"Breaking Moore's law is what we did with Infinity Fabric."

"Breaking Moore's law is what we did with Infinity Fabric."

Is he right, Sup Forums?

youtu.be/NoelgG8JoyQ?t=8m25s

there's literally no limit to "moar cores", so yes

Moores law is just a meme now. it died years ago.

Transistor density can only take you so far.

Moore's Law is a meme now, but AMD casually putting ~780mm^2 of silicon in one package for cheap is a disaster for Intel.
EPYC 7601 competes with 8176 for half the price.
And i'm not even talking about memory or I/O. Or AMD's 1S solutions.

DELET

I wish i could. Truth hurts.
I wish Intel was smart enough to stay fucking silent instead of spreading LITERAL FUCKING FUD about EPYC.
They spent more time shittalking EPYC than leveraging strenghs of SKL-SP.

The entire presentation is a good listen. Lots of info, no shittalking. Ok, very little shittalking.

>"Breaking intel anus is what we did with Infinity Fabric."

...

...

Well played. But RISC-V isn't ready for primetime just yet. It will be the IoT future though.

what is this

the future of memes

Infinity Fabric, which is really just a unified host of multiple different buses built upon a new Hypertransport, is the interconnect IP that will facilitate AMD's ventures into die stacking.

Multi die chips with individual components fabbed separately. Tiny dies each ran on a process with metrics specifically tailored to the part. It starts with simple MCMs, and it will evolve into breaking up unified dies into modular components. The era of fabbing monolithic large dies is coming to an end, at least in the short term.
It fundamentally is breaking Moore's Law since the number of transistors you can put on one chip economically is not totally irrelevant.

AMD started talking about this back in 2011/2012. They had a couple presentations in 2013 and 2014 where they went into more detail.
This has been a long time coming.

So this is why Intel is so afraid, because their bingbus can't into 3d

Moore's Law is long since dead.
Stitching CPU cores together and calling it a day isn't under Moore's Law.

Its true. Intel held back moore's law because no no competition. Now AMD competes and destroys intel's artificial "moore's law is dead" nonsense garbage.

Expect Zen2 to escalate things even more.

AMD's Ryzen architecture doesn't suddenly allow for Moore's Law to continue. In fact, it's an admission that it can't go on.

Sure, it admitting mores law is dead and then not caring and making faster Cpus anyways. Kinda breaking it.

>Kinda breaking it.
If the law is dead, then the "transistor per square inch doubling" isn't going to happen, it may increase, but it can't keep complying to the law, so its not really breaking it.

They didn't break moore's law, they circumvented it.

And it's one hell of a quote.

modular DIY CPUs when?

>glued together

is here to stay apparently even AMD said it
Intel probably not even aware of b the beast they unleashed

This
Everyone forgets amds plan for HSA because of the joke that apus are now, but infinity fabric is the next step. Combining something like a 1700 and a cut down Vega die inside the space of an epyc cpu with 8gb of shared hbm is somewhat feasible, and the performance would be nuts

There's nothing stopping them from using 300mm^2 dies on 7nm next time.
Absolutely nothing.

Holy shit. AMD is going to disrupt CPU manufacturing and make the world a better place.

>meanwhile Intel has trouble bringing out mobile dual cores on 10nm

It'll be a bloodbath

Not really, Intel has been prepared for a long time, in a year or two they'll release what they've been holding back from their R&D billions into CPUs and there won't be a AMD to speak of anymore

>m-muh glued t-together dies, goyim
>glues IHS to CPU themselves

>just w-wait Intel will unleash THEIR SECRET WEAPON

>Source: My uncle works at Nintendo

new arch only in 2021, intel said this themselves, so fuck off

Oh wow, this is just sad.

Dude Intel didn't have any competition in the CPU market for like 8 years.

>b-but muh non-NUMA optimized workloads ..

>y-you should activate NUMA on our scalable xeon processors for maximum performance though

he's right.
now they can just glue the die together and get the same performance for fraction the price of a huge die.

I'm watching some retard on another site right now dance around claiming NUMA on Threadripper is the same as a dual socket system, and that you can fix Skylake-X's shitty memory latency by overclocking the mesh.

so AMD basically managed to make their processors scale at 98% on "8 socket" systems?

>yfw eventually your entire PC case will consist of glued together processors.

Aside from memory latency sensitive workloads I can't think of much that would suffer under NUMA, and Intel is in the same boat now with their mesh shit anyway.

>It will be the IoT future though.
If it succeeds in IoT, it's almost inevitable it will end up in mobile devices, set-top boxes, Chromebooks, etc.

>watching some retard on another site right now dance around claiming NUMA on Threadripper is the same as a dual socket system,

lolfail9001?

Yeah fantastic name.

So thats either Reddit, Hardocp or possibly OCN (I forget if he has an account there). If its Hardocp he is part of a clique of posters that relentlessly aim to purge anything AMD related from the user forums.

gib thread link

He's on anandtech and reddit.

You know what this whole thing is telling me? There has never been a better a time to be a distributed systems programmer. Soon, your cheap ass commodity server will look more and more like one, with, e.g., groups of cores. Remember the switch to multicore? It will be like that times ten. You will have to understand not only concurrency and threading, but the difference in latency between different groups of cores and how to efficiently connect them together. If you don't know Erlang, get studying.

After modular DIY smartphones.

Nobody has time to waste on learning AMD's super special snowflake non-standard CPU architecture, most will settle for the tried and tested and easier to work with Intel architecture

>Microsoft
>Baidu

>nobody

...

You can't break Moore's Law, it's not a kind of "law" that you can break.

>chinks
>Microshits
>Amazon
Who?

I think you didn't understand. I was saying that all CPUs will end up like that, just like all of them ended up 64-bit and multicore.

Stop sniffing glue.

please be b8

Yes, you can circumvent it.
This and Tigerlake-SP will use EMIB anyway.
Intel themselves said that.

Oh boy it's a nice race.
Even nVidia joined it.

They have a whitepaper. AMD already has a GPU with Infinity Fabric connected HBM dies. Granted, it's more of a stepping stone, but they're getting there. I think AMD will have the first MCM GPU.

>I think AMD will have the first MCM GPU.
Just like ATi was the first to unified shaders with Xenos.
nVidia is (almost) always late.
To everything, especially new and shiny memory tech.

Why does it even matter? Now they(CPU manufacturers in general) are spending those transistors on more cores, but what were they spent on before? 2-4 cores were the standard for a fucking decade.

>but what were they spent on before?
For Intel? A shitty GPU

Intel beats AMD at their own game, makes better CPUs and better GPUs

How can this company even function?

t. Brian JUSTnich

>2-4 cores were the standard for a fucking decade.
That has more to do with Intel holding the entire market back to line their own pockets than anything else.

>better CPUs
Gotta hand it to them, they still got those high clockspeed quad cores.
>better GPUs
Absolutely retarded.

Your tears are delicious weebshit

when will intel go full retard and release a quad core with 200W tdp?

Maybe on their shitty 10nm node.

but will they be able to fix their 1.7% yields enough for something like that?

Icelake is what, 2019?
It won't matter since Zen2 on 7nm LP will clock higher.
And Zen on 7nm LP+ (the EUV one) will shit on Intel's grave.
Long live Nazi science.

B A T T L E T E S T E D

lmao

b-but muh 300cb single threaded score

>128 PCIe lanes in a single epyc
There goes the crypto coin market

motherboards with 120 pci-e slots when?

They hace to release it before zen 7nm comes.

>even the lowest end epyc has 128 lanes of going fast
this will change things lmao

Actually, it's quite the opposite here literally.
Ryzen uses a trained neural network to perform it's branch prediction, and while it is "set in stone" on the Ryzen CPUs, AMD can improve it by just running more and more softwares on the simulation, making the neural net learn and improve.
The next iteration of the Ryzen will be faster without any sort of engineering effort, just a guy training the brain.
Of course, would be nice if they actually did some engineering effort and didn't just relied on the spongy branchy thing because that's just lazy.

>implying zen2 is only getting branch prediction improvements

Obviously.
But if they were lazy, they could just sit and let the brain churn.

Let me remind you that each Bulldozer revision was a huge improvement, and it was obvious AMD let 3 people just improve it while their engineers did more important things.
I mean just look at steamroller, internally it had more changes than Sandrybridge up to Kabylake, too bad the arch itself was shitty but it had a lot of changes done to it over the years with basically little effort from AMD.

You don't think they'll double down until Zen3 and make this thing in a fucking beast? Zen>Zen3 hitting over 20% more IPC at far higher clocks doesn't sound too insane at all, the arch is fresh as a virgin.

As intel said, Ryzen is not "battle tested".
Which kinda only spell bad news for intel, as the "green" Ryzen is already blaring all the red alert alarms at their HQ.

>Intel spends like 10 slides shitting on EPYC
Kikes are LITERALLY on suicide watch.

They already struck a deal with Amazon and Microsoft so...

You forgot Baidu.

Yep, but i imagine Amazon and microsoft will use the threat of Ryzen to get quite a few extra favors out of intel, and maybe a quite nice discount.

Amazon? Maybe.
Definetly not micrococks.
You don't fuck someone over Early Purley and expect obidience thereafter.

There's a bunch more than those, and that's launch day.

Others won't be switching yet obviously because the arch is so new, give it a year.

Actually i think i brainfarted here.
Those companies listed actually jumped over to AMD side, cold blooded like that, right?

Is there really "sides" in business?

Yes.
Why the fuck do you think Intel devotes like 10 slides to LITERALLY shit on EPYC in some way or another?
They are panicking.

They are cheaper and do more,you can bet your ass they will "switch sides" like that.

when there is only two x86 companies, yes there are two sides
if market was more varied there won't be sides and actual competition would prevail
CPU tech, except for specific IPs like IF or intel hypoerthreading variant, must be made open source for humanity sake

Not fixed ones no, but you always can use the threat to squeeze more out of your current partner.
Or go coldblooded murder like those did.

You can license SMT from IBM and IF is AMD's secret sauce they will never disclose.
It's too good.

Ouch.

AMD gave the market what it wants.
EPYC does more for less, with lower TCO.
Let's not speak of AMD's 1S solutions.
And AMD can sell you fuckton of these, since yields are not a problem.
And margins on EPYC are going to be so stellar.