The last public details of Intel's 10nm (P1274) showed it was tied with TSMC's 7nm (7FF) in density metrics (54 nm as contacted gate pitch for both processes). As the slide still compares it to TSMC's 7FF is to be expected that their density is still comparable, but we'll only know for sure once somebody cleaves a shipping processor and measures it.
Newsflash: today Intel announced long-awaited release of a new code name. "Specious Lake represents the last 10 years of Intel innovation" said Super Vice President of Marketing Patel Dumborajachurpan.
Not sure where you heard that Intel 10nm is close to TSMC 7nm? Pretty sure that's wrong. Plus TSMC has stated that they are already offering 5nm in 2020, iPhone 2020 already confirmed that it's going to be using it.
you mean you think they have changed it. no one has any information confirming it. besides who cares. its just for mobile. the real piece of info here is 7nm. no more quad patterning. maybe no more dual patterning. better yields. faster chips. cant wait. my old 8700k will last me fine till then. was gonna get some really fast ddr4 but might just wait for ddr5 and 7nm
Intel has long (*very* long) reported a transistor density of ~102 million transistors per mm^2 at 10nm. Intel's 7nm node was originally planned for a staggering 250+ MTr/mm^2 transistor density, but more recently -basically a bit after Brian Krzanich stepped down- I heard about densities in the order of ~200+ MTr/mm^2. Apparently Robert Swan set up a saner and more conservative target because he wants to avoid a repeat of the 10nm node fiasco, just like he should. Or he is listening to what his head engineers are saying is possible for 7nm, at least without blowing all their (publicly announced) deadlines and overshooting their allocated budget.
p.s. According to SemiWiki, to give you a comparison, TSMC at 7nm (non EUV) has a "minimum" transistor density of ~97 MTr/mm^2 but their 2nd gen 7nm node with EUV and their 6nm node will both have a 18% higher density of ~114 MTr/mm^2. Samsung will start at ~95 MTr/mm^2 and end up at 113 MTr/mm^2 in the third iteration of either their 7nm node or 5/6nm node (they say the node number is unclear).
So both Samsung and TSMC at 7nm are comparable to Intel at 10nm in terms of transistor density, but we do not yet really know how the high power variant of either of these nodes performs, or rather the sole such sample we have is Radeon VII. Intel released a sole 15W Core i3 with atrocious clocks (and a disabled iGPU) but that was the first unfixed iteration of their 10nm node. Ice Lake will sport the presumably fixed 10nm+ node. No data yet from Samsung's 7nm node.
The kicker here is in the fine print. The Intel density numbers assume you are using high density libraries, when in fact their 10nm SDK has 3 different libraries of transistors. The high density cells are >100MT/mm^2, but the highest performance library is more like 60MT/mm^2 or less. I don't think TSMC or Samsung have released density figures with that granularity, but I assume all three are using similar marketing tactics to sound the best.
The issue is that unlike Intel, most TSMC chips actually do use the dense libraries. That includes servers, for example Centriq does 45 MT/mm^2 on 10nm. Mobile 10nm SoCs like Kirin 970 get 57 MT/mm^2, while Snapdragon 845 does 56, so they are only ~25% denser.
So 10nm Intel servers should achieve 75 MT/mm^2. However I bet they won't even get 40 MT/mm^2 given 14nm density is around 16 MT/mm^2.
Intel CPU should achieve 50MT/mm² (10 nm) or 65MT at best with 7nm while ARM is achieved 96MT already (Kirin) which goes up to 120MT with usage of UHD libs on same 7nm TSMC node, 170MT on 5 nm and around 200 MT on N5+ (UHD lib probably). Theoretical all the node's are physically the same (really nm) so achievable density is also the same but we seriously need a new gate structure. FinFET never whose good for many things to start with. Whosent good for analogue at all nor mixed circuit's for that matter which means not good for transceivers, MOSFET's cetera. It whose an Intel's child all together on the quest & design methodology to ensure peak higher possible clocks, two fins instead one. Does that sound familiar & then let's tie even more fets together (high performance lib) to ensure even better drain so that pore thing could hit even more MHz. The strategy whose wrong from beginning all together. FinFET brought a modest bump (200 MHz) over the planar in terms of what is industrialy considerd as sustainable leak for a complex structure as transistor but with an almost 2x power cost in a idle state it also enabled higher density of around 20~25% compared to planar in terms of possible miniaturisation but again with much higher both design and manufacturing costs. The AMD's (& IBM's also) idea originally with SOI whose much better one overlay but they didn't had enough money to push it to the end so didn't the Global foundries. It remains to be seen how good will Gate All Around be but FinFET really needs to die for many reasons.
The base DUV 7 nm TSMC is a very good base for calculation as it uses basic rooting lib. The HP lib basically ties more fin's per gate to ensure better connectivity & better drain so it's fair to split HP in half to get the base gate calculation. Intel doesn't give (nor use) base, HD or UHD libs only HP one to ensure brute forced higher clocks & it sips power like hell leaking all around.
Coincidentally, 2021/22 is about when I'm due for a new desktop build; so I really am hoping Intel gets its stuff together and Intel and AMD are able to vigorously compete for my CPU dollars.
10nm desktop sounds like 1st half 2020 not 2021, this means less than year away since shortly it be starting 2nd half. and what does 1st half mean - any time between Jan and July 1.
I did not say that - I say it may not have to - that is different. 5Ghz on current Intel architexture - may be slower than 3Ghz on Sunny Cove. If ALU's on CPU handle twice the instructions per clock cycle as previous architexture than 5 Ghz is not required.
Is is trustworthy (supposed to be an IL 1.4GHz base, 2.3GHz turbo)? Who knows, but it's better done than most fakes. Reveals an IPC very close to current i5 except for AES.
If this is real it must be either from a very early engineering sample or Geekbench has trouble running on the new arch. By the way, the size of the caches is consistent with what Intel has disclosed about Sunny Cove : They said "50% increase in the size of L1 data cache", well the L1D cache of the mystery processor is 48 KB. They said "Larger L2 cache", well that went up to 512 KB, just like Ryzen. I wonder if the L2 cache of the X and SP processors will also increase in size. Perhaps they can afford 1.5 to 2 MB L2 cache due to the 10nm node.
"5Ghz on current Intel architexture - may be slower than 3Ghz on Sunny Cove." No, that's impossible. The architectures are not that different. Sunny Cove is a marginal, evolutionary redesign of the Skylake arch. Intel's true redesign is going to be Ocean Cove, Golden Cove's successor (after Sunny, Willow and Golden Cove, so we are talking about an early 2023 launch assuming zero delays with either of these three Coves).
Bearing also in mind that Intel's 10nm+ node will still not perform as good as Intel's 14nm++ node, a Sunny Cove core clocked at 4.5 GHz will, at best, perform as good as a select Intel 5 GHz clocked core, though at a quite lower TDP.
"5Ghz on current Intel architexture - may be slower than 3Ghz on Sunny Cove"
At this point it is not possible anymore on real software of today (unless you are going to compare some future AVX512 implementation to AVX2 on some artificial tests). If anything, mitigations for various security flaws require LESS OoO, not more.
This said nothing about desktop, only server. That probably means 2H 2020 for 10nm desktop as server has much higher margins to recoup their capital expense.
I don't expect to see 10nm high performance desktops for several reasons. 1. 10nm notebook and 10nm server signals to me that 10nm doesn't hit high speeds. 2. Intel only has one 10nm fab 3. Rumored roadmaps have 14nm go out to 2021. Still no official word on 10nm desktop from Intel. I bet Intel will be limited to 10nm low core count server chips due to yields. Guess we'll know more when the 10nm notebook chips appear this year.
Is this a launch like q4 18 for cascade lake or a real launch like cascade lake q2 19 (big datacenters received some product shipments earlier for mass deployment. Same for skylake-sp)?
I am an optimist, this shows that Intel is learning from it's past with 10nm mess up with Cannon Lake, and other things. They know what processes and products are failing. It is obvious that the Micron deal was working for them and invested in 28 billions into changes for the future and got rid the old stuff - even stuff like quark which I never truly understand why it was there.
But intel has been hurt by bad perception issues that they are a dinosaur and no longer the leader but this bad reception comes from minor area of industry which is gamers on desktop which is actually the dying industry - mobile is where the future is going, ARM s major threat to Intel not AMD.
Don't think Intel is stupid, with all this 10nm/7nm stuff, they know they had troubles with 10nm - and it would be stupid to think that they have not learn from it - 10nm is logical to come out on notebooks first because laptop and tablets are 80% plus of the market. and then servers and desktops.
My thought 10nm is cannon lake with fixes and stability over last 2 years and 7nm is brand new process - 10nm process with cobalt was too aggressive.
No I just don't wear AMD dark shades that make you blind to seeing the future besides AMD propaganda. You have no idea how much an optimist, I am. I am not just talking about stupid cpu architexture.
I forgot exactly where it shown - but keep in mind if you look at Amazon / Newegg CPU's sales those numbers are primary almost 100% sales - because mobile sales are include mostly if not entirely with companies like Dell, HP, Lenovo ….
But there are companies that do it - it not 100% accurate but just visit you local Best Buy and compare mobile laptop vs desktops ( not counting All in ones which is actually mostly mobile in design )
of course you forgot.. cause maybe it doesnt exist ?? HStewart.. just face it.. intel isnt the leader it once was.. its be hind on just about everything now.. you keep bending and twisting things just to make intel look better then it is.. cause you are just a fanboy.. as you said to eva02langley " No I just don't wear AMD dark shades that make you blind to seeing the future besides AMD propaganda " the funny thing is.. you do the EXACT same thing when it comes to intel.. before you accuse some one of doing that.. you REALLY need to see that you do the same thing.. if you were really an optimist.. then you would be able to admit, that intel has screwed up the last few years.. but yet you CAN'T
I think saying that you are a "selective optimist" is imho probably closer to the truth. A bit like people seeing their favorite sports team winning the next trophy simply because it's always been their favorite team.
So, 10m was too aggressive and the way to fix that is claim that you'll do an ADDITIONAL 2.4x density improvement with 7nm?
What EXACTLY did they learn from 10nm? Because it doesn't appear to be "STFU until you KNOW the process works". And it doesn't appear to be "make small improvements to the process every year instead of trying to change everything all at once for a massive density boost".
The real race is about gate-all-around replacing FinFets and delivering much higher performance, not just power and area optimization. That would probably be TSMC and Samsungs 3nm...
Actually no, the lies about great yields, shipping soon, started with Broadwell and 14nm. It's just that in that case the delay was "only" about 6..9 months, so the true believers could ignore then forget it. And remember it's not 10nm that has been the subject of constant lies. What about all those claims regarding Optane that have now been so aggressively scaled back? And 5G?
One error is reasonable, everyone makes mistakes or is too optimistic. But this is a clear pattern that's been going on for at least five years.
" They only failed with Cannon Lake - all of the other stuff - is just BS on Internet. " only failed with cannon lake ?? what about 10nm ??????? what about the fact.. that Epyc is on par.. or faster then Xeon when the software can use the cores it is given ?? the only BS that is there.. is the BS that you keep saying HStewart.. when will you wake up and realize your beloved intel.. is losing... failing.. and is falling behind... oh wait.. cause you cant, and you wont admit it..
Jumping to 7 nm is feasible considering the usage of EUV. This is what Intel should have done for 10 nm but EUV simply wasn't ready and they thought multipatterning could still scale. Sometime you make the wrong bet.
The bigger news is that Xe will indeed be using EMIB. This opens the door for scaling between the low and high end by re-using the same piece of silicon, just numerous instances per package. This is something that should not be underestimated in terms of performance. With Intel having their own fabs, they can afford to throw far more silicon into the GPU race, more than even nVidia's insane 818 mm^2 monolithic die. The question is what size will the base Xe die be before they scale up in number? 300 mm^2? 400 mm^2?
Why do you imagine that Intel is the only one with this sort of tech? As soon as it becomes appropriate, I expect nV and AMD to follow the same path (AMD already has, of course, with Ryzen).
Packaging is currently undergoing an explosion of different possibilities. It's the shortest of short-sightedness to imagine that Intel's particular entries into this space (EMIB and Foveros) represent the only options, and that Samsung and TSMC have no clue about how to do this, or that their options inevitably suck compared to the Intel's...
Intel is the only one who also owns fabs. AMD and nVidia have access to other foundries but those are middle men who are out to make their share of money as well. The point is that Intel by merit of having their own fabs could throw more silicon into a product at a lower cost than either AMD or nVidia could. Intel could brute force their way to the performance crown which would be too costly for AMD or nVidia to match.
The whole chiplet strategy for GPU is indeed being explored by both AMD and nVidia. There is a nice research paper published by nVidia on this very topic. AMD has strongly hinted at a similar concept for their future GPUs and CPU + GPU parts.
There is still a minimum size desirable. Too many smaller dies makes the packing complex. Easier to link together five 200 mm^2 dies than twenty 50 mm^2 dies. It is also simpler to have one side of the die match the dimension of a HBM module to simplify the design.
I'm looking forward to getting a 28/32 core CPU for work. I had been hoping my company would buy it for me, but if they won't, I may end up buying it myself and working from home a bit more.
Translation - Consumer CPUs will be on 14nm until 2022 when high volume manufacturing begins and will be in short supply until 2023. That's a huge window for competitors with ARM chips and, of course, for AMD to exploit because they're using TSMC to fab.
What happened with Intel? Someone goofed up ordering EUV equipment from ASML and they company is going to end up five years behind the competition when its said and done.
Intel already can manufacturer 7nm chips but why ? Intel can not meet demand of it's 14nm+++ and now shipping 10nm for the first time with improved performance over previous gen. So, milk much as possible from 14nm+++ and 10nm than in two years, release 7nm. Good timing Intel.
Where's the "improved performance"? If you look at what Intel put out, they are talking about efficiency....isn't that what all the people putting AMD down used to talk about when it came to AMD products? If you can't improve performance, then work on efficiency, but Intel hasn't improved IPC, clock speeds have been stagnant(if you look at what could be done three years ago if you would delid an Intel chip), so where's the true improvement in performance from Intel?
AMD got its act together with Ryzen in 2017, second generation brought a 300MHz clock speed improvement plus 3-5% IPC improvement, and third generation will bring an estimated 13-15 percent IPC boost plus a significant clock speed improvement compared to second generation. The stuff from AMD isn't speculation about products that are still 1-2 years away from when consumers have access to them, it's only one to two months away.
I am most excited about Lakefield - but it was interesting to see the performance comparing 10nm cpus physically running again existing 8th and 9th computers.
I am desiring a portable always connected machine, and thought about it and ideally it would be compatible with software I run today on Windows. No need for iOS or Android and not emulated with ARM making applications slow. Lakefield is the answer.
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.
Facebook
Twitter
RSS
71 Comments
Back to Article
Marlin1975 - Wednesday, May 8, 2019 - link
But what type of "10nm" is this? Whats the density?Saying 10nm means nothing since they have changed designs to get a better yields.
Arsenica - Wednesday, May 8, 2019 - link
The last public details of Intel's 10nm (P1274) showed it was tied with TSMC's 7nm (7FF) in density metrics (54 nm as contacted gate pitch for both processes).As the slide still compares it to TSMC's 7FF is to be expected that their density is still comparable, but we'll only know for sure once somebody cleaves a shipping processor and measures it.
peevee - Thursday, May 9, 2019 - link
Newsflash: today Intel announced long-awaited release of a new code name. "Specious Lake represents the last 10 years of Intel innovation" said Super Vice President of Marketing Patel Dumborajachurpan.RSAUser - Tuesday, May 14, 2019 - link
Not sure where you heard that Intel 10nm is close to TSMC 7nm? Pretty sure that's wrong.Plus TSMC has stated that they are already offering 5nm in 2020, iPhone 2020 already confirmed that it's going to be using it.
bobhumplick - Wednesday, May 8, 2019 - link
you mean you think they have changed it. no one has any information confirming it. besides who cares. its just for mobile. the real piece of info here is 7nm. no more quad patterning. maybe no more dual patterning. better yields. faster chips. cant wait. my old 8700k will last me fine till then. was gonna get some really fast ddr4 but might just wait for ddr5 and 7nmSantoval - Wednesday, May 8, 2019 - link
Intel has long (*very* long) reported a transistor density of ~102 million transistors per mm^2 at 10nm. Intel's 7nm node was originally planned for a staggering 250+ MTr/mm^2 transistor density, but more recently -basically a bit after Brian Krzanich stepped down- I heard about densities in the order of ~200+ MTr/mm^2.Apparently Robert Swan set up a saner and more conservative target because he wants to avoid a repeat of the 10nm node fiasco, just like he should. Or he is listening to what his head engineers are saying is possible for 7nm, at least without blowing all their (publicly announced) deadlines and overshooting their allocated budget.
Santoval - Wednesday, May 8, 2019 - link
p.s. According to SemiWiki, to give you a comparison, TSMC at 7nm (non EUV) has a "minimum" transistor density of ~97 MTr/mm^2 but their 2nd gen 7nm node with EUV and their 6nm node will both have a 18% higher density of ~114 MTr/mm^2.Samsung will start at ~95 MTr/mm^2 and end up at 113 MTr/mm^2 in the third iteration of either their 7nm node or 5/6nm node (they say the node number is unclear).
So both Samsung and TSMC at 7nm are comparable to Intel at 10nm in terms of transistor density, but we do not yet really know how the high power variant of either of these nodes performs, or rather the sole such sample we have is Radeon VII. Intel released a sole 15W Core i3 with atrocious clocks (and a disabled iGPU) but that was the first unfixed iteration of their 10nm node. Ice Lake will sport the presumably fixed 10nm+ node. No data yet from Samsung's 7nm node.
FullmetalTitan - Thursday, May 9, 2019 - link
The kicker here is in the fine print. The Intel density numbers assume you are using high density libraries, when in fact their 10nm SDK has 3 different libraries of transistors. The high density cells are >100MT/mm^2, but the highest performance library is more like 60MT/mm^2 or less. I don't think TSMC or Samsung have released density figures with that granularity, but I assume all three are using similar marketing tactics to sound the best.Wilco1 - Thursday, May 9, 2019 - link
The issue is that unlike Intel, most TSMC chips actually do use the dense libraries. That includes servers, for example Centriq does 45 MT/mm^2 on 10nm. Mobile 10nm SoCs like Kirin 970 get 57 MT/mm^2, while Snapdragon 845 does 56, so they are only ~25% denser.So 10nm Intel servers should achieve 75 MT/mm^2. However I bet they won't even get 40 MT/mm^2 given 14nm density is around 16 MT/mm^2.
ZolaIII - Thursday, May 9, 2019 - link
Intel CPU should achieve 50MT/mm² (10 nm) or 65MT at best with 7nm while ARM is achieved 96MT already (Kirin) which goes up to 120MT with usage of UHD libs on same 7nm TSMC node, 170MT on 5 nm and around 200 MT on N5+ (UHD lib probably). Theoretical all the node's are physically the same (really nm) so achievable density is also the same but we seriously need a new gate structure. FinFET never whose good for many things to start with. Whosent good for analogue at all nor mixed circuit's for that matter which means not good for transceivers, MOSFET's cetera. It whose an Intel's child all together on the quest & design methodology to ensure peak higher possible clocks, two fins instead one. Does that sound familiar & then let's tie even more fets together (high performance lib) to ensure even better drain so that pore thing could hit even more MHz. The strategy whose wrong from beginning all together. FinFET brought a modest bump (200 MHz) over the planar in terms of what is industrialy considerd as sustainable leak for a complex structure as transistor but with an almost 2x power cost in a idle state it also enabled higher density of around 20~25% compared to planar in terms of possible miniaturisation but again with much higher both design and manufacturing costs. The AMD's (& IBM's also) idea originally with SOI whose much better one overlay but they didn't had enough money to push it to the end so didn't the Global foundries. It remains to be seen how good will Gate All Around be but FinFET really needs to die for many reasons.ZolaIII - Thursday, May 9, 2019 - link
The base DUV 7 nm TSMC is a very good base for calculation as it uses basic rooting lib. The HP lib basically ties more fin's per gate to ensure better connectivity & better drain so it's fair to split HP in half to get the base gate calculation. Intel doesn't give (nor use) base, HD or UHD libs only HP one to ensure brute forced higher clocks & it sips power like hell leaking all around.ajc9988 - Sunday, May 26, 2019 - link
TSMC has. Here are the two things you are looking for:https://fuse.wikichip.org/news/2004/iedm-2018-inte...
https://www.techcenturion.com/7nm-10nm-14nm-fabric...
Then, you should compare that to achieved densities like here:
https://www.anandtech.com/show/13687/qualcomm-snap...
0ldman79 - Friday, May 10, 2019 - link
Density is only part of the equation.Eventually there will be gaps in between the circuits and transistors as the gap will be too small for a dielectric to stop electromigration.
The density doesn't set the size of the transistors, it is just one side effect of the smaller lithography.
DanNeely - Wednesday, May 8, 2019 - link
I'll believe it when I can buy it.Coincidentally, 2021/22 is about when I'm due for a new desktop build; so I really am hoping Intel gets its stuff together and Intel and AMD are able to vigorously compete for my CPU dollars.
HStewart - Wednesday, May 8, 2019 - link
10nm desktop sounds like 1st half 2020 not 2021, this means less than year away since shortly it be starting 2nd half. and what does 1st half mean - any time between Jan and July 1.Wilco1 - Wednesday, May 8, 2019 - link
Do you seriously believe there will be 10nm 5GHz desktop chips next year?HStewart - Wednesday, May 8, 2019 - link
Desktop is now a minor market, but server will have new cpus's but with new architexture 5Ghz may not be necessary on Sunny Cove.Wilco1 - Wednesday, May 8, 2019 - link
Right, so you agree 10nm will never reach 5GHz?HStewart - Wednesday, May 8, 2019 - link
I did not say that - I say it may not have to - that is different. 5Ghz on current Intel architexture - may be slower than 3Ghz on Sunny Cove. If ALU's on CPU handle twice the instructions per clock cycle as previous architexture than 5 Ghz is not required.name99 - Wednesday, May 8, 2019 - link
Not if this is anything close to reality!https://browser.geekbench.com/v4/cpu/compare/11493...
https://browser.geekbench.com/v4/cpu/compare/13056...
Is is trustworthy (supposed to be an IL 1.4GHz base, 2.3GHz turbo)? Who knows, but it's better done than most fakes. Reveals an IPC very close to current i5 except for AES.
Santoval - Wednesday, May 8, 2019 - link
If this is real it must be either from a very early engineering sample or Geekbench has trouble running on the new arch. By the way, the size of the caches is consistent with what Intel has disclosed about Sunny Cove :They said "50% increase in the size of L1 data cache", well the L1D cache of the mystery processor is 48 KB.
They said "Larger L2 cache", well that went up to 512 KB, just like Ryzen.
I wonder if the L2 cache of the X and SP processors will also increase in size. Perhaps they can afford 1.5 to 2 MB L2 cache due to the 10nm node.
Santoval - Wednesday, May 8, 2019 - link
"5Ghz on current Intel architexture - may be slower than 3Ghz on Sunny Cove."No, that's impossible. The architectures are not that different. Sunny Cove is a marginal, evolutionary redesign of the Skylake arch. Intel's true redesign is going to be Ocean Cove, Golden Cove's successor (after Sunny, Willow and Golden Cove, so we are talking about an early 2023 launch assuming zero delays with either of these three Coves).
Bearing also in mind that Intel's 10nm+ node will still not perform as good as Intel's 14nm++ node, a Sunny Cove core clocked at 4.5 GHz will, at best, perform as good as a select Intel 5 GHz clocked core, though at a quite lower TDP.
hMunster - Thursday, May 9, 2019 - link
There is no way at all that Intel will nearly double single-core performance that way. We are lucky if we get more than 10% at the same clock speed.peevee - Thursday, May 9, 2019 - link
"5Ghz on current Intel architexture - may be slower than 3Ghz on Sunny Cove"At this point it is not possible anymore on real software of today (unless you are going to compare some future AVX512 implementation to AVX2 on some artificial tests). If anything, mitigations for various security flaws require LESS OoO, not more.
shabby - Wednesday, May 8, 2019 - link
It will reach 5ghz... on a single core.peevee - Thursday, May 9, 2019 - link
... for 1ns before melting.schujj07 - Wednesday, May 8, 2019 - link
This said nothing about desktop, only server. That probably means 2H 2020 for 10nm desktop as server has much higher margins to recoup their capital expense.TristanSDX - Wednesday, May 8, 2019 - link
look at slides from their financial results. Profit margins for server are on the same level as desktop, suprise.HStewart - Wednesday, May 8, 2019 - link
Desktop sales are lower than server and especially mobile.Korguz - Thursday, May 9, 2019 - link
do you have a link to back this claim up??ZippZ - Wednesday, May 8, 2019 - link
I don't expect to see 10nm high performance desktops for several reasons.1. 10nm notebook and 10nm server signals to me that 10nm doesn't hit high speeds.
2. Intel only has one 10nm fab
3. Rumored roadmaps have 14nm go out to 2021. Still no official word on 10nm desktop from Intel.
I bet Intel will be limited to 10nm low core count server chips due to yields. Guess we'll know more when the 10nm notebook chips appear this year.
peevee - Thursday, May 9, 2019 - link
You have good points."I bet Intel will be limited to 10nm low core count server chips due to yields."
They will use multi-chip packages, just like AMD does with EPYC. And NUMA is a btch on software not optimized for it.
The real question will it even reach 3GHz sustained or will be stuck around 2.
peevee - Thursday, May 9, 2019 - link
"10nm desktop sounds like 1st half 2020 not 2021""10nm desktop sounds like" will not be available ever as it is not going to yield well (better than 14+++) on 5GHz at all.
brakdoo - Wednesday, May 8, 2019 - link
Is this a launch like q4 18 for cascade lake or a real launch like cascade lake q2 19 (big datacenters received some product shipments earlier for mass deployment. Same for skylake-sp)?HStewart - Wednesday, May 8, 2019 - link
This is not cannon lake - but it revisedHStewart - Wednesday, May 8, 2019 - link
I am an optimist, this shows that Intel is learning from it's past with 10nm mess up with Cannon Lake, and other things. They know what processes and products are failing. It is obvious that the Micron deal was working for them and invested in 28 billions into changes for the future and got rid the old stuff - even stuff like quark which I never truly understand why it was there.But intel has been hurt by bad perception issues that they are a dinosaur and no longer the leader but this bad reception comes from minor area of industry which is gamers on desktop which is actually the dying industry - mobile is where the future is going, ARM s major threat to Intel not AMD.
Don't think Intel is stupid, with all this 10nm/7nm stuff, they know they had troubles with 10nm - and it would be stupid to think that they have not learn from it - 10nm is logical to come out on notebooks first because laptop and tablets are 80% plus of the market. and then servers and desktops.
My thought 10nm is cannon lake with fixes and stability over last 2 years and 7nm is brand new process - 10nm process with cobalt was too aggressive.
eva02langley - Wednesday, May 8, 2019 - link
"I am an optimist"Correction, you are wearing Intel goggles.
HStewart - Wednesday, May 8, 2019 - link
No I just don't wear AMD dark shades that make you blind to seeing the future besides AMD propaganda. You have no idea how much an optimist, I am. I am not just talking about stupid cpu architexture.brakdoo - Wednesday, May 8, 2019 - link
No, you are just butthurt and make things up like: " because laptop and tablets are 80% plus of the market" (vs servers and desktops???)Where do you have that number from?
HStewart - Wednesday, May 8, 2019 - link
I forgot exactly where it shown - but keep in mind if you look at Amazon / Newegg CPU's sales those numbers are primary almost 100% sales - because mobile sales are include mostly if not entirely with companies like Dell, HP, Lenovo ….But there are companies that do it - it not 100% accurate but just visit you local Best Buy and compare mobile laptop vs desktops ( not counting All in ones which is actually mostly mobile in design )
Korguz - Wednesday, May 8, 2019 - link
of course you forgot.. cause maybe it doesnt exist ?? HStewart.. just face it.. intel isnt the leader it once was.. its be hind on just about everything now.. you keep bending and twisting things just to make intel look better then it is.. cause you are just a fanboy.. as you said to eva02langley " No I just don't wear AMD dark shades that make you blind to seeing the future besides AMD propaganda " the funny thing is.. you do the EXACT same thing when it comes to intel.. before you accuse some one of doing that.. you REALLY need to see that you do the same thing.. if you were really an optimist.. then you would be able to admit, that intel has screwed up the last few years.. but yet you CAN'TIrata - Thursday, May 9, 2019 - link
I think saying that you are a "selective optimist" is imho probably closer to the truth. A bit like people seeing their favorite sports team winning the next trophy simply because it's always been their favorite team.That is fine though - to each their own.
name99 - Wednesday, May 8, 2019 - link
So, 10m was too aggressive and the way to fix that is claim that you'll do an ADDITIONAL 2.4x density improvement with 7nm?What EXACTLY did they learn from 10nm? Because it doesn't appear to be "STFU until you KNOW the process works". And it doesn't appear to be "make small improvements to the process every year instead of trying to change everything all at once for a massive density boost".
blu42 - Thursday, May 9, 2019 - link
It's the good-old 'double-or-nothing' desperate gambler tactics.eva02langley - Wednesday, May 8, 2019 - link
7nm node in 2021... I will believe it when I will see it.By that time, AMD will be on 5nm.
smilingcrow - Wednesday, May 8, 2019 - link
I have more faith in TSMC's 5nm right now but it will be interesting to see how the two compete.Much too early to say really.
brakdoo - Wednesday, May 8, 2019 - link
TSMC 5 nm is in risk production. They don't do that unless they reach a reasonable yield. Next year Apple and Huawei seems very likely.2021 would probably be a half node advanced in mass production, like N5+ or N4.
brakdoo - Wednesday, May 8, 2019 - link
The real race is about gate-all-around replacing FinFets and delivering much higher performance, not just power and area optimization. That would probably be TSMC and Samsungs 3nm...tipoo - Wednesday, May 8, 2019 - link
The latest density estimate they showed had 10nm tied with TSMC 7nm, so AMD being on 5nm vs Intel 7nm would be pretty comparable.Hifihedgehog - Wednesday, May 8, 2019 - link
> 7nm in 2021Haha... I'll believe it when I see it. They've said this countless times before. Fool me once, shame on you. Fool me twice, shame on me.
HStewart - Wednesday, May 8, 2019 - link
They only failed with Cannon Lake - all of the other stuff - is just BS on Internet.Wilco1 - Wednesday, May 8, 2019 - link
Or maybe you're the one with all the BS?eva02langley - Wednesday, May 8, 2019 - link
Remove your fanboy glasses. 10nm is a huge blow for Intel.name99 - Wednesday, May 8, 2019 - link
Actually no, the lies about great yields, shipping soon, started with Broadwell and 14nm. It's just that in that case the delay was "only" about 6..9 months, so the true believers could ignore then forget it.And remember it's not 10nm that has been the subject of constant lies. What about all those claims regarding Optane that have now been so aggressively scaled back? And 5G?
One error is reasonable, everyone makes mistakes or is too optimistic. But this is a clear pattern that's been going on for at least five years.
Korguz - Wednesday, May 8, 2019 - link
" They only failed with Cannon Lake - all of the other stuff - is just BS on Internet. " only failed with cannon lake ?? what about 10nm ??????? what about the fact.. that Epyc is on par.. or faster then Xeon when the software can use the cores it is given ?? the only BS that is there.. is the BS that you keep saying HStewart.. when will you wake up and realize your beloved intel.. is losing... failing.. and is falling behind... oh wait.. cause you cant, and you wont admit it..Kevin G - Wednesday, May 8, 2019 - link
Jumping to 7 nm is feasible considering the usage of EUV. This is what Intel should have done for 10 nm but EUV simply wasn't ready and they thought multipatterning could still scale. Sometime you make the wrong bet.The bigger news is that Xe will indeed be using EMIB. This opens the door for scaling between the low and high end by re-using the same piece of silicon, just numerous instances per package. This is something that should not be underestimated in terms of performance. With Intel having their own fabs, they can afford to throw far more silicon into the GPU race, more than even nVidia's insane 818 mm^2 monolithic die. The question is what size will the base Xe die be before they scale up in number? 300 mm^2? 400 mm^2?
name99 - Wednesday, May 8, 2019 - link
Why do you imagine that Intel is the only one with this sort of tech? As soon as it becomes appropriate, I expect nV and AMD to follow the same path (AMD already has, of course, with Ryzen).Packaging is currently undergoing an explosion of different possibilities. It's the shortest of short-sightedness to imagine that Intel's particular entries into this space (EMIB and Foveros) represent the only options, and that Samsung and TSMC have no clue about how to do this, or that their options inevitably suck compared to the Intel's...
Kevin G - Thursday, May 9, 2019 - link
Intel is the only one who also owns fabs. AMD and nVidia have access to other foundries but those are middle men who are out to make their share of money as well. The point is that Intel by merit of having their own fabs could throw more silicon into a product at a lower cost than either AMD or nVidia could. Intel could brute force their way to the performance crown which would be too costly for AMD or nVidia to match.The whole chiplet strategy for GPU is indeed being explored by both AMD and nVidia. There is a nice research paper published by nVidia on this very topic. AMD has strongly hinted at a similar concept for their future GPUs and CPU + GPU parts.
Irata - Thursday, May 9, 2019 - link
" This opens the door for scaling between the low and high end by re-using the same piece of silicon, just numerous instances per package"Um, isn't this exactly what AMD has been doing with Ryzen / Epyc / Threadripper in the CPU space for a while now ?
peevee - Thursday, May 9, 2019 - link
And Intel with Core 2 Quad 13 years ago.peevee - Thursday, May 9, 2019 - link
"The question is what size will the base Xe die be before they scale up in number? 300 mm^2? 400 mm^2?"What is the point? They can have better yield on something really small, 50-100mm2, packaging is cheap and reliable.
Kevin G - Thursday, May 9, 2019 - link
There is still a minimum size desirable. Too many smaller dies makes the packing complex. Easier to link together five 200 mm^2 dies than twenty 50 mm^2 dies. It is also simpler to have one side of the die match the dimension of a HBM module to simplify the design.TristanSDX - Wednesday, May 8, 2019 - link
this means that their GPU is pushed to 2021JasonLD - Wednesday, May 8, 2019 - link
no, it means first 7nm product will be their GPU. They will still release their first GPU on 10nm.twtech - Thursday, May 9, 2019 - link
I'm looking forward to getting a 28/32 core CPU for work. I had been hoping my company would buy it for me, but if they won't, I may end up buying it myself and working from home a bit more.PeachNCream - Thursday, May 9, 2019 - link
Translation - Consumer CPUs will be on 14nm until 2022 when high volume manufacturing begins and will be in short supply until 2023. That's a huge window for competitors with ARM chips and, of course, for AMD to exploit because they're using TSMC to fab.What happened with Intel? Someone goofed up ordering EUV equipment from ASML and they company is going to end up five years behind the competition when its said and done.
Rash1 - Thursday, May 9, 2019 - link
Intel already can manufacturer 7nm chips but why ? Intel can not meet demand of it's 14nm+++ and now shipping 10nm for the first time with improved performance over previous gen. So, milk much as possible from 14nm+++ and 10nm than in two years, release 7nm. Good timing Intel.Targon - Thursday, May 9, 2019 - link
Where's the "improved performance"? If you look at what Intel put out, they are talking about efficiency....isn't that what all the people putting AMD down used to talk about when it came to AMD products? If you can't improve performance, then work on efficiency, but Intel hasn't improved IPC, clock speeds have been stagnant(if you look at what could be done three years ago if you would delid an Intel chip), so where's the true improvement in performance from Intel?AMD got its act together with Ryzen in 2017, second generation brought a 300MHz clock speed improvement plus 3-5% IPC improvement, and third generation will bring an estimated 13-15 percent IPC boost plus a significant clock speed improvement compared to second generation. The stuff from AMD isn't speculation about products that are still 1-2 years away from when consumers have access to them, it's only one to two months away.
HStewart - Sunday, May 12, 2019 - link
I think it best to get information about 10nm directly from Source - Intelhttps://www.bing.com/videos/search?q=intel+lakefie...
I am most excited about Lakefield - but it was interesting to see the performance comparing 10nm cpus physically running again existing 8th and 9th computers.
I am desiring a portable always connected machine, and thought about it and ideally it would be compatible with software I run today on Windows. No need for iOS or Android and not emulated with ARM making applications slow. Lakefield is the answer.
HStewart - Sunday, May 12, 2019 - link
One thing is that some rumors that Intel is out of 5G area is totally FALSE, look at the end of video of a product call Snowridge.umano - Tuesday, May 14, 2019 - link
So they say we will have 10 nm desktop and HEDT cpus before 2020? I don't believe that