Intel’s Entire Turnaround Plan Hinges on Clearwater Forest, Intel’s First 18a Chip Slated for High-volume Manufacturing

[fansink]

Forget Lunar Lake; If there's one chip family that serves as a lynchpin for Intel’s entire turnaround plan, it's this – the company’s coming Clearwater Forest Xeon that it revealed for the first time at a recent event. The chip family is exceedingly important because this is the first high-volume chip to be fabbed on the Intel 18A process node, a node so critical that Intel CEO Pat Gelsinger has said he has bet the entire company on it.

Yes, Intel will make plenty of other processors with the 18A node, but mass-producing the Clearwater Forest chips on time is paramount to building confidence in Intel Foundry for potential customers, and that's the key to Gelsinger's entire turnaround plan. It also marks the culmination of Gelsinger’s audacious and now seemingly last-ditch effort to develop five nodes in four years to spark a resurgence at the ailing chipmaker.

Intel’s entire turnaround plan hinges on this one new chip family – Clearwater Forest pictured, Intel’s first 18A chip slated for high-volume manufacturing

Clearwater Forest is key to Intel’s future.

www.tomshardware.com

 

[Fred Chen]

Forget Lunar Lake; If there's one chip family that serves as a lynchpin for Intel’s entire turnaround plan, it's this – the company’s coming Clearwater Forest Xeon that it revealed for the first time at a recent event. The chip family is exceedingly important because this is the first high-volume chip to be fabbed on the Intel 18A process node, a node so critical that Intel CEO Pat Gelsinger has said he has bet the entire company on it.

Yes, Intel will make plenty of other processors with the 18A node, but mass-producing the Clearwater Forest chips on time is paramount to building confidence in Intel Foundry for potential customers, and that's the key to Gelsinger's entire turnaround plan. It also marks the culmination of Gelsinger’s audacious and now seemingly last-ditch effort to develop five nodes in four years to spark a resurgence at the ailing chipmaker.

Intel’s entire turnaround plan hinges on this one new chip family – Clearwater Forest pictured, Intel’s first 18A chip slated for high-volume manufacturing

Clearwater Forest is key to Intel’s future.

www.tomshardware.com

More important than Panther Lake?

 

[samwilde]

It is kind of bonkers betting the farm on a huge core Xeon design on a brand new process.
Intel always used to start the new process with the mobile chips.

 

[hist78]

I thought Intel 18A-related processors would be manufactured at Fab 52 in Chandler, Arizona. Has Intel completed Fab 52 yet?

 

[MKWVentures]

I thought Intel 18A-related processors would be manufactured at Fab 52 in Chandler, Arizona. Has Intel completed Fab 52 yet?

My Favorite question. What does Intel officially say?

 

[kevin01]

More important than Panther Lake?

I suspect that CWF (consisting of 12 computing tiles) will utilize some advanced packaging technology, which is not mature enough yet. In that sense, PTL will show if 18A is successful, and CWF will show if 18A and also advanced packaging are successful.

 

[kevin01]

It is kind of bonkers betting the farm on a huge core Xeon design on a brand new process.
Intel always used to start the new process with the mobile chips.

CWF is chiplet design, consisting of 12 computing tiles. https://hardwaretimes.com/intel-xeon-clearwater-forest-18a/amp/

 

[Daniel Nenni]

Forget Lunar Lake; If there's one chip family that serves as a lynchpin for Intel’s entire turnaround plan, it's this – the company’s coming Clearwater Forest Xeon that it revealed for the first time at a recent event. The chip family is exceedingly important because this is the first high-volume chip to be fabbed on the Intel 18A process node, a node so critical that Intel CEO Pat Gelsinger has said he has bet the entire company on it.

Yes, Intel will make plenty of other processors with the 18A node, but mass-producing the Clearwater Forest chips on time is paramount to building confidence in Intel Foundry for potential customers, and that's the key to Gelsinger's entire turnaround plan. It also marks the culmination of Gelsinger’s audacious and now seemingly last-ditch effort to develop five nodes in four years to spark a resurgence at the ailing chipmaker.

Intel’s entire turnaround plan hinges on this one new chip family – Clearwater Forest pictured, Intel’s first 18A chip slated for high-volume manufacturing

Clearwater Forest is key to Intel’s future.

www.tomshardware.com

Do we know how many die is on Intel 18A versus TSMC N3 or Intel 3?

Asked and answered, thank you Kevin:

Intel Xeon Clearwater Forest “E-Core” CPUs Consist of 17 Chiplets: 18A Compute, 3nm Base, and 7nm I/O Dies
Clearwater Forest will be a cloud-centric design, the successor to Sierra Forest, which is slated to land later this year. It’ll feature up to 288 cores leveraging the Atom-class Darkmont architecture. Like Lunar Lake’s compute tile, these “Darkmont” E-cores will be spread across 12 chiplets (or tiles) fabbed on the Intel 18A process.

The 18A compute tiles will be 3D stacked on three base tiles fabbed on the Intel 3 node in groups of four. The base tiles will hold the CPU L3 cache and VRMs, components getting minimal gains with node shrinks. Finally, the I/O dies will be placed on the top and bottom ends of the package. Fabbed on the Intel 7 process, it’ll control the PCIe, CXL, memory, and other links crucial for connectivity and inter-chip data transfer.

So no TSMC? That would be quite impressive!

 

[siliconbruh999]

I'm guessing Clearwater Forest is chiplet based? Do we know how many die is on Intel 18A versus TSMC N3 or Intel 3?

Asked and answered, thank you Kevin:
Intel Xeon Clearwater Forest “E-Core” CPUs Consist of 17 Chiplets: 18A Compute, 3nm Base, and 7nm I/O Dies
Clearwater Forest will be a cloud-centric design, the successor to Sierra Forest, which is slated to land later this year. It’ll feature up to 288 cores leveraging the Atom-class Darkmont architecture. Like Lunar Lake’s compute tile, these “Darkmont” E-cores will be spread across 12 chiplets (or tiles) fabbed on the Intel 18A process.

The 18A compute tiles will be 3D stacked on three base tiles fabbed on the Intel 3 node in groups of four. The base tiles will hold the CPU L3 cache and VRMs, components getting minimal gains with node shrinks. Finally, the I/O dies will be placed on the top and bottom ends of the package. Fabbed on the Intel 7 process, it’ll control the PCIe, CXL, memory, and other links crucial for connectivity and inter-chip data transfer.

So no TSMC? That would be quite impressive!

Yes all Intel 3/18A/7 here is the appearance for Clearwater Forest

 

[MKWVentures]

Datacenter chips are very slow to ramp.... and the fact that Intel has released 3 DC products in the last year makes it even more difficult.

Hence, Clearwater forest will not determine 18A.

The current paradox Intel faces is that ramping 18A is not good financially and NOT ramping 18A is not good financially. There are some options but they all have tradeoffs.

Other comments in the article are not accurate interpretations of what decisions were made in the last 15 years either.... but that is water under the bridge.

 

[kevin01]

Datacenter chips are very slow to ramp.... and the fact that Intel has released 3 DC products in the last year makes it even more difficult.

Hence, Clearwater forest will not determine 18A.

The current paradox Intel faces is that ramping 18A is not good financially and NOT ramping 18A is not good financially. There are some options but they all have tradeoffs.

Other comments in the article are not accurate interpretations of what decisions were made in the last 15 years either.... but that is water under the bridge.

Mind explaining why DC chips are very slow to ramp? I can think of two things: 1. DC chips yield are lower as they usually have big die sizes, 2. Customers are slow to make purchase decisions, as it takes a while to validate these new chips, especially when the chips are launched on a new platform.

 

[MKWVentures]

Mind explaining why DC chips are very slow to ramp? I can think of two things: 1. DC chips yield are lower as they usually have big die sizes, 2. Customers are slow to make purchase decisions, as it takes a while to validate these new chips, especially when the chips are launched on a new platform.

DC has always been a slow business. They are changing multi billion dollar facilities and they often dont want mix and match systems. So they sample, validate, do a limited pilot, optimize, plan for expansion, etc. If Intel sells 15M server CPUs it would be typical to sell 1-2M in the first year of a new product. If it is REALLY over the top high nd like Clearwater forest seems to be ... it might take 2 years to sell 1M units. Intel doesnt know all the skus they will do but they will make adjustments.

The market often PEAKs for a DC CPU about 2 years after launch.

With Intels new plan of lots of processor releases, plus the DC issues with spending a ton on nvidia GPUS, it could make things worse..... Even Intel is not sure how the response will go.

Same is true for GPUs, HDD, SSD.

AI servers took off like a rocket but I would expect that to slow in ramp speed as well. We will see on broadwell.

again when you do the math, its well under a fab of capacity even when fully ramped.... plus you cannot force DC to convert using price as easily like you can with client.
Client+foundry is needed to make Intel win on 18A IMO.

 

[kevin01]

DC has always been a slow business. They are changing multi billion dollar facilities and they often dont want mix and match systems. So they sample, validate, do a limited pilot, optimize, plan for expansion, etc. If Intel sells 15M server CPUs it would be typical to sell 1-2M in the first year of a new product. If it is REALLY over the top high nd like Clearwater forest seems to be ... it might take 2 years to sell 1M units. Intel doesnt know all the skus they will do but they will make adjustments.

The market often PEAKs for a DC CPU about 2 years after launch.

With Intels new plan of lots of processor releases, plus the DC issues with spending a ton on nvidia GPUS, it could make things worse..... Even Intel is not sure how the response will go.

Same is true for GPUs, HDD, SSD.

AI servers took off like a rocket but I would expect that to slow in ramp speed as well. We will see on broadwell.

again when you do the math, its well under a fab of capacity even when fully ramped.... plus you cannot force DC to convert using price as easily like you can with client.
Client+foundry is needed to make Intel win on 18A IMO.

Thanks for explaining!

I agree that smaller, higher volume chips are better customers for foundries, examples are PC cpu chips (e.g. Panther Lake), smartphone chips, etc.

It would be interesting to see if Samsung or Qualcomm could become a 18A, or, more likely, 14A customer.

Another source of potential customers could be hyperscalers, either buying custom x86 chips, or making their own chips. They are more nimble than typical enterprise customers, as hyperscalers are more technically capable.

 

[TICin852]

It is certainly interesting if you look at mobile chips in the way that Intel seems to be releasing them.

Intel Core Ultra is a dumpster fire of a brand - but (speculation on my part) maybe it was renamed because Pat wanted to focus on getting the Tick-Tock strategy going again.

Core Ultra 100 (Meteor Lake) - Arguably a Tick, bringing in the Intel 4/3 node and noted as really being a die shrink - at least Redwood Cove, Intel's P cores. Released Q42023/Q12024

Core Ultra 200 (Lunar Lake) - Tock - a whole new microarch and layout of the CPU, made on N3 - different node, but still about the same in terms of node class. Arguably, I would point to TSMC being the "Trusted Foundry" and Apple already having products on N3 as ensuring the node worked. Released Q32024, about 9 months after a late Meteor Lake.

Core Ultra 300 (Panther Lake) - Tick? I am unsure of the architectural changes involved here, and don't try to keep up with the rumors on it - but I would assume that Intel would be rather conservative on microarch changes to prevent massive failures. Releasing 2025, likely 2H.

If Intel is able to do this successfully, I'd say Pat seems to have gotten them back to launching competitive products regularly, on time - being able to have CWF and PNL on time is critical for Pat's survival as CEO, seemingly moving Intel back to being an execution machine.

The other important aspect of financials - as others have noted - is a completely different thing that I won't get in to, but Pat's survival as CEO is clearly linked to it. Seems Pat's hope for survival is that competitive products and successful execution (in addition to foundry) convince the board that financials will eventually improve in 2025 that they stick through with him until 2027.

 

[MKWVentures]

Now that Intel is releasing multiple products for Server and client (I think this should be their total focus). It will be interesting to see how conversions and tiering go. Sapphire rapids has just peaked in volume and Intel has released Emerald, Sierra Forest, now Granite in last year. Lots of choices for customers.... product group is delivering. I would still expect Emerald and Sapphire rapids to outsell granite and sierra Forest over the next 12 months but lets see what the uptake is.

 

[siliconbruh999]

Now that Intel is releasing multiple products for Server and client (I think this should be their total focus). It will be interesting to see how conversions and tiering go. Sapphire rapids has just peaked in volume and Intel has released Emerald, Sierra Forest, now Granite in last year. Lots of choices for customers.... product group is delivering. I would still expect Emerald and Sapphire rapids to outsell granite and sierra Forest over the next 12 months but lets see what the uptake is.

I would expect the reverse no one will buy those unless at massive discounts
GNR/SRF are actually good products unlike SRF/EMR

 

[MKWVentures]

I would expect the reverse no one will buy those unless at massive discounts
GNR/SRF are actually good products unlike SRF/EMR

DC Product conversions and ramps do not work like that. but lets see what happens.

 

[OneEng]

Datacenter chips are very slow to ramp.... and the fact that Intel has released 3 DC products in the last year makes it even more difficult.

Hence, Clearwater forest will not determine 18A.

The current paradox Intel faces is that ramping 18A is not good financially and NOT ramping 18A is not good financially. There are some options but they all have tradeoffs.

Other comments in the article are not accurate interpretations of what decisions were made in the last 15 years either.... but that is water under the bridge.

I am guessing you make this statement because you believe (maybe rightfully so) that Intel 18A yields are quite bad. If this is your assertion, then there is certainly some evidence to support this from Samsung's GAA experience on 3nm.

So the Clearwater forest tiles are decently big and being produced on a brand new process at Intel. That does indeed seem quite risky. GAA and BSPD all with a new server chip. Kind-of a "Tock-Tock-Tick" all in one generation .

Intel once could do no wrong in the process side. They kept 1 to 2 full process nodes ahead of anyone else. I used to argue that it doesn't take much of a CPU designer to create a superior CPU when you have twice the transistor budget, and double the power envelope to do it in.

Give the most recent process history of 14+++++++ and 10++++++ at Intel, and the most recent drop of 20A (de-risking node) I guess I feel like it is pretty hard to believe that 18A and CWF will be on time. To your point, even if they are, it might cost them so much to produce that it hurts the bottom line.

Intel 3 seems to have decent characteristics, but I have no idea what their yields are on that process.

 

[OneEng]

Now that Intel is releasing multiple products for Server and client (I think this should be their total focus). It will be interesting to see how conversions and tiering go. Sapphire rapids has just peaked in volume and Intel has released Emerald, Sierra Forest, now Granite in last year. Lots of choices for customers.... product group is delivering. I would still expect Emerald and Sapphire rapids to outsell granite and sierra Forest over the next 12 months but lets see what the uptake is.

I am not sure the business case is there for that. Many a "big iron" processing company bit the dust trying to do exactly that. Maybe the world has changed though and the market along with it.

Still, conventional wisdom is that you produce lower volume less cost effective designs on a new process node first where you can get very high margins for each chip produced. Once the process matures, you use the process and core technology trimmed down for your desktop and mobile designs.

Of course, that was how it USED to be done.

Seems now days that even phone processors (Apple) are out there on the cutting edge process node (or close to it). This is an insanely high volume market compared to computers.

 

[count]

I am not sure the business case is there for that. Many a "big iron" processing company bit the dust trying to do exactly that. Maybe the world has changed though and the market along with it.

Still, conventional wisdom is that you produce lower volume less cost effective designs on a new process node first where you can get very high margins for each chip produced. Once the process matures, you use the process and core technology trimmed down for your desktop and mobile designs.

Of course, that was how it USED to be done.

Seems now days that even phone processors (Apple) are out there on the cutting edge process node (or close to it). This is an insanely high volume market compared to computers.

It’s usually the other way around. You produce smaller tiles first because your yield will be higher for smaller tiles (defects increase to the square of die size). As you ramp and improve yields you start making bigger tiles.