Discussion Intel current and future Lakes & Rapids thread

[Joe NYC]

Cause DMR-HD fixes the issue with Clear water forest of not having AVX-512 and botched Mesh/L3

Broadwell eDRAM Lakefield for foveros and Kaby Lake G for EMIB

I don't think the "Foveros Direct" - hybrind bond variety of Foveros was available until maybe last year or two.

 

[511]

I don't think the "Foveros Direct" - hybrind bond variety of Foveros was available until maybe last year or two.

Yeah but they have been working on HBI for some time it's the reason for delay of Clearwater forest.

 

[msj10]

If Coral was originally scheduled to be on 14A, then either Coral will have to (instead) target 18A or 14A needs some acceleration too. Or at minimum, no delays.

I wouldn't be surprised if It's N2.

 

[511]

I would be surprised if it's N2 though it would likely reuse DMR's platform a lot if they want increased Time to Market

 

[Geddagod]

I wouldn't be surprised if It's N2.

I used to think it was unified core, but if it's not, it being external might not make it on par/better than AMD's platform anyway. Worse core IP and all.
I think Intel only goes external if they think they have a chance on being on par or better than the competition. Otherwise, just compete on costs and make it internal for better margins.
Also, I think if Intel does go external for anything in the 28' timeframe, it would be on A16.

 

[Joe NYC]

I wouldn't be surprised if It's N2.

Intel has so far kept the server CPUs in Intel's own foundry. It would be a very bad advertisement for Intel Foundry if server chips were to switch to TSMC.

 

[adroc_thurston]

Worse core IP and all.

Worse literally everything.

Intel has so far kept the server CPUs in Intel's own foundry

SRF was TSM N3 originally.

 

[msj10]

I used to think it was unified core, but if it's not, it being external might not make it on par/better than AMD's platform anyway. Worse core IP and all.
I think Intel only goes external if they think they have a chance on being on par or better than the competition. Otherwise, just compete on costs and make it internal for better margins.
Also, I think if Intel does go external for anything in the 28' timeframe, it would be on A16.

idk they seem pretty positive about coral rapids in their earnings calls and all of the negativity surrounding DMR seem to be centered around the lack of SMT.

 

[511]

idk they seem pretty positive about coral rapids in their earnings calls and all of the negativity surrounding DMR seem to be centered around the lack of SMT.

cause it is not having SMT hurts in servers though there are servers without SMT where DMR might perform decently well

I used to think it was unified core, but if it's not, it being external might not make it on par/better than AMD's platform anyway. Worse core IP and all.

Doubt Griffin Cove is as worse as having 3 gens of Golden Cove it's another tock

I think Intel only goes external if they think they have a chance on being on par or better than the competition. Otherwise, just compete on costs and make it internal for better margins.

yeah

Also, I think if Intel does go external for anything in the 28' timeframe, it would be on A16.

might as well use 14A if it's 28

 

[DrMrLordX]

all of the negativity surrounding DMR seem to be centered around the lack of SMT.

Certainly there must be some other things wrong with the product. SMT alone wouldn't explain why DMR-SP is essentially dead.

 

[wayner.]

Certainly there must be some other things wrong with the product. SMT alone wouldn't explain why DMR-SP is essentially dead.

Agreed. SMT alone can’t explain it. DMR-SP seems to suffer from a mix of issues: weak perf/W versus Genoa, platform cost and complexity, late availability, and limited real-world uplift. When power, pricing, and ecosystem all lag, even decent core counts don’t save the product.

 

[adroc_thurston]

Agreed. SMT alone can’t explain it. DMR-SP seems to suffer from a mix of issues: weak perf/W versus Genoa, platform cost and complexity, late availability, and limited real-world uplift. When power, pricing, and ecosystem all lag, even decent core counts don’t save the product.

it suffers from toiletbowl-tier hybrid bonding yields.