Zen 5nm

What if Epyc on 5nm will pack 128 cores?

Attached: 1527690603812.jpg (494x345, 29K)

heat efficient spacing

Amd will likely segment the market and improve yields untill the market is saturated. Then they'll release 64 cores and eventually 128.

They told us they will focus on 5nm . Epyc yields are pretty good already from what I have heard (which is the main reason AMD doubled their flagship CPU's core count in a year,much to everyone's surprise)

>5nm 128 cores

unlikely for the simple reason that even 8 channel DDR5 won't be nearly enough bandwidth.

maybe something like 96 (24c * 4) with assloads of L3 cache and more AVX could happen though.

I'm not sure, but the prospect of Zen 2 coming with 6c-8c CCXs with 4 MB L3/core makes my dick diamonds.

We could realistically see CCXs with as much L3 as an entire 28c Xeon Platinum (24.75 MB L3, for $9k-$13) today. Each.
And

Ok, this is Epyc

TSMC in 2020 5nm
>Intel my 10nm are equal to 7nm TSMC.

Intel 10nm densities are still virtually speculative/fictitious until they can actually make fully enabled 100mm^2 chips with >10% yields.
It took them half a year to make

dude that would be epic lmao

It would be an epyc win

you missed the newest one, they had to fuck around with 10nm to save face, and now it loos more like a 12nm instead

Anyone seen the upcoming Zen SOC with 24 Vega CU's??

Imagine those inside tablets and netbooks holymoly mastacholli.
Maybe some mini ITX's or game consoles, SBCs.

Thank you mummy Su

Not to mention Intel's market entrance is supposedly dated in "holidays 2019" meaning the end of the damn year in novemeber/december

Intel really fucked up badly..

Yeah, its for the Chinese market for domestic consoles. You'll probably never see one.

They're already planning on releasing 64 core on EYPC with 7nm.

Time to do something about the memory wall, then?

But if they're good, western brands will also commission similar custom chips for gaming tablets and such.

Just a reminder that zen so called 7nm is actually bigger than Intel 10nm.
7nm or whatever shit global foundries want to say is just market glimmick

They'll probably make it wider (8-issue or more) and add SMT4. The memory wall won't let them add more than 64 cores per socket.

Some will get sold on aliexpress or alibaba.

a monolithic design will never reach the yields to beat amd

I think we are at least 3 years off from 128 cores. I think we may see 40-56 cores by 2020. I think the main improvements of 7nm Wil be to speed and efficiency. Intel is fucked until 2022 (as the new reports are revealing. This is likely due go the need to develop proprietary MCM's vs monolithic dies (with their famously low yields on high core parts.) Intel still has plenty of gas in the tank for consumer parts for the next 3 years or so.l, but are fucked in the prosumer department.

it literally isn't, even before they had to gut it and turn it into 12nm

it could, but you'd need to be almost a decade ahead of the competition on your foundries

If AMD sticks to the same basic model for their HEDT/server chips, then we can always expect the flagships to have a core count that's a multiple of 4. While 56 is possible (14/die), 14 is a really weird number for a single die (since we're talking the very top-of-the-line, I'll ignore the disabled die thing). I would say 48 is more likely, but 56 certainly isn't out of the question.

>I think we may see 40-56 cores by 2020.

Nigger we're seeing at least 48c by next Summer and 64c by Summer 2020 at the latest.
The main question is whether 7nm yields are good enough that AMD goes straight for the jugular with 16c Zen 2 parts next year. Even with 4 MB L3/core, such a part at 7 nm wouldn't be all that much bigger than Zeppelin. And remember that Rome was designed to fight successfully against 10 nm Ice Lake in 2019 at somewhere in the 60-70 core range.

it wont amd is already up to 90%+

At best, 2P Servers with Zen 2 can deliver 128 cores since it may have 8 core CCX

is 5nm even a real node and not 7nm with some marketing-sprinkles?

>They'll probably make it wider (8-issue or more) and add SMT4.
given that this is even at all possible, why would they bother?

>intel makes glued cpus
>now intel will automatically also be affected by latency
BASED

not necessarily
they've got their EMIB shit

Which is inferior to infinity fabric, IT'S EVEN WORSE

they have to spend a lot of billions to their fabs to get their shit together
thats the problem
otherwise they gonna force to outsource it

Those aren't the same things. EMIB is meant to replace TSVs. They're just smaller and faster versions of it. Unlike AMD, they aren't going to put giant spaces between dies for it. It can also be used for 3D die stacking ( cache on top of the cores instead of adjacent ).

I think it'd drink power like a motherfucker
would need the cooling capacity of a 45W mobile workstation

epyc, epyc for the win!!!

Having a CPU 4 times faster than your competitor's is a great case for its adoption. And would embarrass Intel even more.

since there is no standard for it everything is just a marketing shite