottlenecked

>> No.62308262

>>62304887
>Zen 2 is far away, q4 2018-q1 2019 id guess.
Why some people say it's coming early 2017?

>> No.62308269

>>62308262
2018 I mean of course.

>> No.62308344

>>62308262
It's not coming in early 2017, best case scenario is q3 2018

>> No.62308555

>>62308148
My 3600mhz G.Skill runs perfectly fine on my stock clocked 1700.

>>62308262
>>62308269
Early 2018 Global Foundries will be doing nothing other than running test chips during the ramp up phase. Though mid year there is a chance that they'll be producing some small 7nm GPUs for AMD. The second generation Zen chips won't be coming any time without fair warning. Its going to be a huge press matter when their launch gets near.

>> No.62308669

>>62298939
>20% ipc bump 5.0ghz clock
Are people actually stupid enough to believe any of that?

Most likely, we will see a 5-7% IPC jump with 4.5ghz being a reasonable clock to reach.

>> No.62308673

>>62298939
Even in your worst-case scenario, AMD will cuck Intel.

>>
Name (leave empty)
Comment (leave empty)
Name
E-mail
Subject
Comment
Password [?]Password used for file deletion.
Captcha
Action