Solved! Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Joe NYC]

Oh they have competition, they managed to stave off mass hyperscaler exodus from x86 for now. Intel should be grateful to AMD on that one.

Very point.

Because they need to have healthy GM to invest in R&D. Semiconductor industry is capital intensive.

One variable that gives you better GMs is volume. If a design (that required a fixed amount of R&D resources to complete) sells 5 million units, it will have higher GMs than if it sells only 1 million units, and the design cost is spread over smaller number of units.

 

[Joe NYC]

I would say until AMD gets 50% market share in servers, anything goes to help them to that goal. Once there, I want competition ! I want to see Intel fight back ! Right now they can't in servers.

Edit: I hope the same for desktop, Ryzen 4. I want to see 50% market share.

Yup. The only way to have sustainable competition is when there is no single player with 80-90% market share. The sooner it gets corrected, the greater the likelihood of vigorous competition in the future.

 

[Markfw]

Yup. The only way to have sustainable competition is when there is no single player with 80-90% market share. The sooner it gets corrected, the greater the likelihood of vigorous competition in the future.

Exactly my point. I don't care who is in the lead, as I am not a fan of either company. But I want good competition. Right now servers is way lopsided, Intel has a larger base, but AMD has a better product and is selling more (I think). At least they are selling every one they make. And somebody pointed out there is no low end Ryzens, because AMD is using the chiplets for servers where the GM is much higher. I believe that. They just have to keep up the pace until they can get something close to 50% market share in all areas. I would love to see the resultant "war". Intel is trying hard with Alder lake, but there are problems.

 

[Joe NYC]

Ahh, that's a good point. Not all revenue gains has to come from more sold wafers, which makes sense since TSMC probably doesn't have more capacity anyways.

But TSMC is adding capacity. There will be more capacity in 2022 than there was in 2021.

Unlike Intel, which always kept the same capacity (more or less) just upgraded to newer and newer process nodes, TSMC runs the fabs almost indefinitely, and keeps adding more and more capacity on the new nodes.

And in addition to normal process, which would be to just expand by bringing on new capacity on N5 and N3, what TSMC is doing now (2021-2022) is also expanding their "old" 7nm and 28nm capacity.

So, Imagine Intel's 14nm fabs. TSMC builds nearly as much capacity on each new node and continues to run the old fabs on old nodes.

That's how TSMC got to > 50% market share.

 

[Joe NYC]

Intel 4 is low-volume and may never be used for anything outside of Meteor Lake and Granite Rapids.

But aren't these #1 and #3 highest volume products?

#1 Meteor Lake

(big gap)

#2 desktop
#3 Granite Rapids

What else would Intel be doing with their "Intel 4"

We will see how competitive these will be with AMD's Zen4 based on TSMC N5...

 

[DrMrLordX]

but AMD has a better product and is selling more (I think).

. . . not yet. Definitely not by revenue. Intel is still raking in a surprising amount of cash considering that most of their server shipments in 2021 were still Cascade Lake-SP.

But aren't these #1 and #3 highest volume products?

They aren't anything volume products, yet. And that volume will be constrained by Intel's own manufacturing problems. Unless they second-source it to N3.

 

[Joe NYC]

Exactly my point. I don't care who is in the lead, as I am not a fan of either company. But I want good competition. Right now servers is way lopsided, Intel has a larger base, but AMD has a better product and is selling more (I think). At least they are selling every one they make. And somebody pointed out there is no low end Ryzens, because AMD is using the chiplets for servers where the GM is much higher. I believe that. They just have to keep up the pace until they can get something close to 50% market share in all areas. I would love to see the resultant "war". Intel is trying hard with Alder lake, but there are problems.

There are different companies and methodologies to compute market share, and server market in particular, since it is the most strategic one. One of the analyses, by units, had AMD at 18% at the end of Q3, gaining 2% over previous quarter.

Extending the trend, while nothing meaningful changes in the market (SPR + Genoa likely shipping in Q3 in volume), AMD could be at:
Q4 2021: 20%
Q1 2022: 22%
Q2 2022: 25%

Unless something disruptive happens.

From there, Q3 22, AMD will be getting products from 2 buckets: N7 and Zen 4 on N5, both in volume. It's possible that in the server market, Intel dominance may be effectively over by the end of 2022 if AMD is at 30% in units, and greater than 30% in revenue.

One thing, in the category of "disruptive" may be excess capacity at TSMC at 7nm (possibly due to some cancellations of price sensitive customers after price increases).

It appears that TSMC did not sell out their N7 capacity last quarter (and none of the idiot analysts asked management why). It is possible that some extra capacity may have become available, and it fell in AMD's lap - which could accelerate the share gains.

 

[Markfw]

It just really surprises me that after years of EPYC beating Intel, that they are gaining so slowly. I mean the majort part of what Intel is shipping is 28 core cascade lake-SP that take as much power as the 64 core EPYC boxes ? Power/performance is critical in the data center, where twice the power is at least 4 times the cost, due to AC. It amazes me how stupid these data center people are.

 

[Joe NYC]

They aren't anything volume products, yet. And that volume will be constrained by Intel's own manufacturing problems. Unless they second-source it to N3.

Intel has secured N3 capacity from TSMC, but if Intel has to rely on TSMC for its mainstream processors, that will be a whole new set of troubles ahead for Intel.

I wonder how Intel breaks this to the investors in next week or two... Just a wild guess: There will be a lot of talk about Mobileye...

 

[jpiniero]

It just really surprises me that after years of EPYC beating Intel, that they are gaining so slowly. I mean the majort part of what Intel is shipping is 28 core cascade lake-SP that take as much power as the 64 core EPYC boxes ? Power/performance is critical in the data center, where twice the power is at least 4 times the cost, due to AC. It amazes me how stupid these data center people are.

You got to have product to be able to sell it.

 

[Markfw]

You got to have product to be able to sell it.

They have it. By the last quarterly report, they were shipping TONS of them. Now I hear there is spare 7nm capacity. You know they asked for all there is to spare... Its also available, Rome, Milan, all of them on ebay, newegg, Amazon... You know they are available. Why the post ?

 

[jpiniero]

They have it. By the last quarterly report, they were shipping TONS of them. Now I hear there is spare 7nm capacity. You know they asked for all there is to spare... Its also available, Rome, Milan, all of them on ebay, newegg, Amazon... You know they are available. Why the post ?

It's nowhere near enough.

 

[Joe NYC]

It just really surprises me that after years of EPYC beating Intel, that they are gaining so slowly. I mean the majort part of what Intel is shipping is 28 core cascade lake-SP that take as much power as the 64 core EPYC boxes ? Power/performance is critical in the data center, where twice the power is at least 4 times the cost, due to AC. It amazes me how stupid these data center people are.

AMD was just starting from such a small base in servers. Overall (for the whole company), there was one of the recent quarters when AMD grew 99% over the same quarter year before.

This level of growth is quite remarkable for products that have long lead time and in middle of shortage / supply chain disruptions.

But even doubling the shipments in a span of a year, if the starting point was, say 7%, it only got AMD to 14% in a span of an entire year.

 

[Joe NYC]

It's nowhere near enough.

There were a lot of rumors of long lead time to get Milan chips.

There are 2 ways to deal with shortages: queuing (long lead times) or price increases.

Maybe AMD realized that queuing (long lead times) makes people more unhappy than price increases...

 

[Joe NYC]

a bit of crazy/fun speculation from me. Recent leaks point to Ryzen using a Fan-out package, so it's possible Epyc is using it too. Looking at the leak from Gygabyte, I'm estimating the size of such fan-out area has to be around 31x65mm big. Looking online at what different companies offer, found an ASE fan-out packging page, in which they state that with "FOCoS", which is aimed at servers and networking, it's possible to make a package with a length of 67mm, barely enough for potential Epyc package. And on a standard wafer, about 20 such packages can be made. Edit: TSMC offers InFO_OS, which seems similar, and is qualified for 65x65mm packages in 2019.

View attachment 55674

Nice job with this.

There seems to be a quite a bit of empty space on 2 sides of this arrangement. It makes me wonder if AMD might be adding some additional silicon there. Could this support possibly putting some HBM memory in the package? Or some sort of FPGA or customized chips for things like encoding / decoding / networking...

If there is nothing else, just 12 chiplets connected to IO Die, is this the most optimized approach, given that there will be:
- additional cost to cost to the fan out package and assembly costs
- AMD still has the SerDes connection with its limitations
- bandwidth increase will only be linear,
- power savings may be limited.
- may limit the desktop / AM5 to higher market segments, no entry level CPUs

It seems that EMIB / EFB could be:
- cheaper
- no more SerDes, latency hit
- lower power use
- greater bandwidth potential

I wonder if this will end up an interim solution for a single generation.

Or, there may be more to this solution that we don't know yet. Something beyond just IOD-CCD links

One idea: Cutress had an article / video speculating that AMD could use interposer, which could provide a mesh interconnect between cores of the package, replacing the ring bus

Alternatively, quadrants could be NUMA, groups, and there could be a faster chiplet to chiplet interconnect within a quadrant constituting NUMA group.

Another AMD exec was hypothesizing about a mesh chiplet to chiplet interconnect, possibly providing very fast access to each other's L3 caches. But that could be quite messy, and if all the links would need a SerDes, also power hungry.

Alternatively, it could be only quadrant to quadrant mesh of SerDes based connections, as opposed to chiplet to chiplet...

 

[DrMrLordX]

It just really surprises me that after years of EPYC beating Intel, that they are gaining so slowly.

For a lot of datacentre people, buying AMD means waiting on a list. And there are vendors out there that still have loyalty to Intel, so if you've got a multi-year contract with them then you're stuck. Asking why datacentre people stick with Intel is like asking why AMD stuck with GF for so long.

 

[biostud]

But what about my RDRAM kit?!

Say it three times in front of a mirror to summon a horde of angry lawyers.

 

[Ajay]

One variable that gives you better GMs is volume. If a design (that required a fixed amount of R&D resources to complete) sells 5 million units, it will have higher GMs than if it sells only 1 million units, and the design cost is spread over smaller number of units.

Not necessarily. A company can reduced margins purposefully to to gain market share. AMD doesn't really have this option right now. Intel does to some extent (which they are doing in the enterprise server market) - well, to maintain customer loyalty in this case.

 

[scannall]

For a lot of datacentre people, buying AMD means waiting on a list. And there are vendors out there that still have loyalty to Intel, so if you've got a multi-year contract with them then you're stuck. Asking why datacentre people stick with Intel is like asking why AMD stuck with GF for so long.

Another part of it is the sales force. They don't just show shiny pages of products. They build relationships with the purchasing agent. A sales call is also visiting a friend, so you get that messy contract and paperwork out of the way so you can go have a nice lunch, talk about family. Your last fishing trip or vacation, things like that.

 

[Joe NYC]

Not necessarily. A company can reduced margins purposefully to to gain market share. AMD doesn't really have this option right now. Intel does to some extent (which they are doing in the enterprise server market) - well, to maintain customer loyalty in this case.

Well, when you discuss impact of one variable (market share, quantity of chips sold), you generally do it with assumption of "everything else being equal".

 

[Mopetar]

They have it. By the last quarterly report, they were shipping TONS of them. Now I hear there is spare 7nm capacity. You know they asked for all there is to spare... Its also available, Rome, Milan, all of them on ebay, newegg, Amazon... You know they are available. Why the post ?

Yes, they're shipping a lot, but if AMD is selling everything they have (which seems to be the case) then there's not a lot of room for them to grow unless they divert wafers from other product lines. The new consoles have eaten up a ton of wafers and judging from the fact that over a year after launch we still don't have any Zen 3 CPUs below the 3600X that AMD is prioritizing their servers at the expense of other product lines.

There's also a question of cost and replacement. If in the last three years you just purchased a large number of new Intel servers, they probably won't be replaced for a few more years. Even if AMD does offer better performance per dollar with Zen 3 in terms of both raw computation and power/cooling costs, it still may not come out ahead when considering the cost of replacing the existing hardware. Businesses are unlikely to replacing existing equipment that isn't at its end of life yet.

AMD has and continues to gain traction in that market and it's likely that many data centers are considering them if they hadn't before. But if there isn't hardware to replace right now or AMD can't guarantee it in the amount that customer needs, then it doesn't matter if they want to consider AMD or not. AMD's gains in the server market are quite impressive considering that they basically started from 0%. It likely won't be too long before they eclipse the market share that they managed with Opteron, which itself took several years to grow AMD's market share to similar levels to what we see now.

All of this is also being done during a time that's incredibly disruptive for the market. Keep in mind that the market itself is also expanding so it's not just a matter of AMD being able to take away sales that would have otherwise gone to Intel, but having to be able to constantly supply even more chips to keep up with that growth.

 

[biostud]

They have it. By the last quarterly report, they were shipping TONS of them. Now I hear there is spare 7nm capacity. You know they asked for all there is to spare... Its also available, Rome, Milan, all of them on ebay, newegg, Amazon... You know they are available. Why the post ?

Spare 7nm capacity? Then I would very much like them to ramp up RDNA2 production...

 

[Atari2600]

Spare 7nm capacity? Then I would very much like them to ramp up RDNA2 production...

I'm sure you would - and so would many others.

AMD's bean counters on the other hand, laugh at your opinion then drool at the thought of more EPYC dies!!

 

[Saylick]

I don't know what Pat is smoking but I want some.
https://www.tomshardware.com/uk/news/intel-ceo-gelsinger-says-amd-is-in-the-rear-view-mirror-after-alder-lake

"Alder Lake. All of a sudden...Boom! We are back in the game," exclaims the impish tech CEO. "AMD in the rearview mirror in clients [consumer market]," he adds, "and never again will they be in the windshield; we are just leading the market."

Pat is going to be in for a rude awakening when Zen 4 launches later this year.

 

[scannall]

I don't know what Pat is smoking but I want some.
https://www.tomshardware.com/uk/news/intel-ceo-gelsinger-says-amd-is-in-the-rear-view-mirror-after-alder-lake



Pat is going to be in for a rude awakening when Zen 4 launches later this year.

In all fairness he really can't say anything else. Though it probably would have been better to say nothing at all in that regard.