Yields of Intel Sapphire Rapids Processors Are Low, Mass Production to Start in 1H2023

[Daven]

I understood this news article as "the individial tile yields are only 50-60%" which has nothing to do with the packaging.

If the reality is that Foveros packaging itself is ruining perfectly good tiles, then Intel have an even bigger problem, because tile-based Foveros is their entire roadmap going forward. If that doesn't work, they have nothing to sell beyond what's already in the market right now!

That’s a great point. No matter how well Intel is executing outside the enterprise market, all products after Raptor Lake are tile based. The inability to make this work freezes all of Intel’s product lines.

 

[Punkenjoy]

I understood this news article as "the individial tile yields are only 50-60%" which has nothing to do with the packaging.

If the reality is that Foveros packaging itself is ruining perfectly good tiles, then Intel have an even bigger problem, because tile-based Foveros is their entire roadmap going forward. If that doesn't work, they have nothing to sell beyond what's already in the market right now!

well a 50% yields of such a large tiles can be expected, but it still very low. ADA is 600 mm2 vs 400 mm2 for those tiles. If they get 50% good die, then have to loose again more in packaging, that is just disastrous.

Also, by putting the IMC and other I/O in the same tiles, they have to waste silicon space of a better process on things that do not scale well.

 

[Fouquin]

Do you mean that they will release it in 70th week of this year, with the date code 2270?

There are 52 weeks in a year. SPR was supposed to be out somewhere in the April-June time frame of 2021; (end of) June 2021 until now is 70 weeks. At worst, it's been 79 weeks since we first expected SPR to launch.

 

[Wirko]

If the reality is that Foveros packaging itself is ruining perfectly good tiles, then Intel have an even bigger problem

Maybe you mean EMIB here. Or both. Sure, every advanced packaging process destroys some chips, including AMD's X3D (and I can only hope to ever get some data about the percentages).

There's another (possibly big) issue with such packaging. Can individual chiplets be tested and binned with good enough precision and reliability before they become part of a MCM?

Remember that, in Alder for example, there's an incredible number of SKUs in all shades of blue made from same silicon, to maximise the % of sellable parts with imperfections. Imperfections are common, they are not just defects but also inability of some cores to reach high frequency and/or low power.

Now does Intel build a 60-core monster with 4 tiles that are proven to have 15 good cores each, not just working but reaching, like, 3.2 GHz? Or is that a blind matching game? Or partially blind? Does packaging sometimes impair the characteristics of a tile without destroying it? Many questions.

Same goes for AMD of course, hence those 7600X chips with 2 chiplets.

 

[trsttte]

... by three manufacturers on two continents, and with three supply chains over all continents. Hm, is the Suez canal better managed now than it was in March of 2021?

I don't think that's really a factor, their CPUs already cross the world before being finished with a monolithic design. But integrating very different parts might indeed be a challenge

 

[Minus Infinity]

What about Ponte Vecchio? This computing unit is also super delayed and there hasn’t been a whisper about it in awhile.

That's why they were hyping it's successor. They have nothing good to report on PV.

AMD is about to release Zen 4 Epyc at 96 cores on Nov 10, 2022 with presumably Zen 4 Epyc-X with 3d v-cache on the way. As long as SPR is delayed, Intel’s best enterprise CPU is the 40 core dual socket Ice Lake based Xeon. I don’t see SPR catching up even with HBM.

But but accelerators!

Just to see how delayed SR is, Anandtech did the deep dive analysis in August 2021! And it seems it another 6 months until release at best.

 

[toddincabo]

This period of time we read
- bad news for AMD in the retail market (AM5 sales)
- bad news for Intel in the server market (SP)

At least there is a balance this way.

Not even close. I'm sure AMD would gladly give up more retail for server market if they could. Margins are that huge.

 

[john_]

For AMD they can just shift more CCDs towards EPYC.
For Intel their Xeons and Desktop chips are different dies so it is more of an issue.

AMD is doing this all these last years. But not in the extend I was thinking in the past. Their financial results showed that. Last quarter they made 2.2 billions from their client segment and probably they where hoping for over 1.8 billions this quarter. They only managed to pocket 1 billion instead. They probably sold much less GPUs than they expected, less APUs for laptops, considering I don't see Ryzen 6000 series to be the success most of us probably expected. And obviously it's not simply a switch where Ryzen chips are transformed into Epyc CPUs. You have to have demand for Epyc chips. In fact you have to have big orders, because Epyc chips are not meant to make money by just waiting on shop selves. Thankfully for AMD, those Sapphire Rapids delays continue, meaning increasing Epyc sales.

AMD Is still killing it in retail market with am4

I believe they are not. They only made 1 billion this quarter from the client segment. Whatever mine or your opinion is, the fact is that they made much less than they where expecting. AM4 sales will just give them an extra quarter or two where they can paint a more rosy picture. But selling AM4 upgrades wouldn't generate meaningful income for a company that is trying to become bigger. They have to fix this.

Not even close. I'm sure AMD would gladly give up more retail for server market if they could. Margins are that huge.

I am sure AMD would want server AND retail market. Their server and gaming(console) sales did kept them from sinking, because of their failure to compete in the retail market. And they DO need retail market. We can see this in their latest financial report. They probably made over half a billion less, maybe even close to a billion less than what they where expecting to make from their client segment.



In gaming, even selling better performing products than Nvidia at lower prices than Nvidia, doesn't seems to make a significant difference. People go and buy Nvidia cards blindly. And I don't believe that Raytracing performance and DLSS 2 alone is the reason driving people to buy overpriced Nvidia cards. AMD has not pushed it's latest RX 6000 series as it should. Tech sites and big YouTube tech channels had gone back in Nvidia's pockets lately. AMD needs to do more in marketing.
In laptops we see some models with 6000 Ryzen APUs in high end, but everything else is 5000 series or Intel. Many OEMs have turned to Intel to advertise higher number of cores/threads, even knowing that their newer products will be worst in battery, thermals, even performance.
In desktops AMD keeps selling CPU upgrades mostly for the AM4. The problem is that with upgrades people only buy the CPU. AMD is making money from also selling chipsets. But if people only upgrade and don't buy enough AM4 motherboards or jump to AM5, that means that AMD is not making enough profits from chipsets. And one day all those who wanted to upgrade will done their upgrades and then what?

 

[Wirko]

AMD is doing this all these last years. But not in the extend I was thinking in the past. Their financial results showed that. Last quarter they made 2.2 billions from their client segment and probably they where hoping for over 1.8 billions this quarter. They only managed to pocket 1 billion instead. They probably sold much less GPUs than they expected, less APUs for laptops, considering I don't see Ryzen 6000 series to be the success most of us probably expected. And obviously it's not simply a switch where Ryzen chips are transformed into Epyc CPUs. You have to have demand for Epyc chips. In fact you have to have big orders, because Epyc chips are not meant to make money by just waiting on shop selves. Thankfully for AMD, those Sapphire Rapids delays continue, meaning increasing Epyc sales.

I believe they are not. They only made 1 billion this quarter from the client segment. Whatever mine or your opinion is, the fact is that they made much less than they where expecting. AM4 sales will just give them an extra quarter or two where they can paint a more rosy picture. But selling AM4 upgrades wouldn't generate meaningful income for a company that is trying to become bigger. They have to fix this.

I am sure AMD would want server AND retail market. Their server and gaming(console) sales did kept them from sinking, because of their failure to compete in the retail market. And they DO need retail market. We can see this in their latest financial report. They probably made over half a billion less, maybe even close to a billion less than what they where expecting to make from their client segment.



In gaming, even selling better performing products than Nvidia at lower prices than Nvidia, doesn't seems to make a significant difference. People go and buy Nvidia cards blindly. And I don't believe that Raytracing performance and DLSS 2 alone is the reason driving people to buy overpriced Nvidia cards. AMD has not pushed it's latest RX 6000 series as it should. Tech sites and big YouTube tech channels had gone back in Nvidia's pockets lately. AMD needs to do more in marketing.
In laptops we see some models with 6000 Ryzen APUs in high end, but everything else is 5000 series or Intel. Many OEMs have turned to Intel to advertise higher number of cores/threads, even knowing that their newer products will be worst in battery, thermals, even performance.
In desktops AMD keeps selling CPU upgrades mostly for the AM4. The problem is that with upgrades people only buy the CPU. AMD is making money from also selling chipsets. But if people only upgrade and don't buy enough AM4 motherboards or jump to AM5, that means that AMD is not making enough profits from chipsets. And one day all those who wanted to upgrade will done their upgrades and then what?

I mostly agree but why do you think AM4 sales are only, or mostly, upgrades? People and companies buy new PCs all the time, at all price points. AMD has some new cheap chips in 2022, and deep price drops for old chips - an appropriate response, I think.

 

[stimpy88]

Oh shit, this is SERIOUS.

Sapphire Rapids is the first product using Intels Foveros "tile-based chiplets"

The point of this is to make smaller, easier-to-manufacture tiles, rather than big monolithic dies - because yields on smaller tiles are always going to be higher than on big monolithic dies.

Maybe I'm missing something here, but if yields are only 50-60% on small tiles, Intel's foundry execution is even worse than I thought, and I was already pretty cynical.

Sapphire Rapids largest tile is apparently 400mm^2. That's in the same ballpark as Navi22 (RX 6700XT) on TSMC N7 and GA104 (3070) on Samsung 8, so it's not as if yields should be that low....

Maybe Intel is having trouble with its glue...

 

[john_]

I mostly agree but why do you think AM4 sales are only, or mostly, upgrades? People and companies buy new PCs all the time, at all price points. AMD has some new cheap chips in 2022, and deep price drops for old chips - an appropriate response, I think.

Even a company or someone searching for an office PC will probably choose to pay for a new Intel platform than an old AMD platform, if the end price is not that much different. But let's say it goes with AM4. We still talk about a cheap office solution, so probably not enough money there for AMD to pocket. From upgrades AMD is losing the income from chipset sales and AM4 CPUs are already seeing reduction in their prices, meaning AMD will probably have to accept lower profit margin. Even if TSMC is lowering the cost for 7nm wafers, AMD is probably losing much more profit from those price reductions compare to what it saves from lower TSMC costs. So, unless they start selling R9 5950X by the dozens of thousands to AM4 owners, I can't see them making up from the disappointing launch of 7000. R7 5800X3D probably is not cheap enough to manufacture to hope that AMD is getting a nice profit margin from that chip. And how many X3Ds are they going to sell? AM4 users are not all gamers. In a few weeks we might see X3D sales slow down considerably.

 

[AleksandarK]

This needs some context. Is this the % of chips with all cores good, or % of chips with uncore logic OK and enough functional cores for Intel to not be ashamed of selling them?

Unclear, TrendForce reported it as such. I assume that it means a completely functional die.

 

[Hardware Geek]

Maybe Intel is having trouble with its glue...

I told them they should have used duct tape!