News Intel Bartlett Lake-S: up to 12P-Core or up to 8P-Core +16E-core

[Thunder 57]

Well, looking at my post, I missed several words that would clarify my point. As late as Bartlett Lake is to market, and given that Intel has all the needed Xtor IP blocks for Intel3 for the same P and E cores as well as all the needed IO, it's almost silly that they didn't just make Bartlett Lake on Intel3 with 12 P cores. It might not have moved the needle a whole lot for 1t as I'm not sure they could have clocked it much higher, but, the nT benchmarks would have loved the improved node, and everyone would have loved the better Perf/Watt numbers, especially utility scale buildouts that have to pay for all that power usage over the long haul...

Gracemont was made on Intel 3? Also, the cost of designing a new die with Raptor Cove + whatever E core on Intel 3 for what would be a niche market at best doesn't seem like it would have a good ROI and LBT seems to want 50%+ margins which wouldn't happen.

 

[LightningZ71]

No, Meteor Lake's CPU core die was on Intel 4, then ported to Intel 3 for Arrow Lake refresh. Meteor Lake's cores are Redwood Cove, which are barely touched raptor cove cores for the P core and a VERY mildly tweaked Gracemont called Crestmont for the e cores. To do a 12 P core die for Bartlett Lake starting from Raptor Lake STILL requires changes to the floorplan and power delivery on the Raptor Cove die, which is NOT trivial. If you're going to do all that work, and you've already got the needed blocks for Intel 3, why are you using Intel7?

 

[Thunder 57]

No, Meteor Lake's CPU core die was on Intel 4, then ported to Intel 3 for Arrow Lake refresh. Meteor Lake's cores are Redwood Cove, which are barely touched raptor cove cores for the P core and a VERY mildly tweaked Gracemont called Crestmont for the e cores. To do a 12 P core die for Bartlett Lake starting from Raptor Lake STILL requires changes to the floorplan and power delivery on the Raptor Cove die, which is NOT trivial. If you're going to do all that work, and you've already got the needed blocks for Intel 3, why are you using Intel7?

Well I had a brain fart since this mythical Bartlett Lake 12 P core wouldn't have E cores. Derp. It would be a nice swan song for LGA 1700 but I think Intel has other priorities. No ODM's would use it so Intel would have to make money on a small market of enthusiets. It just doesn't seem to be worth the effort.

 

[LightningZ71]

We say that and yet Intel just introduced a new Comet Lake SKU the other day. Stranger things have happened.

 

[jpiniero]

If you're going to do all that work, and you've already got the needed blocks for Intel 3, why are you using Intel7?

It's cheap and they have the capacity for it (unlike Intel 3)

 

[LightningZ71]

While they are most likely not as capacity constrained on Intel7 as they are on Intel3, it is most certainly NOT cheap. Intel 7 apparently requires a LOT of multi-patterning to get their DUV machines to give them the power/performance/density targets that they want. In addition, to achieve the clock speeds that they managed with Raptor Lake, they had to relax their density requirements multiple times on what was originally 10esf+. It's not area efficient, it's not power efficient on the end of the V/f curve that they play in to get their performance numbers, and as a result, it's not cheap.

 

[jpiniero]

While they are most likely not as capacity constrained on Intel7 as they are on Intel3, it is most certainly NOT cheap. Intel 7 apparently requires a LOT of multi-patterning to get their DUV machines to give them the power/performance/density targets that they want. In addition, to achieve the clock speeds that they managed with Raptor Lake, they had to relax their density requirements multiple times on what was originally 10esf+. It's not area efficient, it's not power efficient on the end of the V/f curve that they play in to get their performance numbers, and as a result, it's not cheap.

But the tools have long been paid off.

The main purpose was mainly because Embedded wanted something cheap that only had P cores and didn't want to resort to crazy things like buying AMD. But since it'd be faster in games than Arrow Lake, it may still get a regular desktop release.

 

[dullard]

But since it'd be faster in SOME games than Arrow Lake, it may still get a regular desktop release.

Fixed that for you. Many games don't benefit from the additional cores. Other games will be hurt by the lower top all-core frequencies that Barlett Lake would have. There is a fundamental power constraint. As you add more and more cores without changing the node or the power limit, each core gets less and less power.

 

[jpiniero]

Fixed that for you. Many games don't benefit from the additional cores.

Arrow Lake... not Raptor Lake.

 

[LightningZ71]

Fixed that for you. Many games don't benefit from the additional cores. Other games will be hurt by the lower top all-core frequencies that Barlett Lake would have. There is a fundamental power constraint. As you add more and more cores without changing the node or the power limit, each core gets less and less power.

Which is why I suggested doing it on Intel 3 instead. The node improvement would show up most in all core boost scenarios. Intel could finally have a 12 core AVX512 workstation class processor. Heck, that target market probably wouldn't mind if Intel omitted the iGPU and made them all -F chips to save for space, or to trade off increasing the L3 size by at least 50%.

 

[jpiniero]

Which is why I suggested doing it on Intel 3 instead. The node improvement would show up most in all core boost scenarios.

And Bartlett Lake was supposed to be done on the cheap.