"Gelsinger also said that he expects the majority of Intel’s 2023 products to be fabbed internally." - how ? Three 10nm fabs are too low, and their yield is uncertain, may not reach level of 14nm. Maybe they plan to extend 14nm capacity up to 2023 at full volume. Seems that Rocket Lake end up as lower variants in Alder / Meteor Lake lines.
Gelsinger also said the majority of products in 2023 will be on 7nm, and that is on track. I expect, that Intel will aggressively retool fabs for 7nm, and take advantage of higher densities to produce more chips per wafer, thereby hitting their targets. Gelsinger is having an aura of confidence reminiscent of old Intel, which makes me optimistic for competition in 2023+.
you can redesign 7nm and reduce complexity to a point that density is hardly better than the previous node :) that is what happened with 10nm and that seems to be the case for 7nm.
bye bye leap forward. look at density increase on tsmc 7-5-3nm
the only problem with what you are saying is that you are assuming 7nm will work right and be on time and jump over 10nm problems. almost 10nm was an elective class you dropped in college and has no effect on the remaining courses, if anything you can now focus on 7nm.
We have never seen a big semi maker fall behind nearly 1.5 generations now and then recover and get back to the leading edge in the next node. maybe intel can do it maybe it cant. time will tell.
As 7nm (P1276) is now apparently a 2023 product they should at the very least rebrand it as 5nm or 5N, 5i or something to at least show that it is in the same class as TSMC´s 5N (which by then will likely be shipping 3N).
But knowing Intel they are more likely to name it something stupid as 7nm Improved UltraFin++
There was NEVER a point when 7nm WASN'T a 2023 product - other than Xe HPC.
Rocket Lake Q1 2021 Alder Lake Q42021 to Q12022 (1 year after Rocket Lake Meteor Lake Q12023 (1 year after Alder Lake)
Ice Lake SP - Q22021 Sapphire Rapids - Q42021 to Q12022 (3 quarters after Ice Lake SP) Granite Rapids - Q12023 (1 year after Sapphire Rapids)
Ice Lake SP will be a short term product - but quite a few businesses will snap them up in volume in the uncontested 2 Socket servers - with 128 lanes of PCIe4 (1st mainstream server implementation of PCIe4) and 8 channels of DDR4 3200 ECC... will be a VMWare match made in Silicon Valley
I for one, for my use, do not see a compelling case for Sapphire Rapids over Ice Lake SP - don't need PCIe5 and DDR5 - with 32 lanes of PCIe4 I can attach the 2 dual port 100Gb/s Ethernet NICs I will need... amd on the single CPU Workstations - only need 48 lanes for 2 dual port 100Gb/s Ethernet NICs + the carry over RTX6000.
So Intel 7nm had issues with Xe HPC - since TSMC has not even transitioned from Copper to Cobalt and has not even attempted to fab ALL layers of a piece with EUV - the issue COULD be in the ASML machines themselves - and incompetent Murthy and to a lesser degree Bob Swan.
And so you know, 2H2021 - Intel will be moving some i3 class CPUs to TSMC 5nm and then some to 3nm in 2022... And since it can supply the materials that TSMC can't get in volume - Intel will get full speed production while AMD is trying to supply enough Consoles SOCs so it can actually ship some of it's PC CPUs and GPUs...
You show your ignorance - SuperFin is named as such because of the structure allowed by moving to Cobalt instead of copper - and not just COAG - but the entire M0/M1 layers... TSMC isn't making that small of a feature with copper - which becomes a resistor rather than a conductor at that scale.
TSMC calls it's 10nm class process "7nm" which is far more ridiculous...
"There was NEVER a point when 7nm WASN'T a 2023 product" Then why did they announce delays? Come on, at some point it would have had to have been scheduled for earlier release. 10nm was supposed to be out in 2015, there must have been a time when 7nm was planned a lot sooner than 2023.
"Ice Lake SP ... with 128 lanes of PCIe4 (1st mainstream server implementation of PCIe4)" Rome has had that since August 2019.
"TSMC calls it's 10nm class process "7nm" which is far more ridiculous..." It's more dense and power-efficient than both Intel's in-production 10nm (not the fantasy football version they advertised) and their own 10nm. What else should they call it, and why would they do that?
The announced delay was for Xe HPC which was the lead off product for 7nm.
Enough of the Supposed to be this and should have been then
EPYC is niche and does not represent mainstream - my statement stands. Ice Lake SP will sell more than the entire history of Epyc combined. So many Intel VM Servers will be replaced by new Intel VM Servers.
TSMC "7nm" and Samsung "8nm" and Intel 10nm are all 10nm class processes - both TSMC and Samsung make every tiny little improvement into a new Product - because they are contract foundries and need to do so.
Not more dense - esp with 10nm SF. more power efficient in the Laptop SOC space than AMD - longer battery life - etc - not that there are very many AMD laptops to actually test...
Must be tough since all your old "go tos" have been OBE (overtaken by events).
5 record years in a row - they must be doing something right.
I can't get too worked up about marketing names for processes even as someone who generally likes Intel. The last time process names from anyone had a decent correlation to feature sizes process names were in microns, not nanometers.
Does Anandtech think the good news is shiping new tech this quarter or Intel says that????
There are a lot of OEM sales who advice to just skip Ice Lake because they expect a new CPU very fast as this is just an inbetween stopgap. Cascade Lake-R will continue to ship next to Ice Lake, that says enough.
Rocket Lake is to be seen if competitive with up to 8 cores... for some gamers sure the 2-3-4-5fps will do the trick to keep buying an Intel desktop on a dead platform.
In both cases ASP is down due to competition and will go down even further, because Intel had to drop prices massively and give cores for free (mainly in Cascade Lake -R to stay competitive with EPYC). Ice lake is not going to change that as they are only capable of half the core count vs AMD. 10nm HAs no longer the density and nobody is sure about the ghz scaling which was a 14nm favorits....
We as a consumer can only gain from this competition.
Of course Cooper Lake / Comet Lake will continue to sell alongside Ice Lake SP - same way Ice Lake SP will sell alongside Sapphire Rapids...
Dell, HPE and Supermicro are all in on Ice Lake SP - the market for Ice Lake SP is different than for Sapphire Rapids. VM farms will not have a need this gen for PCIe5 and the added cost of DDR5 - which during a transition is always a mixed bag
Thing is - no one is buying the 64 core Epycs (top selling SKUs are 32c) - and Ice Lake SP is a shorter term platform - Golden Cove (Alder, Sapphire and un named laptop SOC) is the long term unified platform - same Golden Cove across Laptop, Desktop and Servers (1-8 sockets). Sapphire Rapids will be 4 tile - initially 14 cores per tile (out of a possible 16) each with an HBM2x stack per tile. So 56 at launch and 64 as a refresh.
10nm vanilla had/has frequency issues - big IPC uplift but offset by low clocks
10nm SF (Cobalt) has no frequency issues - the TGL H are 5Ghz - my TGL 1165G7 can hit 4.8Ghz - and one can assume that 10nm ESF (Golden Cove) will not have issues with frequency either
Intel does not need to stay competitive with AMD - it's the other way around - thing about Ice Lake SP and Rocket Lake - Intel has the advantages of being an IDM will be able to supply the materials needed for super high volume - TSMC's long fragile supply chain has already broken as it has to rely on suppliers for many more products than Intel does - and in the case of substrates - Intel has the advantage of being the substrate manufacturer's largest customer for 30 years - while TSMC has been high volume for a couple of years now (and no the 16/14/12 were NEVER in the same league as their "7nm" as far as volume) - Intel gets all they want - what's left goes to TSMC - and that goes to Apple 1st, AMD with what's left - and AMD must produce the high volume, low margin console SOCs over it's much higher margin PC CPUs and GPUs.
Intel Xeon sapphire lake needs to be out sooner & Intel must not finish 14nm+++, why release i3 & Pentium on non-14nm of Intel (suppose 8 cores i3 after rocket lake?)? Intel KNOWs it is still very good compared to SAMSUNG & TSMC.. do i3 on 14nm+++, Xeon & i5 & rest on 10nm++, Altera on 14nm & 10nm, etc..
Sapphire Rapids will be out this year - Ice Lake SP will come first, and be made concurrently with Sapphire Rapids - both 10nm - but on separate processes (ICL is 10nm and SR is 10nm ESF).
Rocket Lake is the last 14nm Desktop CPU and has been on the roadmap for a couple of years Cooper Lake is the last 14nm Server CPUs Comet Lake was the last 14nm Laptop SOC - both Ice Lake and Tiger Lake since both 10nm
Core i3 is going to TSMC's "5nm" 2nd half of this year and follow up to Alder Lake may be "3nm" next year. Intel 7nm is between TSMC's 5nm and 3nm.
Altera is on 10nm mostly - the newer large stuff is and has been 10nm for some time.
Habana was originally on TSMC before Intel bought it - new product was already in risk production - so Habana will stay at TSMC until next gen.
Xe HP is 10nm SF or ESF - compute card
Xe HPG is TSMC 6nm for it's initial release - and 2nd gen likely on TSMC 3nm.
Once there is 14nm excess cap - the chipsets and network cards will be on 14nm (22&28 now)
7nm is intel's last chance. if it doesnt "work right" AND "on time" AND "in great abundance". intel will have to lay off most of workforce and reinvent itself as a fabless player 10% its current size.
LMAO - all 110K employees will be laid off? What are you like 12yo? - and who will provide all that fab space for 50% of the market?
Get a grip. $77B+ a year company - record 5 years in a row - all the while TSMC and AMD destroying it.. LMAO
TSMC is in the middle of a crisis with not being able to secure all the items in it's long supply chain which leads to shortages - specifically of AMD CPUs and GPUs
you dont seem to be familiar with semiconductor. "who will provide ... 50% something" first of all you are delusional if you think intel makes 50% of world transistor output. 2nd the answer to your question is the next gen node.
its not 50% of world's auto output or 50% of rare earth minerals. and yes if you will have to lay off 90% of their 110k workforce if manufacturing is gone and uarch is all they have left. its called free market. it doesnt allow you to make an inferior product and pay 10 folds as many ppl.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
18 Comments
Back to Article
TristanSDX - Thursday, January 21, 2021 - link
"Gelsinger also said that he expects the majority of Intel’s 2023 products to be fabbed internally." - how ? Three 10nm fabs are too low, and their yield is uncertain, may not reach level of 14nm. Maybe they plan to extend 14nm capacity up to 2023 at full volume. Seems that Rocket Lake end up as lower variants in Alder / Meteor Lake lines.Otritus - Thursday, January 21, 2021 - link
Gelsinger also said the majority of products in 2023 will be on 7nm, and that is on track. I expect, that Intel will aggressively retool fabs for 7nm, and take advantage of higher densities to produce more chips per wafer, thereby hitting their targets. Gelsinger is having an aura of confidence reminiscent of old Intel, which makes me optimistic for competition in 2023+.duploxxx - Friday, January 22, 2021 - link
you can redesign 7nm and reduce complexity to a point that density is hardly better than the previous node :) that is what happened with 10nm and that seems to be the case for 7nm.bye bye leap forward.
look at density increase on tsmc 7-5-3nm
azfacea - Friday, January 22, 2021 - link
the only problem with what you are saying is that you are assuming 7nm will work right and be on time and jump over 10nm problems. almost 10nm was an elective class you dropped in college and has no effect on the remaining courses, if anything you can now focus on 7nm.We have never seen a big semi maker fall behind nearly 1.5 generations now and then recover and get back to the leading edge in the next node. maybe intel can do it maybe it cant. time will tell.
azfacea - Friday, January 22, 2021 - link
almost as if*Arsenica - Thursday, January 21, 2021 - link
As 7nm (P1276) is now apparently a 2023 product they should at the very least rebrand it as 5nm or 5N, 5i or something to at least show that it is in the same class as TSMC´s 5N (which by then will likely be shipping 3N).But knowing Intel they are more likely to name it something stupid as 7nm Improved UltraFin++
Deicidium369 - Thursday, January 21, 2021 - link
There was NEVER a point when 7nm WASN'T a 2023 product - other than Xe HPC.Rocket Lake Q1 2021
Alder Lake Q42021 to Q12022 (1 year after Rocket Lake
Meteor Lake Q12023 (1 year after Alder Lake)
Ice Lake SP - Q22021
Sapphire Rapids - Q42021 to Q12022 (3 quarters after Ice Lake SP)
Granite Rapids - Q12023 (1 year after Sapphire Rapids)
Ice Lake SP will be a short term product - but quite a few businesses will snap them up in volume in the uncontested 2 Socket servers - with 128 lanes of PCIe4 (1st mainstream server implementation of PCIe4) and 8 channels of DDR4 3200 ECC... will be a VMWare match made in Silicon Valley
I for one, for my use, do not see a compelling case for Sapphire Rapids over Ice Lake SP - don't need PCIe5 and DDR5 - with 32 lanes of PCIe4 I can attach the 2 dual port 100Gb/s Ethernet NICs I will need... amd on the single CPU Workstations - only need 48 lanes for 2 dual port 100Gb/s Ethernet NICs + the carry over RTX6000.
So Intel 7nm had issues with Xe HPC - since TSMC has not even transitioned from Copper to Cobalt and has not even attempted to fab ALL layers of a piece with EUV - the issue COULD be in the ASML machines themselves - and incompetent Murthy and to a lesser degree Bob Swan.
And so you know, 2H2021 - Intel will be moving some i3 class CPUs to TSMC 5nm and then some to 3nm in 2022... And since it can supply the materials that TSMC can't get in volume - Intel will get full speed production while AMD is trying to supply enough Consoles SOCs so it can actually ship some of it's PC CPUs and GPUs...
You show your ignorance - SuperFin is named as such because of the structure allowed by moving to Cobalt instead of copper - and not just COAG - but the entire M0/M1 layers... TSMC isn't making that small of a feature with copper - which becomes a resistor rather than a conductor at that scale.
TSMC calls it's 10nm class process "7nm" which is far more ridiculous...
Spunjji - Friday, January 22, 2021 - link
"There was NEVER a point when 7nm WASN'T a 2023 product"Then why did they announce delays? Come on, at some point it would have had to have been scheduled for earlier release. 10nm was supposed to be out in 2015, there must have been a time when 7nm was planned a lot sooner than 2023.
"Ice Lake SP ... with 128 lanes of PCIe4 (1st mainstream server implementation of PCIe4)"
Rome has had that since August 2019.
"TSMC calls it's 10nm class process "7nm" which is far more ridiculous..."
It's more dense and power-efficient than both Intel's in-production 10nm (not the fantasy football version they advertised) and their own 10nm. What else should they call it, and why would they do that?
Deicidium369 - Saturday, January 23, 2021 - link
The announced delay was for Xe HPC which was the lead off product for 7nm.Enough of the Supposed to be this and should have been then
EPYC is niche and does not represent mainstream - my statement stands. Ice Lake SP will sell more than the entire history of Epyc combined. So many Intel VM Servers will be replaced by new Intel VM Servers.
TSMC "7nm" and Samsung "8nm" and Intel 10nm are all 10nm class processes - both TSMC and Samsung make every tiny little improvement into a new Product - because they are contract foundries and need to do so.
Not more dense - esp with 10nm SF. more power efficient in the Laptop SOC space than AMD - longer battery life - etc - not that there are very many AMD laptops to actually test...
Must be tough since all your old "go tos" have been OBE (overtaken by events).
5 record years in a row - they must be doing something right.
drothgery - Friday, January 22, 2021 - link
I can't get too worked up about marketing names for processes even as someone who generally likes Intel. The last time process names from anyone had a decent correlation to feature sizes process names were in microns, not nanometers.duploxxx - Friday, January 22, 2021 - link
Does Anandtech think the good news is shiping new tech this quarter or Intel says that????There are a lot of OEM sales who advice to just skip Ice Lake because they expect a new CPU very fast as this is just an inbetween stopgap. Cascade Lake-R will continue to ship next to Ice Lake, that says enough.
Rocket Lake is to be seen if competitive with up to 8 cores... for some gamers sure the 2-3-4-5fps will do the trick to keep buying an Intel desktop on a dead platform.
In both cases ASP is down due to competition and will go down even further, because Intel had to drop prices massively and give cores for free (mainly in Cascade Lake -R to stay competitive with EPYC). Ice lake is not going to change that as they are only capable of half the core count vs AMD. 10nm HAs no longer the density and nobody is sure about the ghz scaling which was a 14nm favorits....
We as a consumer can only gain from this competition.
Deicidium369 - Saturday, January 23, 2021 - link
Of course Cooper Lake / Comet Lake will continue to sell alongside Ice Lake SP - same way Ice Lake SP will sell alongside Sapphire Rapids...Dell, HPE and Supermicro are all in on Ice Lake SP - the market for Ice Lake SP is different than for Sapphire Rapids. VM farms will not have a need this gen for PCIe5 and the added cost of DDR5 - which during a transition is always a mixed bag
Thing is - no one is buying the 64 core Epycs (top selling SKUs are 32c) - and Ice Lake SP is a shorter term platform - Golden Cove (Alder, Sapphire and un named laptop SOC) is the long term unified platform - same Golden Cove across Laptop, Desktop and Servers (1-8 sockets). Sapphire Rapids will be 4 tile - initially 14 cores per tile (out of a possible 16) each with an HBM2x stack per tile. So 56 at launch and 64 as a refresh.
10nm vanilla had/has frequency issues - big IPC uplift but offset by low clocks
10nm SF (Cobalt) has no frequency issues - the TGL H are 5Ghz - my TGL 1165G7 can hit 4.8Ghz - and one can assume that 10nm ESF (Golden Cove) will not have issues with frequency either
Intel does not need to stay competitive with AMD - it's the other way around - thing about Ice Lake SP and Rocket Lake - Intel has the advantages of being an IDM will be able to supply the materials needed for super high volume - TSMC's long fragile supply chain has already broken as it has to rely on suppliers for many more products than Intel does - and in the case of substrates - Intel has the advantage of being the substrate manufacturer's largest customer for 30 years - while TSMC has been high volume for a couple of years now (and no the 16/14/12 were NEVER in the same league as their "7nm" as far as volume) - Intel gets all they want - what's left goes to TSMC - and that goes to Apple 1st, AMD with what's left - and AMD must produce the high volume, low margin console SOCs over it's much higher margin PC CPUs and GPUs.
nobodyblog - Friday, January 22, 2021 - link
Intel Xeon sapphire lake needs to be out sooner & Intel must not finish 14nm+++, why release i3 & Pentium on non-14nm of Intel (suppose 8 cores i3 after rocket lake?)? Intel KNOWs it is still very good compared to SAMSUNG & TSMC.. do i3 on 14nm+++, Xeon & i5 & rest on 10nm++, Altera on 14nm & 10nm, etc..Deicidium369 - Saturday, January 23, 2021 - link
Sapphire Rapids will be out this year - Ice Lake SP will come first, and be made concurrently with Sapphire Rapids - both 10nm - but on separate processes (ICL is 10nm and SR is 10nm ESF).Rocket Lake is the last 14nm Desktop CPU and has been on the roadmap for a couple of years
Cooper Lake is the last 14nm Server CPUs
Comet Lake was the last 14nm Laptop SOC - both Ice Lake and Tiger Lake since both 10nm
Core i3 is going to TSMC's "5nm" 2nd half of this year and follow up to Alder Lake may be "3nm" next year. Intel 7nm is between TSMC's 5nm and 3nm.
Altera is on 10nm mostly - the newer large stuff is and has been 10nm for some time.
Habana was originally on TSMC before Intel bought it - new product was already in risk production - so Habana will stay at TSMC until next gen.
Xe HP is 10nm SF or ESF - compute card
Xe HPG is TSMC 6nm for it's initial release - and 2nd gen likely on TSMC 3nm.
Once there is 14nm excess cap - the chipsets and network cards will be on 14nm (22&28 now)
azfacea - Friday, January 22, 2021 - link
7nm is intel's last chance. if it doesnt "work right" AND "on time" AND "in great abundance". intel will have to lay off most of workforce and reinvent itself as a fabless player 10% its current size.Deicidium369 - Saturday, January 23, 2021 - link
LMAO - all 110K employees will be laid off? What are you like 12yo? - and who will provide all that fab space for 50% of the market?Get a grip. $77B+ a year company - record 5 years in a row - all the while TSMC and AMD destroying it.. LMAO
TSMC is in the middle of a crisis with not being able to secure all the items in it's long supply chain which leads to shortages - specifically of AMD CPUs and GPUs
azfacea - Monday, January 25, 2021 - link
you dont seem to be familiar with semiconductor. "who will provide ... 50% something"first of all you are delusional if you think intel makes 50% of world transistor output. 2nd the answer to your question is the next gen node.
its not 50% of world's auto output or 50% of rare earth minerals. and yes if you will have to lay off 90% of their 110k workforce if manufacturing is gone and uarch is all they have left. its called free market. it doesnt allow you to make an inferior product and pay 10 folds as many ppl.
azfacea - Monday, January 25, 2021 - link
and you showed to talk about " - record 5 year in row - " in an article you dont seem to have read.