[EETimes] TSMC starting 5 nm EUV Risk Production in April?

[moinmoin]

@krumme do you mean Zen 2 or are you just looking ahead further?

One thing to keep in mind with servers is that even though Intel's prices may seem outrageous at first, due to the high cost for memory and other system components the CPU prices are comparatively negligible. So while AMD's Epyc CPUs are clearly better positioned price/performance wise, they don't do much about the overall server cost. That why AMD so far focused on TOC due to offering more cores in less sockets (which affects common per socket software licensing cost).
With Epyc 2 this will change, AMD now likely to offer equal performance per core as Intel. As a result we already see Intel now focusing on their remaining USPs (like AVX512, Optane and their own approach to interconnects).

 

[Yotsugi]

One thing to keep in mind with servers is that even though Intel's prices may seem outrageous at first, due to the high cost for memory and other system components the CPU prices are comparatively negligible

Plus volume discounts.

AMD now likely to offer equal performance per core as Intel

Also more cores and I/O, plus hands down the craziest 1p/2p platform ever made with commodity h/w in mind.

like AVX512

The ISA is nice, but still niche.

Optane

Nice, but again, niche and a whole list of caveats so far.

their own approach to interconnects

The what?
They're behind in both PCIe and socket to socket speeds and feeds versus Rome.
xGMI v2 is almost 2.5 times the bandwidth per lane over UPI.

 

[moinmoin]

Also more cores and I/O, plus hands down the craziest 1p/2p platform ever made with commodity h/w in mind.

That's down to the server board manufacturer to exploit, which didn't happen that much so far and often comes with a premium. For companies designing servers in-house that's a definitive advantage though, I think Epyc already widespread availability from many cloud services is a testament to that.

The what?
They're behind in both PCIe and socket to socket speeds and feeds versus Rome.
xGMI v2 is almost 2.5 times the bandwidth per lane over UPI.

Yeah, with Zen 2 I/O should be clearly in favor of AMD in pretty much all areas. Nevertheless Intel is seen pushing their Compute Express Link (instead joining CCIX), I guess some potential customers will still be impressed by that. *shrugs*

 

[Yotsugi]

That's down to the server board manufacturer to exploit,

Which they will, Rome makes for a killer accelerator host and that's not even my words.

Intel is seen pushing their Compute Express Link

AMD can easily join the CXL group, and it takes exactly one I/O spin for them to start shipping CXL soon thereafter.

 

[moinmoin]

AMD can easily join the CXL group, and it takes exactly one I/O spin for them to start shipping CXL soon thereafter.

If Intel allows AMD to join I'm sure AMD can easily add support for it. So far CXL just looks like Intel reacting to Gen-Z and CCIX. Market wise the latter two have far more widespread backing, and AMD is already part of them. We will see how much CXL will matter, Intel certainly likes to push it (like so many of their tech they have announced but have not yet released).

 

[Yotsugi]

If Intel allows AMD to join I'm sure AMD can easily add support for it.

It's an open group.

 

[moinmoin]

It's an open group.

So? The few members of it looks like Intel and its biggest customers. Why would AMD help push it when the good parts of it are better introduced to Gen-Z and CCIX?

 

[Yotsugi]

The few members of it looks like Intel and its biggest customers.

That's already the lion's share of DC space.

Why would AMD help push it when the good parts of it are better introduced to Gen-Z and CCIX?

AMD holds no allegiance and does anything if the customer demand is there.
If customers want CXL, they will get it.

 

[moinmoin]

That's already the lion's share of DC space.

All of which are already working with AMD as well, most of which already participate in Gen-Z and CCIX as well.

AMD holds no allegiance and does anything if the customer demand is there.
If customers want CXL, they will get it.

Indeed. But why should AMD join before customers want CXL? Both Gen-Z and CCIX are older and more importantly both have multiple CPU manufacturer backing it whereas Intel is the sole one for CXL. Do you want AMD to help Intel justifying CXL?

 

[Yotsugi]

All of which are already working with AMD as well, most of which already participate in Gen-Z and CCIX as well.

Not really relevant.

But why should AMD join before customers want CXL?

Why not?

Both Gen-Z and CCIX are older and more importantly both have multiple CPU manufacturer backing it whereas Intel is the sole one for CXL

Multiple irrelevant CPU vendors versus the one and only relevant one.

Do you want AMD to help Intel justifying CXL?

Whatever, as long as either CCIX or CXL dies and the market goes with the winner.

 

[Ajay]

@krumme do you mean Zen 2 or are you just looking ahead further?

I think he meant Zen4, since this thread is on TSMC 5nm. Zen3 will be on 7nm+.

Edit: brain fart.

 

[maddie]

I think he meant Zen3, since this thread is on TSMC 5nm. Zen2 will be on 7nm+.

What?

Zen2 is on 7nm and 7nm+ will not be ready for AMD until next year.

 

[Ajay]

What?

Zen2 is on 7nm and 7nm+ will not be ready for AMD until next year.

Thanks for catching my mistake.

 

[krumme]

@krumme do you mean Zen 2 or are you just looking ahead further?

One thing to keep in mind with servers is that even though Intel's prices may seem outrageous at first, due to the high cost for memory and other system components the CPU prices are comparatively negligible. So while AMD's Epyc CPUs are clearly better positioned price/performance wise, they don't do much about the overall server cost. That why AMD so far focused on TOC due to offering more cores in less sockets (which affects common per socket software licensing cost).
With Epyc 2 this will change, AMD now likely to offer equal performance per core as Intel. As a result we already see Intel now focusing on their remaining USPs (like AVX512, Optane and their own approach to interconnects).

Meant zen 2. Sorry for not being precise enough.

I worked with tco for a few years many years back so know a bit about it so I know where you are going.

Still read my poins. There is a change in market even a 5% matters. Every penny counts.

And yes Intel knows and is already targeting more precisely other markets.

 

[name99]

If this is true, then surely 5nm EUV for TSMC is pushed back as well. Seems kind of strange to announce risk production for a lithography process that is now likely two years out (unless their big foundry customers are all skipping 7nm+.

That's because you don't understand the foundry business model.
TSMC does not sell one leading edge process and done! They sell a variety of processes, and part of the sale includes continual upgrades of the non-leading edge processes, so that even if you don't have the budget to fully redesign and run a new mask set for the newest process, you may be able to get some improvement (slightly denser, or lower power, or faster) though the much smaller modifications required to move from say 7nm to 7nm+ or 6nm.
Just because these sorts of improvements (like ongoing improvements at 10 or 16, or even 28nm) are not of interest to, say, Apple, doesn't mean they're of no interest to anyone!

 

[Hans de Vries]

From the Q1, TSMC, conference call:

https://seekingalpha.com/article/42...-wei-q1-2019-results-earnings-call-transcript

C. C. Wei:

I want to reassure everybody that N5 as a business will grow bigger than N7 because our expanding the product portfolio, and you know what I mean. In the HPC, we have very good opportunities and on the smartphone we are gaining market share.

 

[Hans de Vries]

From the Q1, TSMC, conference call:

https://seekingalpha.com/article/42...-wei-q1-2019-results-earnings-call-transcript

C. C. Wei:

I want to reassure everybody that N5 as a business will grow bigger than N7 because our expanding the product portfolio, and you know what I mean. In the HPC, we have very good opportunities and on the smartphone we are gaining market share.

"HPC" would point to a very high volume expected for AMD on 5nm ....

They also currently have an all time high N7 inventory build-up for a product to be released.

 

[DrMrLordX]

"HPC" would point to a very high volume expected for AMD on 5nm ....

Is IBM looking to use TSMC 5nm for a future POWER product?

 

[jpiniero]

"HPC" would point to a very high volume expected for AMD on 5nm ....

Could be several companies, even Intel. Intel might actually be using Samsung for the GPUs though.

 

[Hans de Vries]

I don't have any additional information but since TSMC produces hunderds of millions of high end SOC's per year on N7 and they expect a sizable increase of that number on N5

Is IBM looking to use TSMC 5nm for a future POWER product?

IBM POWER is extremely low volume when compared with the smartphone market in both numbers and silicon revenue.

 

[jpiniero]

To me, it would make sense if Intel fabbed at both companies. Maybe something like GPU tiles at Samsung and CPU tiles at TSMC.

 

[Ajay]

To me, it would make sense if Intel fabbed at both companies. Maybe something like GPU tiles at Samsung and CPU tiles at TSMC.

GPU, at least to start, makes sense. But, I’ve mentioned b4, no other fab can replace the 8(?) fabs that Intel has. Intel gets 7nm EUV right or it’s the beginning of the end of their existence as a major semiconductor company.

 

[krumme]

To me, it would make sense if Intel fabbed at both companies. Maybe something like GPU tiles at Samsung and CPU tiles at TSMC.

Well Intel identify themselves as foremost a fab company.
Fastest densest and cheapest process.
4 years ago in this sub you were flat out ridiculed if you said it wasn't so rosy.
I wouldn't expect their cultural identity to change without some big bang.

 

[JasonLD]

Is IBM looking to use TSMC 5nm for a future POWER product?

Afaik, Power 10 is going to be on Samsung's 7nm EUV, which is in line with its expected release in 2020.

 

[maddie]

Well Intel identify themselves as foremost a fab company.
Fastest densest and cheapest process.
4 years ago in this sub you were flat out ridiculed if you said it wasn't so rosy.
I wouldn't expect their cultural identity to change without some big bang.

Bizzarro world indeed. Intel possibly fabbing die at TSMC and Samsung,and being discussed as such an ordinary event. Wake me up.