> http://wccftech.com/amd-naples-32-c ore-zen/ > AMD Z

Don't link wccftech shit
Use this instead:
http://www.anandtech.com/show/10581/early-amd-zen-server-cpu-and-motherboard-details-codename-naples-32cores-dual-socket-platforms-q2-2017

>>56194687
I just took the most recent of the first three Google search results. What is wrong with wccftech?

Kys OP seriously
There are xeons with 20+ cores
All CPUs with 4+ cores are for workstation shit, nobody needs more for gaymen or general use

>>56194704
/thread

That's a server CPU, made to compete with high end Xeons with just as many cores. None of us will own one in our home PCs, unless you're someone like deadmau5 and can afford to pointlessly spend $8k on a CPU.

>>56194700
Its a clickbait blog ran by a bunch of Indians who just steal articles from other sites and treat forum rumors as if they're fact.

>>56194716
>>56194704
It was a joke autismos.

>>56194667
Because you can't parallelize all workloads, retard. I doubt you're encoding video 24/7.

>>56194667
>In 20 years, people are gonna laugh at us for thinking 8 cores was a lot.
who the fuck thinks 8 is a lot ?

>>56194687
I always think it says analtech

>>56194802
>people are going to be laughing about numbers of cores
Autism

>>56194667
>When are we finally going to reach the kilocore mark?
in your gpu?
in datacenters and supercomputers?

Pfft, i will stick with my faildozer dual Opteron 6348 setup. I see lots of problems with am4, starting with: no sb on mobos, like on am1, which have horribly long post times. It also starts to be obvious that 14nm die has some problems, which might realize as low freqs on zen. Because of no south bridge on mobo, board manufacturers will probably try to cheapout on everything possible, including vrm components. AMD will also use PSP on it, just like on the last bulldozer archs.
And the question: will 14nm Zen chips have soldered heatspreader and especially the desktop chips? If not, no buy.

>>56194900
/g/'s amazing sense of humor shines again.

>>56194735
basically this.

>>56194727
the highest core xeon is 22 and that bitch is clocked at 2.2ghz and costs over 5kusd. This shit is incredibly important if it has the potential to shake up the server/compute market.

>>56195148
24 core 48 thread actually.
Its north of $7,000 for that Xeon E7.

>>56194667
>in 20 years we will have 100-core cpus as the norm
>in 20 years we will have 100k gflops for entry-level graphics cards
>1378x720

File: 1463686594903.jpg (36KB, 480x534px) Image search: [Google]

36KB, 480x534px

>>56195053
Youre retarded

>>56195175
20 years until 100 tflops? I give it 10 years at most.

>>56195189
Care to elaborate on that?
T. Different guy

>>56194667
Hey Nigger, GPUs are processors with shitloads of cores that are designed for work that scales well with concurrency, such as solving a bunch of geometry equations at the same time.
The same thing doesn't apply to your CPU, in which most applications rely more on clock speed than anything else.

In dumbed-down nigger terms, we already have what you are talking about.

>>56195227
IT WAS A JOKE.

I IMPLIED THAT IN TWO OF MY REPLIES NOW.

YOU HAVE NO PLACE CALLING ME A RETARD WHEN YOU CAN'T EVEN SEE HUMOR WHEN IT'S RUBBED IN YOUR FACE.

>>56195192
100TFLOPS is a 10X increase in compute power over the Pascal Titan X.
A 10X increase over a span of 20 years? Seems pretty pessimistic.

I think the Pascal Titan X is just below a 10X increase in compute power over the Fermi GTX 480. Those were about 6 years apart.
If we manage the same increase in compute power for the next 5~ years we'll have 100TFLOP GPUs on the high end. Give it another 3-4 and that power will be entry level.

File: 1471601399421.jpg (32KB, 500x500px) Image search: [Google]

32KB, 500x500px

>>56195257

So what's the release date?

>>56195257
>le "Im only pretending to be retarded"
He's right, you are a fucking nigger

>>56195171
Yikes. That shit is even worse value than the 22 core one. How the fuck does Intel even operate like this?

>>56195539
Nobody to compete.

>naples
Confirmed trash

>>56195539
Businesses foot the bill for all kinds of absurd shit. Ever checked the price of enterprise SSDs?
Server racks can easily sink several million bucks.

>>56195539
It's a niche market. Do you complain that Bugattis' cost $1m+?

Unless there's a radical change in how software works there's no reason to expect CPU cores to go that far. You can't parallelize everything and just expect it to work. More cores is beneficial right up until you run out of processes to run on those cores and then they're just a waste of die space and it would have been a better investment to work on new instruction set extensions to improve performance for the other cores instead.

>>56195882
>Unless there's a radical change in how software works there's no reason to expect CPU cores to go that far.

Yes, there is. If we could fab a super computer on a single chip we could. Intel would charge $100,000,000 per unit if they could fit 1,000 high clocking high performance X86 cores on a single chip.
Even if the damn thing needed to be cooled by a constant flow of LN2 it would be the most attractive chip in existence.

>>56196162
You're a fucking retard.

>>56196240
Says the spastic who has literally no idea what the HPC and server markets demand. Nobody builds a 100+ socket rack of servers because they want the sockets. They want the fucking CPU threads to do work.

Stick to /v/, kiddie. You don't belong here.

>>56196162
OP is talking about regular consumer processors not server processors. As it is anything past like 16 cores/threads would probably be waste for us plebs.

File: Intel-Coffee-Lake-Roadmap-635x255.jpg (58KB, 635x255px) Image search: [Google]

58KB, 635x255px

>>56195350
the challenge is that moore law feels like it's bleeding out pretty fast.

for example, intel's tick-tock became tick-tock-ohfuckrespin and apparently has slipped even further now.

> Cannon Lake (10nm) for non-mobile slipping to at least 2H'18...

AMD in 2011
>MOAR CORES

AMD in 2016
>*barely avoids bankruptcy*
>EVEN MOAR CORES

>>56196490
Zen appears to be at parity with Broadwell-E's IPC somehow, so there's not too much to really complain at this point besides the schedule slip to Q1'17.

>>56196528
>Zen appears to be at parity with Broadwell-E's IPC somehow
Wishful thinking lad. Maybe in multiple make workloads, with all gcc optimations in effect.

>>56196489
GPUs don't have the same issues with performance scaling that CPUs do.
So long as you can provide adequate bandwidth you can simply increase ALU(TMUs and all other associated logic too)count and GPU performance will scale accordingly. Nvidia could keep the Pascal arch forever, and just keep adding units to it. Only issue there is lowering power consumption as well.
GloFo at least has its next two high performance nodes lined up. Samsung as they'll go to 5nm without issue, I'm sure TSMC claims similarly. Once we're fabbing ultra low voltage GAAs companies are going to go crazy with high performance designs. The performance of a Titan inside of 10w.

>>56196528
It matched Broadwell-E core for core and clock per clock in one specific Blender scene. This isn't indicative of general performance in anyway.
Zen will be around Ivy Bridge - Haswell in most things sans heavy FPU ops.

>>56194776
I work in Computer vision and this would work wonders tbqh

>>56196595
what you're saying about TSMC, GloFo, GAA, etc. could all be true, but I'm extremely pessimistic about them being able to deliver meaningful improvements on a timely basic at this point after the market leader's slip.

we need improvements in static and dynamic power draw as well as density to continue going where we want.

I would love to see 1 TFLOPS/W and <10W/cm^2 chips within a decade, but I see little convincing layman-tier evidence to give me any confidence in that happening.

>>56195226
Intel also has the SB on die and boot times are fine for one

>>56195148
It depends on the power consumption

File: Screen Shot 2016-08-21 at 3.17.28 PM.png (79KB, 1962x404px) Image search: [Google]

79KB, 1962x404px

>>56195873
>>56195832
>>56195622
Now that you mention it, your right. I just finished accepting a quote for pic related.

So I guess I'm part of the group that supports it.


Damn I'm pullin for AMD to get a good chip out.

>>56194667
>server CPU
>loads of cores
Yeah sounds about right. It will probably perform about the same as a previous gen 16 core xeon if you go by past disappointments from AMD.

>>56196955
wtf is even on a SB in this age of on-CPU PCIe, just a handful of SATA and USB controllers?

>>56197004
Not even. I've got racks right now that have plenty of power, but incredibly limited space.

I can cool anything, but I aint got the space.

>>56194735

>indians

Try a bit slightly north west.

>>56194687
>implying anandtech is any better

jesus man.

>>56196490
>workstation/server hardware
>MOAR COARS XDDDDXXDDDDD
More cores are retarded for the average gaymer because the average pajeet doesn't know a shit about multithreading so gaymes barely use 1 core, but this is server shit, it actually benefits from more cores.

>>56197732
Moar coars are great for me personally. It brings down prices on 8+ core intel CPUs. I run a PCI passthrough setup for gaymen and I'd like to have at least 6 cores dedicated to the gaymen machine.

>>56197732

moar corez are only good for better perf/watt. less cores + high clocks can often outperform moar corez even in properly parallelized workloads.

Who the fuck thinks 8 cores is a lot?

>>56197970
People in the 90s.

We will get 8 cores CPUs with the core performance of say a 4690?

>>56198121
And DDR4. Don't forget DDR4.

>>56198121
Maybe a little less, depends on final clocks.

>>56195350
We are reaching density wall rather quickly now. Just like we reached clock wall some five or so years ago. We might never see 100 TFlop cpu/gpu. Just like Intel's phantasm of 10ghz by 2006 never happend.

>>56198312
>We are reaching density wall rather quickly now.
Not really. Conventional CMOS scaling only has a finite limit on a 2D plane. Stacked logic just like stacked DRAM and NAND is entirely possible.

>>56196162
We already can. Old school Pentium chips are approx~ 1mill trannies. You literally can fit a 1000 of those on a single chip these days and clock them 1-2ghz.

>>56197732
Moar coars are the only avenue of progress in terms of processing power we have at this moment. Unless you fix the leaky trannies in current chips, so we can go up, up and away to 20ghz land.

>>56198343
With the current tech we'll struggle to go past 6-7 layers of logic, and at some point there is going to be an issue with getting heat out of the chip.

Sure - fabrication process will improve, and we will see more potent chips in the future, but there is a definite wall to be reached in our lifetimes.

>>56198346
>You literally can fit a 1000 of those on a single chip these days and clock them 1-2ghz.
Holy shit

>>56198401
Conventional physics puts the limit around 7-8Ghz for bigger chips in absolute perfect scenarios.
We're never going to have a perfect scenario with the materials but let's assume some mature III-IV GAA design gets 99.999999% the leakage reduction of today's fabrication processes, it still wouldn't be enough to make a clock gain after two node jumps, the power density is still too high.

We need more durable materials with better control and better fabrication. Atom layer deposition is the future but we won't see that for decades

>>56198516
True that. But we might get 5.5-6 ghz cpus on the desktop with sensible power envelopes. My first PC was entirely passively cooled, I want that back :)

File: 123.jpg (132KB, 1300x866px) Image search: [Google]

132KB, 1300x866px

>>56194667
>In 20 years, people are gonna laugh at us for thinking 8 cores was a lot.
in 100+ years people are gonna laugh at us for having cores at all

>>56198590
>lol grandpa you didn't have neural nets in your brain, how backwards
kill me.

>>56198590
>in 100+ years people are gonna laugh at us
>in 100+ years
what people? the earth will belong to skynet by then

File: whomightitbe.png (11KB, 274x290px) Image search: [Google]

11KB, 274x290px

>>56198738
>skynet
I agree.

>>56198738
t-the goyim werent s-supposed to know Chaim

>>56198346
>You literally can fit a 1000 of those on a single chip these days and clock them 1-2ghz.

LOL good luck clocking 5 stage pipeline x86 core to 1 GHz, retard.

>>56194667
>When are we finally going to reach the kilocore mark?
10 years ago.

>>56196583
Single threaded tasks Zen is going to beat Southlake sky lake and clock higher

>>56194735
mfw they've been right about almost every rumour

>>56196528
>besides the schedule slip
what schedule slip?
I thought it was limited availability 16Q4 and normal release 17
what have changed?

>>56203744
nothing recent, just that Zen had a soft 2016 launch window until late last year

>>56203744

AMD has to respin before real mass production which means reviewers will get A0 engineering samples on launch day and stores won't have the real silicon under 3-4 months later.

>>56194667
And yet here we are with software that still isn't multithread or can't be multithread.

>>56194704
>>56194716
Don't tell me what I need fagot. This is America, this country is based on WANTS not NEEDS. I WANT 100 cores bitch

>>56204136
Go fuck yourself

I fucking hate that /pol/ always shits up my AMD threads

>>56200064
do you no know what Core M processors are?

>>56194687
>>56194667
MOAR
O
A
R

COARS
O
A
R
S

>>56194667
> implying most apps can use more than one core

AMD is at again.
Goddamnit.

>>56206134
>I don't need it so nobody does

The 12 year old's at it again.
Goddamnit.

>>56195175
Core count will hit odd price tier boundaries much like HDD capacity where you can have a 256GB drive for (example) $45, but the 512 is only $55 and the 1TB might be 65-70

Need of performance and market saturation will reach a boundary. Sure, those core counts will indeed go up. But I cannot see honestly in 15-20 years where the typical consumer will need more than 8-16 cores.

What we're going to see are 3D stacked designs with water or heatpipe style microchannels.

>>56206134
most apps don't need more than one core
what about the ones that do?
or what if your ran more than one at a time?

>not having at least four cores in your CPU
toplol
i hope you got good cpu-schedulers KEK

>>56206134
open your process monitor and count the number of processes currently running.

now think about it.

>>56194667
so why did amd design new cpus but still dont sell them?

>>56194667
In datacenters what you usually have is large clustered virtualized environments.

Let's use a smallish datacenter.

Their virtualization environment will have, lets' say, 10 hosts hooked up to a SAN. Those 10 hosts will have 2 CPUs each with, lets say 16 cores (32 threads each). The 10 hosts will be clustered and running a few hundred to a few thousand VMs that dynamically move around between the hosts depending on the work load.

So you have a cluster that can run 640 threads simultaneously, but hardly anything needs that. So most of the VMs will have 1-4 cores assigned to them.

That's not to say that that's all there is in a DC. For instance I setup a cluster of 6 cisco c240s to run as a single database for an engineer not too long ago, 1 leader node with 5 slave nodes.

We don't do that at home, no need for more than 8 cores really.

>>56194700
Remember Kotaku?

Imagine if Kotaku only got things right 10% of the time, thats wccftech.

They publish shit like " New Pascal card has HMB2 memory! Confirmed!!!!"

1.5 years pass and only the 1080 has GDDR5X

>>56206769
Let's all take a moment to revel in the fact that Gawker and kotaku got fucked harder than any of us could have ever predicted.

Feels glorious.

File: 1471779601291.jpg (66KB, 450x600px) Image search: [Google]

66KB, 450x600px

>>56206775

>>56195539
Businesses need to comply with far more regulations, include fail-safes and prioritise what's going to work most reliably with the largest range of compatibility, that's why enterprise tech is always multitudes more expensive than consumer products, plus large enterprise actually has the capital to buy the equipment in the first place.

Another way to think of it is to compare something like data loss, if a consumer drive fails and someone loses their data there's one pissed off customer that can either claim warranty or deal with it, if a drive in a data centre containing user data fails and losses their data they'd face possible lawsuits and reputation damage which would far outweigh the initial cost of the equipment needed to keep user data safe in the first place.

>>56206782
This is true. DCs are set up with the assumption that hardware will fail often.

When I was a sysadmin I would have to replace a disk on average every other week.

File: owH9F64.jpg (320KB, 1924x1200px) Image search: [Google]

320KB, 1924x1200px

>>56206781

32 AMD cores, still gets beaten by 2 core atom.

>>56196595

>Zen will be around Ivy Bridge - Haswell in most things sans heavy FPU ops.

I want The Stilt to go and stay go.

>>56194667
What's the point of having 32 cores if the output is still weaker than a single core cpu

>>56194667
>32 cores
>on a single die
enjoy your 1 GHz clock rate I guess?

>>56207202
Two 16 core dies actually.

>>56207260
Which still have to stay synced over the interconnect
I bet towards 2Ghz base operating 2.2-3.5 since entire CCX can shut down

>>56194727

Well. I've got 4*e5-2670s, which retailed somewhere along $1500-2000, I guess.

If naples will be somewhere 1000-2000 usd, I could consider getting one.

File: xeon.png (26KB, 1332x1063px) Image search: [Google]

26KB, 1332x1063px

>>56194727
>None of us will own one in our home PCs
Challenge accepted

>>56207111
I can't think of even a single multicore vs single core cpu where this statement would be true

>>56207435
>E5-2670
>Not even v2

Enjoy your shit singlethread

>>56207501
Enjoy spending $1000 instead of $70 for your miniscule performance increase

MOAR CORES

MOAR COARS

>>56207260
NOPE
They're 8 core dies.
The 32c Opterons are 4 die MCMs.

File: Zen_Summit_Ridge_First.jpg (73KB, 369x676px) Image search: [Google]

73KB, 369x676px

>>56207650
Reminder:

Consumer Summit Ridge and enterprise Opterons are all based off of this one die
Single die = dual channel
Dual die = quad channel
Quad die = octa channel

The 32c/64t Opterons have 8 channel memory
Dual socket boards offer up to 64c and 128 threads

>>56208154

> 64c and 128 threads

Even if we assume (for the sake of argument) IPC is lower than the likes of broadwell-e and skylake a 128 thread system is gonig to get a lot of people with lots of money to spend sitting up and paying attention. Hell at that point its a matter of keeping power consumption at sane levels.

>>56208217
Enterprise chips use the highest of the high binned dies when it comes to voltage scaling at lower clocks. That and they don't target high clocks to begin with.
Intel's 24c/48t Xeon E7 8890-v4 has a base frequency of 2.2ghz, peak turbo of 3.4ghz, and its a 165w part.
I think AMD's 32c/64t chips are alleged to be 180w. Not too far off, though TDP isn't some universally consistent thing.

Regardless the high end of the enterprise market has some astoundingly cool offerings these days. It'll be fun seeing how the Top500 list changes over the next few years. Number 1 is that new MIPS arch from China, number 2 is using Ivy Bridge 12c Xeons, and number 3 is using Piledriver Opterons.
That list is over due for a big shake up.

>>56208217
desperately hoping for a 16c/32t that undercuts the intel jew so i can build a comfy workstation without paying xeon e5 v4 prices

>>56194667
>When are we finally going to reach the kilocore mark?
We already have, they're called GPUs.

>>56208394
I'll be interesting to see what if anything is available around $1,500.

Inb4 Zen powered Mac Pro with dual Radeon SSGs

>>56194667
We already had some kilo core processor projects

>>56208809
Mac Pro (Late 2016) will come with a cluster of ARMs to help you create the most creative iOS apps ever.

>>56208396
> counting individual shader ALUs as cores

shiggy

File: authors-robert-martin-v0.jpg (63KB, 300x300px) Image search: [Google]

63KB, 300x300px

>>56194667
Uncle Bobbo was right, guys.

>>56208809

You joke but if AMD is in a position to sell to Apple it will be a huge win for them - they already supply apple with most (all?) of their gpus so getting the cpu side as well is going to fucking hurt Intel.

SoC is the future and all that holds AMD back is their fairly meh cpu performance. Get that upto snuff and AMD could very be the solution to a lot of needs (i'm looking at you laptops).

>>56209137
Their APUs are supposed to be nice. They're what the current generation of consoles use, right?

Apple likes intel chips because of their low power consumption. It's one reason why they can squeeze out so much life from their hardware even though it's typically underpowered compared to Windows machines. If they can get these multi-cored chips to do as well (or better) than intel's offering, then Apple would switch.

But, being that the Apple empire is built on mobile tech, they'll likely use ARM for everything except desktops. X86 will be for stationary hardware.

>>56209137
It wasn't a joke, its a real possibility.
If AMD's moar coars offer the better performance proposition this time around Apple would be compelled to use one. The current Mac Pro does opt for dual sockets, and a single MCM could handedly destroy it in multithreaded throughput.
If AMD decided to make them a deal for their SSGs coupled with their new CPUs? It'd be a no brainer. The performance of those things is ridiculous.

>>56208154
so what good exactly does 4x SB do for a big Opteron?

off a bazillion USB, SATA, and PCIe lanes that no mobo could ever hope to support?

the more memory lanes seems smart, but the massive peripheral fanout sounds wasteful to me...

>>56209319
Apparently they don't have a dual socket offering, just dual GPU.
Tops out at a 12c Ivy Bridge Xeon.


>>56209396
Disabling unused IO is laughably trivial.

>>56209396
They could just make it so that the unneeded southbridges dont have anything hard-connected to them, and disable them via BIOS.

As >>56209434
said, its laughably trivial to block them off.

>>56209434
>>56209447
It should be obvious I'm talking about die space, not power, geniuses.

>>56209591
>worrying about a trivial amount of die area
>on an MCM

Not too bright are ya, champ?

>>56209591
Its ultimately cheaper to re-use one CPU die, turn shit off through bios (and waste a little die space) than it is to make an entirely dedicated piece of silicon for external IO

>>56194667
8 core is enough
this is starting to become a fashion in the industry where people are just doing it because its hip and in
the task management overhead with 32 cores is somewhat questionable as well though there can exist smart solutions
still, simple, serial processing is the heart of normal computers. Special computing projects have special needs but thats something else

>>56209621
>>56209624
AMD needs to offer enough IO per die to make a competitive desktop/workstation part but not waste too much wafer space for parts that might ultimately go predominantly towards MCM Opterons.

There's no guarantee that they'll be able to command Intel-level price premiums for high-core count CPUs, especially if their perf/W is considered any worse.

I was disappointed by the Broadwell-EP Xeons and want AMD to be competitive, but I'm not convinced yet that they're actually being smart again in their designs yet.

>>56207526
>miniscule

Something something buyers remorse

>make -j64

My body is re@dy.

>>56209695
Again, the die area you're worrying about is trivial.
The Summit Ridge die is in the range of 150mm2~, the number of candidates per wafer is tremendous.

>>56209695
well, there are AMD server boards out there where they have 2 northbridges, with the second one there with the explicit purpose of adding another 24 PCI-E lanes for a total of 64 lanes.

Believe me, if the lanes are available, there will be customers and boards available to make use of them.


Does anyone have any hard numbers on the number of PCI-E lanes Zen is supposed to support per-die?

>>56209833
No word on that yet.

>>56209739
>Summit Ridge die is in the range of 150mm2~

fuck I had no idea.
I really hope 16c MCMs sell for sane prices.

>>56209719
>Buyer's remorse
Sounds like you're the one with buyer's remorse considering you spent $2000 instead of $100.

My investment is not high, yours is.

Personally though I hope Zen in general beats the crap out of Sandy Bridge-E, because if it does I intend on upgrading from my E5-2690 to a Zen setup.

>>56208154
>Single die = dual channel

I'm a bit less excited for Summit Ridge desktop if this is the case.
Will I need to pay a 5x or whatever premium to get a 4 channel MCM workstation chip?

>>56194704
>u dont need more than 4 cores faggots!
>100 fps
>can't open up fucking calculator without dropping frames

>>56210276
seconded.

RAM is cheap enough that I can keep a shitton of stuff running in the background without worrying, but games can get hamstrung when doing this for lack of cores. (or at least smart enough process scheduling by the OS)

>>56195053
>no sb on mobos, like on am1, which have horribly long post times.
not sure how long post times are a result of this. wouldn't current day intel chips post noticeably slower than the pre-IMC ones?

>It also starts to be obvious that 14nm die has some problems, which might realize as low freqs on zen.
yeah, the 480 results have me worried too.

>board manufacturers will probably try to cheapout on everything possible, including vrm components
what prevents manufacturers from doing this right now? because they have to bundle Premium Intel® Chipsets?

>AMD will also use PSP on it, just like on the last bulldozer archs.
fuckin' hell I hope not. hopefully some desktop chips ship without it.

>>56206797
Only every other week? Must have been a pretty small operation all things considered.

>>56206782
Oh you dont need to tell me, as much as I complain, I just finished putting in an order for two data visualization machines totaling well over 30k.

Small fries for many, but our team is a small part of a relatively new group, so it was pretty significant leap for us.

Too bad shit's gonna take like 5 months to get cleared through acquisition to be built and shipped out. I need my damn server.

>>56203954
Cough all browsers and renderers other than servo cough

>>56210276
Most games aren't optimized for moarcoars, it's generally better to go for fewer coars with higher clocks. Even if games start getting good use out of muhcoars, the current CPUs with lotsacoars will probably be obsolete by then.

>>56206740
> we don't do that at home
Speak for yourself /v/

We need to see more multithreaded programming paradigms.

What a good processor. Too bad it's designed for a chipset and socket which will only get server motherboards.

>>56210190
Yeah you're not getting quad channel on socket AM4.
The difference in memory bandwidth doesn't matter in 99% of workloads though.

File: GCN vcore-clock.png (37KB, 1171x703px) Image search: [Google]

37KB, 1171x703px

>>56210412
There is nothing about the RX 480 that indicates any sort of problem with the 14nm LPP process.
The short comings of the card are entirely due to design choices made to keep cost down, the process itself is incredibly solid.
Its beyond ridiculous thats a board of supposed tech enthusiasts understand so little about tech.

>>56211009
what's public knowledge about AM4 at this point besides that it will do DDR4?

can someone explain why i would only need 4 cores for gaming but 6-8 is good for using CAD, photoshop, video editing, etc?

does it have to do with multitasking or something?

>>56211609
Because things like video editing are trivially easy to parallelize - the manipulation you're doing to a pixel in one point of an image doesn't have any relevance to another point a hundred pixels away, so you can do both at the same time. Games, meanwhile, are highly interconnected - what you need to render next depends on the AI behaviour, the AI behaviour depends on a number of factors, and everything depends to a lesser or greater degree on the player input. So if you try to spread that around, everything else can all too easily come to wait on one thread - and before information from that thread gets to the others, it'll have to travel back to L2 or L3 cache at least, which slows them down even more, so when the first thread suddenly needs input from the other threads, there's yet another delay...And then everything has to go through the graphics API which one of the most CPU intensive tasks especially when it comes to FPS games and the like. The graphics API itself is heavily single threaded pre-DX12/vulcan and there's jackshit you can do about that without switching to one of the new APIs and THAT on the other hand almost certainly requires rewriting large portions of any graphics-related tools you use.

>>56211609
Games are developed around 2-4 cores so they tend to do just fine around that area. More than that doesn't necessarily improve performance and the money probably would be better spent on a better graphics card.

>>56209728
>POWER8
>12 core, 8 threads per core
>make -j96

File: amd-soc-and-promontory-chipset-io-coverage-1024x724.jpg (195KB, 1024x724px) Image search: [Google]

195KB, 1024x724px

>>56211558
Stuff from the leak in early 2015

>>56211886
POWER9 detailed at HotChips tomorrow.
The Zen core arch presentation is right after.

Get. Fucking. Hype

>>56211609
game engines have a couple big challenges:

> they're highly latency sensitive: go past 33/16/8/whatever ms? drop a frame
> they're generally not doing data parallel computations for easy efficiency
> there are a lot of subsystems in play, including a lot of resource juggling
> parallel programming is hard, and not many games devs are good at it at all

>>56211921
POWER9's coming out? Hot damn

File: ooh baby.png (6KB, 498x95px) Image search: [Google]

6KB, 498x95px

>>56212046
Tomorrow night. Starting at 5:45 PDT.

>>56211889
that's not as much PCIe on a single-die SR as I had been lead to believe AMD would support.

I guess SLI/CF is falling out of fashion even with Vulkan/DX12 explicit multi-GPU.

> and yes, even though PCIe bw barely affects games normally, CFX shunts all frame data through the PCIe bus, 16x may not cut it

>>56211921
>>56212046
>>56212097
do the presentations and proceedings for things like HC get released/leaked quickly?

don't want to wait forever to hear more about these

>>56212097
How far is this in relative time? 20 hours away?

File: 14-1080.3338864818.png (71KB, 1920x1076px) Image search: [Google]

71KB, 1920x1076px

Zen details here:

http://www.pcgameshardware.de/AMD-Zen-Codename-261795/Specials/Architekturdetails-Benchmark-IPC-1205041/
http://www.pcgameshardware.de/AMD-Zen-Codename-261795/News/CPU-Kern-Groesse-1205423/
https://www.computerbase.de/2016-08/amd-zen-architektur/

Shills who claimed L3 was limited to one CCX are officially BTFO

File: 15-1080.2713563378.png (94KB, 1920x1076px) Image search: [Google]

94KB, 1920x1076px

>>56212199

>>56212199
IT'S OVER AMD IS FINISHED

>>56212199
>nazi propaganda

File: tmp_9004-20160725_160530-229265336.jpg (4MB, 3984x2988px) Image search: [Google]

4MB, 3984x2988px

>>56194727
Ok

File: tmp_9004-20160720_161039157681785.jpg (3MB, 3984x2988px) Image search: [Google]

3MB, 3984x2988px

>>56212309

File: 12-1080.2968623445.png (72KB, 1920x1079px) Image search: [Google]

72KB, 1920x1079px

FPU in the Zen core is substantial compared to the FlexFPU from the BD family

>32 cores

Holy shit boys imagine how many cows we can turn into beef

Since most of /g/ is woefully ignorant of how cpus work i'm fully expecting the intel defence force to be out in numbers after the hot chips presentation.

>>56212365
>mmx/x87
>2016
>semiconductor companies still waste silicon on that crap

File: 11-1080.1838658295.png (81KB, 1916x1080px) Image search: [Google]

81KB, 1916x1080px

>>56212394
>close to 2X the bandwidth on L1 and L2 cache
>L3 5X faster
>has micro op cache
>whole front end massively improved
>can issue, decode, execute, and retire more instructions than Excavator flat out

Neat.

>>56212435
The core size might be about 5mm2, Broadwell cores on intel's denser 14nm node are a bit under 7mm2. Zen is comparatively still a small core, the die area taken up by those legacy FPU instructions is too small to matter.

>>56210050
>you spent
Nope

Also
>$2000
Not even close to their real price

>>56212394
>5x L3 bandwidth

Piledriver was truly a mess.

>>56212465
>Not even close to their real price
Please show me a E5-2670v2 listed for less than $1000

>>56211042
>understand so little about tech
silicon is not my area of expertise. I don't have to be well versed in every single fucking subset of "tech" there is.

now if you actually spread all that knowledge of yours instead of being a pompous cunt about it you might fix the problem you're bitching about.

>>56212559

There's no point because trying to educate anon gets drowned about by memes and shitposting. One that stands out is a few posters on /g/ said not to buy a 970 and go with the 390 due to a shift in in the industry but most of /g/ just meme'd about it and didn't listen - these days the 390 is by far the better card of the two.

>>56212618
the signal-to-noise ratio is only so bad because there's just not enough signal.

>>56212559
https://www.techpowerup.com/gpudb/2848/radeon-rx-480
https://www.techpowerup.com/gpudb/2663/radeon-r9-390x

The Polaris 10 die in the RX 480 only has half the renderer backends of the Hawaii/Grenada die used in the R9 290X-390X. That means per clock it has half of the pixel throughput, but its a much simpler design.
To over come some of the lowered pixel throughput, the RX 480 pushes clocks as high as possible to make up for it.
If the RX 480 had 64 ROPs like Hawaii/Grenada it would absolutely destroy the R9 390X in performance in most benches.
Its performance, or perceived lack there of, is entirely from the architecture.

The 14nm LPP process clocks incredibly well, and volts/clock shows ideal characteristics for a FinFET. Theres no problem to be found here.

File: cachemem-vishera.png (255KB, 539x485px) Image search: [Google]

255KB, 539x485px

>>56212394
for reference, here's Piledriver at 4ghz.

Quick and dirty find, if someone has a piledriver chip here and AIDA64, could you be so kind as to give us better numbers?

>>56212618
>the 390 is by far the better card of the two
Well people weren't exactly expecting the 3,5gb lie. Regardless 900 series as a whole was pretty shit and barely better than 700, 10xx has a much bigger leap in performance.

>>56209739
>The Summit Ridge die is in the range of 150mm2~,

[citation needed]

File: 5-1080.2494205094.png (64KB, 1920x1074px) Image search: [Google]

64KB, 1920x1074px

This might be worth reading into, AMD is now giving us two comparable metrics.
IPC uplift, and energy per cycle.

Energy per cycle is shown equal to Excavator.

>>56212618
>these days the 390 is by far the better card of the two.

a 390 is just an OCed 290. an OCed 970 still outperforms the 290/390 handily. the 390 is massively behind an OCed 970 in both performance and perf/watt.

>>56212957

Its really not behind in performance.

Christ even this year old video has a 390 besting a fairly highly clocked 970.

https://www.youtube.com/watch?v=k9cKZiJw6Pk

File: IMG0046463_1.jpg (350KB, 1450x816px) Image search: [Google]

350KB, 1450x816px

>>56212938
The voltage curves shown here are 1:1 accurate to a middle binned chip.
Its normalized to 3.5ghz.
These are the extrapolated values:

Power per Core Pair

XVR with AVFS
3700mhz: 22.5w~
3500mhz: 16w~
3150mhz: 10w~
2800mhz: 7w~
2450mhz: >5w~
2100mhz: 3w~
1750mhz: >2.5w~

STR
3700mhz: 25w~
3500mhz: 20w~
3150mhz: 14w~
2800mhz: 10w~
2450mhz: 7.5w~
2100mhz: 6w~
1750mhz: 4.5w~


The leaked Summit Ridge ES was shown to have a 2.8ghz base clock on all cores, and a non specific turbo of 3.2ghz. It has a 95w TDP.
An Excavator module takes roughly 10w to hit 3.15ghz. At equal energy per cycle 8 Zen cores at the same frequency would pull 80w give or take.

Worth noting that power consumption in FinFET devices increases more linearly than in planar ones.

File: 35ap4hx.png (603KB, 800x600px) Image search: [Google]

603KB, 800x600px

>>56213019
Also of note
The Excavator module sans L2 is 14.48mm2. If ported to 14nm with maximum density it would take up roughly 4.3mm2.
The Zen core is extrapolated to be around 5mm2, if this holds true then Zen is comparatively larger than the entire Excavator module.

Pretty neat.

File: 8320e 4.7ghz aida64.png (262KB, 1420x604px) Image search: [Google]

262KB, 1420x604px

>>56212919

Since I was bored here is what the trial version of aida64 extreme spat out at me.

>>56212097
Where do I watch

>>56200064
>LOL good luck clocking 5 stage pipeline x86 core to 1 GHz
What is the Pentium 4, for 10$

>>56212516
EBay $70

>>56204675
Eurocuck detected

>>56213100
20-30 stage pipeline. General rule of thumb is that the longer the pipe is the simpler the stages in it are and the faster it can go in terms of clock.

Its why a Prescott Celeron held the world overclocking record for years until Bulldozer showed up.

>>56213019
>>56213077
Using this information we could then guesstimate ballpark clocks for AMD's upcoming Opterons.

At a nominal power of 5.5w~ per core at 2.45ghz we get 44w~ per 8 core die.
4 of these dies in an MCM works out to 176w~
The 32c/64t Opteron was alleged to have a 180w TDP.
If that holds true then it looks like its clocked around 2.4 to 2.5ghz.

File: ebay.png (286KB, 1068x1512px) Image search: [Google]

286KB, 1068x1512px

>>56213147
Link?

Cheapest I could find is $1000

>>56213003
>sponsored youtube review

he's being paid to say it's a good card, in reality the 970 was better performance for less money for most of the 390s lifespan as long as you weren't a goyim and knew where to find deals.

>>56208154
Die would be closer to 180mm2 if the core size was 5mm2

File: CoD BO 3 1080p.png (28KB, 500x610px) Image search: [Google]

28KB, 500x610px

>>56213321

>jayztwocents
>favouring AMD

See, this is the shit i'm talking about. We have someone who flatout states he doesn't give a fuck about AMD reccommending an AMD product and suddenly he's bought off.

>>56213492

He's a youtuber and he's visibly old. Of course he's bought off, he wants to make money.

>>56213518

The question is do you really think the likes of AMD or its partners bought him off given he reviews what, 1 AMD product for every 10 Nvidia ones?

>>56213540

Yeah? AMD and NVIDIA buy off youtube '''reviewers''' all the time. Just look at linus shill tips, same business model.

File: 1422917684557.png (433KB, 1780x1408px) Image search: [Google]

433KB, 1780x1408px

>>56213566

So where are we left then? The belief that in the here and now the 970 is the superior card because everyone knows it (much like koreans believe in fan death) or a lot more evidence suggesting the 390 is the faster card?

I'm not going to argue perf/watt as that is blatantly in Nvidia's favour (though not as much as /g/ believes).

>>56213624
>The belief that in the here and now the 970 is the superior card because everyone knows it (much like koreans believe in fan death) or a lot more evidence suggesting the 390 is the faster card?

OCed 970 > 390 for 1080p gaming
390 > 970 for 1440p

>I'm not going to argue perf/watt as that is blatantly in Nvidia's favour

perf/watt is a big deal if you care about OC headroom and not turning your home into a sauna.

File: certified shit wrecker.jpg (13KB, 381x286px) Image search: [Google]

13KB, 381x286px

Made up drunken benches hot off the press

Zen has a normalized performance of 1.
Broadwell has a normalized general performance of 1.3x.

At 2.2ghz base clock the E7 8890v4 has a throughput of 2.86 Napkin Math Units per core.
At 2.45ghz base the 32c Zen Opteron has a throughput of 2.45 Napkin Math Units per core.

With 24 cores the Xeon E7 has a perfect scaling throughput of 68.64 NMUs
With 32 cores the Opteron has a perfect scaling throughput of 78.4 NMUs

165w for the Xeon
180w for the Opteron

Quad channel memory for the Xeon
Octa channel memory for the Opteron

File: 1448892427415.jpg (367KB, 599x900px) Image search: [Google]

367KB, 599x900px

>>56212334
Good god do I want that CPU.

>>56213799

>perf/watt is a big deal if you care about OC headroom and not turning your home into a sauna.

Unless you're harry potter its really not.

File: nv_online_marketing3.png (36KB, 500x850px) Image search: [Google]

36KB, 500x850px

>>56213956
>power doesn't matter

kek

>>56214053
It matters far less than Nvidiots would have you believe.

File: 10-R9-295X2-Power-Consumption-Gaming.png (53KB, 591x934px) Image search: [Google]

53KB, 591x934px

>>56214053

>posting tpu to try and refute a tom's hardware pwoer measurement

This is how you can spot the noise on /g/. Hint: tom's (especially their german arm) does far more accurate measuring than just about any other site as they have the gear to measure exactly where the pwoer is going.

How do you think the 480's power draw over motherboard was found - tom's was testing the card at 4k in metro LL with their expensive equipment.

tl;dr you're using rather inaccurate numbers compared to what was provided earlier.

File: Zen team.jpg (75KB, 715x573px) Image search: [Google]

75KB, 715x573px

>>56213815
AMD delivers more napkin math per watt
Intel is finished

AMD is being too quiet on the interconnect/coherent bus issue, 4 die MCM is no joke from a engineering standpoint.

>>56214509
GMI links
Big beefy cache units sitting in the middle of the L3
Undoubtedly they are all linked together on a common bus

>>56214509

I wouldn't be surprised they are quiet on the specifics to not tip Intel off on their implementation.

>>56214509
Is it gonna be HT 4.0 or something new? HT 3.0 even how old it is still has more bandwidth than QPI

>>56214124

100w extra at load makes a big difference when gaming (i've used both, even in a well ventilated home that extra heat builds up), not to mention that hawaii cards draw close to 80w at idle if you're using multiple monitors as well.

switching from a 290 to a 1060 was a huge upgrade in this regard, my computer is entirely silent and doesn't increase ambient temperature by 10c like before.

>>56214802
Latency is a bigger issue than raw bandwidth, I don't think there is any known update to the Hypertransport spec beyond 3.1.

I believe they have a new fabric.

>>56214815
>150w difference in GPU makes room twmp go up by 10C

You are so full of shit

>>56212394
With all of these improvements I don't doubt that Zen will BTFO of the Bulldozer series. I hope that's enough to become competitive with Intel though.

>>56214944

b-b-but The Stilt said Zen is shit!

>>56214944
Depending on the workload Zen will perform as well as Broadwell, in the specific case of Blender, or as poor as Nehalem.
Average will probably be around Ivy Bridge to Haswell since Zen lacks an enormous FPU to compete with intel on that front.

Thing to consider is that Zen will be doing this at comparatively lower power.

>>56214954
I mean Zen is definitely a significant upgrade on paper. Basically every single one of those fixes has the potential to boost performance in a fairly significant way. I can't see any possible way for Zen's performance to be worse than any of the Bulldozer series chips, the power usage might be slightly worse but performance/watt Zen will probably have a very significant advantage over the Bulldozer series.

>>56215036

> the power usage might be slightly worse

I significantly doubt that given the absurd power draw the 8 core fx chips have.

>>56194704
>There are xeons with 20+ cores

I know I can't expect intelligence on /g/, but those Xeons cost an arm and a leg.

>>56215036
>the power usage might be slightly worse
AMD is claiming Zen has the same power per clock as Excavator.

See:
>>56213019
>>56212938

If AMD can pull another 15% IPC next year then that will be a real winner.
No 6 core mainstream Intela until late 2018

>>56213282
I'm quite certain he is referring to engineering samples. Which I certainly wouldn't buy.

>>56215101
>AMD is claiming Zen has the same power per clock as Excavator.
Jesus that's a huge step up then.

>>56215155
Their Zen+ chips will probably clock significantly higher.
Both of GloFo's next high performance nodes entering production were designed by IBM for their own POWER line of chips. IBM has been pushing their arch to 5ghz since 65nm, I don't foresee that stopping with their 14nm SOI FinFET process either.

AMD's next gen chips will likely use this process. The other possibility is IBM's 7nm, but that likely won't be online until 2018-2019. Quad patterning is too expensive.

File: Certified shit-wrecker, world saver, anal annihilator.png (777KB, 1280x720px) Image search: [Google]

777KB, 1280x720px

>>56215186

All in a day's work.

Now it feels like BD family was shelved and used as a dumping ground for Power related IP for future product use.

>>56215209

>possibility of 5ghz zen+ chips

Please anon, I can only get so erect. Still I don't see that happening too easily as the power draw would be insane - intel's mainstream and enthusiast chips get thirsty as fuck when clocked so high as well.

>>56215293
Target frequency range is a product of the transistor libraries used. Intel isn't trying to make 5ghz beasts, they're aiming for reasonable TDP chips to butter up OEMs who design products around a given size, power usage, and cooling capacity. They ultimately want their mainstream desktop chips to scale to mobile power levels to fill all the odd 35w-45w TDP roles that OEMs want for SFF systems and AIOs. A high clocking chip with more leakage above 4ghz won't scale down in clocks as well, though you would end up with a higher stable clock ceiling. They aren't building chips to appease the internet, they're building chips to sell the most volume possible for the least investment possible.

SOI FinFETs have significantly lower power per clock than current bulk FinFETs. They're inherently more consistent structures which lends to lower drive current to compensate for variation. They have a ton of advantages over devices built from a bulk starting wafer. Scalability and even lower leakage are two big ones.
We might see 8 core Summit Ridge ship with a 3.4ghz all core turbo. The next generation 8 core Zen+ chip might ship with a stock clock of over 4ghz without increasing power consumption.

>>56215461

> Intel isn't trying to make 5ghz beasts, they're aiming for reasonable TDP chips to butter up OEMs

Well yeah, thats what AMD will be doing as well. Equally though we know AMD has been willing in the past to release balls-to-the-wall clocked chips just because they can (most notably the 9590). I wonder if there is enough of a market (notably within the gaming community) to justify a 4c/8t zen+ chip clocked to the fucking limited (in this hypothetical, somewhere approaching 5ghz). Certainly such a hypothetical chip - assuming zen and its descendants have good ST performance - would be quite nice for those gamers willing to spend the money on high end parts, at the very least it would potentially make a lot of i7 models look less enticing.

>>56215224
They did add an additional decoder to the cores in streamroller so they were probably still working on the cores up to that stage. Maybe Keller came in around that time and didn't like what he was seeing with steamroller and decided to start over.

>>56215542
AMD has an answer for that in their mainstream chip, the Raven Ridge APU.
In Carrizo/Bristol Ridge they finally introduced a third voltage, so they have active per part voltage, and that carries forward. CPU, IGP, and NB all have separate voltages now, so they could release a 150w 4c/8t APU with clocks pushed as high as possible. Let each core pull 25w, and have IGP and uncore take up the remaining 50w.

Though 25w being pulled over a 5mm2 area is absurd heat density. Getting near impossible to cool.

>>56215681

I was more thinking systems that have a high end discrete gpu rather than a monster APU. AMD already slaughters Intel on the apu front with their current offerings.

http://www.relaxedtech.com/reviews/amd/a10-7860k/4

>>56215166

Depends on the ES. Release stepping ES are perfectly safe to use, maybe even preferable since sometimes they'll have stuff like unlocked multiplier.

>>56197475
Pakis?