IBM Power11 appears 50% denser on the same Samsung 7nm process?

[Xebec]

I saw this post on X, and was curious how Power11 is achieving such a significant increase in density over Power 10, with both (in theory) being on the same Samsung 7nm node? Clock speeds are actually up a bit, so I'm thinking this isn't a HD vs HP library situation..

Power11: 45,9 MTr/mm²
Power10: 29,9 MTr/mm²
Power9: 11,5 MTr/mm²

Source:

https://twitter.com/x/status/1943005449659322848

 

[Fred Chen]

It looks like they weren't maxing out the transistor density to begin with.

 

[semiman]

It's still less dense than average(not theoretical maximum) Intel 7(former 10ESF) based CPUs, so It's kinda doable.

 

[hist78]

It looks like they weren't maxing out the transistor density to begin with.

According to Scotten Jones' analysis back in 2018 about Samsung 7nm node:

"The calculated transistor density is 101.23 MTx/mm[SUP]2[/SUP] based on what we know today although there is some uncertainty in M2P and track height in that calculation that will hopefully be cleared up in the full paper. The calculated density improvement versus 10nm is 1.86.

101.23 MTx/mm[SUP]2[/SUP] is slightly less than our calculated value for Intel’s 10nm process and slightly more than our calculated density for GLOBALFOUNDRIES and TSMC’s 7nm processes."

Samsung 10nm 8nm and 7nm at VLSIT - Semiwiki

I got a tip sheet today for the upcoming 2018 Symposia on VLSI Technology & Circuits to be held June 19th through 21st in Honolulu, Hawaii. There is some interesting information on Samsung's 10nm, 8nm and 7nm processes in the tip sheet: 10nm At 10nm the Samsung process has a Contacted Poly Pitch...

semiwiki.com

 

[Xebec]

According to Scotten Jones' analysis back in 2018 about Samsung 7nm node:

"The calculated transistor density is 101.23 MTx/mm[SUP]2[/SUP] based on what we know today although there is some uncertainty in M2P and track height in that calculation that will hopefully be cleared up in the full paper. The calculated density improvement versus 10nm is 1.86.

101.23 MTx/mm[SUP]2[/SUP] is slightly less than our calculated value for Intel’s 10nm process and slightly more than our calculated density for GLOBALFOUNDRIES and TSMC’s 7nm processes."

Samsung 10nm 8nm and 7nm at VLSIT - Semiwiki

I got a tip sheet today for the upcoming 2018 Symposia on VLSI Technology & Circuits to be held June 19th through 21st in Honolulu, Hawaii. There is some interesting information on Samsung's 10nm, 8nm and 7nm processes in the tip sheet: 10nm At 10nm the Samsung process has a Contacted Poly Pitch...

semiwiki.com

Thanks!

I think while the Samsung comparison is accurate (and a great reference point), this might have been looking at the density of the 'old' Intel 10nm process before it was reworked to actually yield. I don't recall how much the density slipped though for what eventually became Intel 7. Ice Lake - the first actually yielding Intel 10nm product launched Sept 2019.

 

[hist78]

Thanks!

I think while the Samsung comparison is accurate (and a great reference point), this might have been looking at the density of the 'old' Intel 10nm process before it was reworked to actually yield. I don't recall how much the density slipped though for what eventually became Intel 7. Ice Lake - the first actually yielding Intel 10nm product launched Sept 2019.

Although we are quickly approaching 2026, there may still be valid technical, licensing, contractual, and business reasons for IBM to continue using Samsung’s 7nm node for its Power11 processors. However in the near future, unless Samsung makes drastic improvements in its 3nm and 2nm nodes, IBM will need to find a new foundry partner to keep its Power Server product line competitive.

Who will that new partner be? Intel, TSMC, or Rapidus?

 

[Xebec]

Although we are quickly approaching 2026, there may still be valid technical, licensing, contractual, and business reasons for IBM to continue using Samsung’s 7nm node for its Power11 processors. However in the near future, unless Samsung makes drastic improvements in its 3nm and 2nm nodes, IBM will need to find a new foundry partner to keep its Power Server product line competitive.

Who will that new partner be? Intel, TSMC, or Rapidus?

TSMC is certainly the lazy option...

Though I suspect Rapidus given IBM is helping them make 2nm.

 

[osnium]

Although we are quickly approaching 2026, there may still be valid technical, licensing, contractual, and business reasons for IBM to continue using Samsung’s 7nm node for its Power11 processors. However in the near future, unless Samsung makes drastic improvements in its 3nm and 2nm nodes, IBM will need to find a new foundry partner to keep its Power Server product line competitive.

Who will that new partner be? Intel, TSMC, or Rapidus?

I would assume Rapidus if they ramp well by 2027.

 

[Maximus]

TSMC is certainly the lazy option...

Though I suspect Rapidus given IBM is helping them make 2nm.

IBM’s adoption will serve as a key indicator of the node health of Rapidus’ 2nm process. Similarly, since Samsung has yet to use its 3nm technology for Exynos, it raises questions about how to assure external customers of its maturity and reliability.