It should come as little surprise to our readers that have followed news about Kaby Lake, Intel's extension of the Skylake architecture that officially broke nearly a decade of tick-tock processor design. With tick-tock, Intel would iterate in subsequent years between a new processor microarchitecture (Sandy Bridge, Ivy Bridge, etc.) and a new process technology (45nm, 32nm, 22nm, etc.). According to this story over at Fool.com, Intel's officially ending that pattern of production.
From the company's latest K-10 filing:
"We expect to lengthen the amount of time we will utilize our 14 [nanometer] and our next-generation 10 [nanometer] process technologies, further optimizing our products and process technologies while meeting the yearly market cadence for product introductions."
It is likely that that graphic above that showcases the changes from Tick-Tock to what is going on now isn't "to scale" and we may see more than three steps in each iteration along the way. Intel still believes that it has and will continue to have the best process technology in the world and that its processors will benefit.
Continuing further, the company indicates that "this competitive advantage will be extended in the future as the costs to build leading-edge fabrication facilities increase, and as fewer semiconductor companies will be able to leverage platform design and manufacturing."
Kaby Lake details leaking out…
As Scott pointed out in our discussions about this news, it might mean consumers will see advantages in longer socket compatibility going forward though I would still see this as a net-negative for technology. As process technology improvements slow down, either due to complexity or lack of competition in the market, we will see less innovation in key areas of performance and power consumption.
Kaby Lake graphic
Kaby Lake graphic above.
5K@30 one display, 5K@60 w/two displays.
That makes no sense…. shouldn’t it be like 5K at 60 with one and @ 30 with two?
Heh weird, maybe a typo or
Heh weird, maybe a typo or they are combining the two 30Hz into one for some unknown reason 0.o hehe. Learning from wireless router marketing?
Unless they mean 5k@30 with
Unless they mean 5k@30 with one display output, and 5K@60 with two display outputs? As in, 60 is only available in multi-link? Just speculating, though. No idea.
Well AMD optimized its
Well AMD optimized its Carrizo APU at 28nm and got around 30% x86 core die space savings by using its high density GPU design libraries on the layout design of Carrizo’s excavator CPU cores. So there is nothing wrong with Intel having an optimization phase with its designs if it saves space without needing a node shrink. Intel really was relying on its fab process die shrinks a little too much and not paying attention to other methods to save die space or get more functionality using design layout libraries.
Intel needs to devote more time to its graphics drivers and maybe other things, before AMD’s APUs on an interposer put Intel at a disadvantage for primary RAM memory bandwidth. AMD’s use of high density design libraries for any 14nm CPU cores will still net AMD that around 30% space savings over CPU cores that use the standard low density design libraries for CPU core layouts. AMD will be able to fit more of its x86 Zen cores in that 14nm space and offer either smaller dies, more CPU cores, or more GPU cores to its APU graphics.
The Zen CPU core still may not be able to best Intel’s offerings in IPC, but AMD will be able to get SKUs at a more affordable price points with 8 cores than Intel can offer with 4 of its cores, with AMD using optimizations like High Density design libraries, and creating APUs on an interposer with a separately fabricated Zen cores die and a separately fabricated GPU die, in addition to HBM stacks all being wired up via an interposer. That single CPU core IPC metric is becoming less important as the major graphics APIs(Vulkan/DX12) will allow for better CPU multi-core utilization, and also for more compute workloads to be moved off on the CPU cores and onto AMD’s ACE units where AMD has the lead in asynchronous compute done on the GPU.
The affordable die shrink era is coming to a close, even for Intel’s deep pockets, and atoms are not getting any smaller so it’s time for space savings through other means like 3d die stacking, interposers, and creative circuit design layouts.
Spoken like a true amd
Spoken like a true amd employee , right?
All Intel CPUs in my laptops,
All Intel CPUs in my laptops, but looking to get something with an AMD SOC if the Linux laptop OEMs will start offering the SKUs. So I do not work for any chip makers, the filthy parts suppliers that they are. AMD will have some very competative Zen SKUs and single core IPC is not as necessary, what with the software stack allowing for more compute to be offloaded to the GPU. Intel’s graphcs still suck if you want to do more than just gaming or your device. AMD’s and Nvidia’s graphics will always be better than Intel’s if you do non gaming graphics workloads with any device.
AMD, Intel, and Nvidia are merely parts suppliers to the PC/laptop/other devices OEMs, so I’m nore intrested in the technology part of thier offerings and AMDs Interposer technology has great potential to offer much better onboard graphics for AMDs Zen based APU SKUs, to go along with HBM’s ability to provide massive amounts of effective bandwidth while saving on power usage, way the hell more than any standard DDR/GDDR memory over narrow memory channels. So It’s about the technology, and not only about just the technology that is good for gaming, but technology that is good for non-gaming graphics uses also!
Nvidia get Your IN Hardware Async compute deficiency fixed, or you will be behind, fix your hardware’s compute resources for better processor utilization. It’s about the technology and who is better on the latest.
And should we expect that AMD
And should we expect that AMD will continue to produce massively inefficient chips that require cooling devices that can only fit inside a XL tower chassis, and consume 300% more watts for 30% less processing?
So, this just splits out the
So, this just splits out the ‘optimization’ step that was normally folded into Tock?
Looks like 10nm is the new 28nm.
Yes the ‘optimization’ step
Yes the ‘optimization’ step will be done over a longer time period for 10nm and below process that are going to cost more! Moore’s law was more about the economics driving transistor counts doubling every 18mo to 2 years. So even At Intel’s economy of scale it will take longer to fully amortize the R&D and chip fabrication costs for process shrinks 10nm and smaller. And the mobile market is where the unit sales volumes are that allow for the R&D/engineering and chip fabrication costs to be quickly amortized. But Intel’s mobile market sales are not going to be at as large enough volume owning to the fact that the ARM based industries has the majority of the mobile devises market sales.
So the third party fab businesses that is supplying the needs for all that is ARM/MIPS/Other ISA based markets is seeing the sales volumes of mobile at an even larger scale than Intel saw making x86 based SKUs for the PC/laptop market. Furthermore the third party chip fabrication industry is able to spread its R&D and chip fabrication facilities costs across an entire industry of makers while Intel has its fabs mostly producing for Intel’s products alone with some minor custom fabrication business between Intel and some third party makers of SOCs. The third party fab industry is also now supplying IBMs in house power8/8+ and power9 fabrication needs, along with any third party licensed from Openpower power8/power8+ and power9 needs for those power* licensees.
Intel was only to able to outspend the other chip makers for the expensive third party chip fabrication machines and the associated R&D cost that allowed Intel to outspend all others when the x86 market was the only game in town but things have changed. The chip fabrication IP/equipment that the entire world gets from the 6 companies that provide the world with the chip etching machinery and other fabrication IP. So the third party fabs are now at 14nm and already talking 7nm for some future Apple contracts.
Intel will have to keep up with a mobile market sales volume that produces many more chips for much more unit sales volumes and companies like Apple do not have to spend billions for chip fab plant upkeep or chip fab process node R&D, as the third party chip fab business will spread those development cost across the not only for Apple’s SOC needs but for many other SOC makers also. Intel will have a big problem not just only from AMD’s new Zen based APUs, but also from any and all of the ARM market players including from AMD’s upcoming custom K12 ARMv8A running APU SKUs that will be coming online in 2017.
Intel is loosing billions on the mobile devices market and having to hide those losses by folding the its mobile devises division into a larger division to hide the bleeding on the books. The x86 market is about to become more competitive with Zen’s release, and add to that from AMD’s APUs and APUs on an interposer as HSA will allow for even more compute workloads to be done on AMD’s APUs(x86 or ARM ISA based) for enhanced GPGPU acceleration on the APU, and AMD future Interposer based APUs! HSA is stating to pay off for AMD with both DX12 and Vulkan allowing more gaming compute workloads to be accelerated for VR on AMDs APUs. And HSA will pay off in the server market once AMD reenters that market also for x86 and ARMv8A based SKUs from AMD.
The sales volumes are just not there anymore for Intel to keep tick-tock-ing along and not able to as quickly amortize R&D and other engineering and facilities costs! So that era is over. The margins are going to have to come down for Intel also for the Cash Cow server market also from more than AMDs x86 but also ARM and power8/Power8+ and power9 from more than just IBM’s power sales but third party power* sales also!
Edit: folding the its mobile
Edit: folding the its mobile devises division
To: folding its mobile devices division
I don’t think “loosing” means
I don’t think “loosing” means what you think it means.
I think most people
I think most people understand that a change was coming, because ever since Bulldozer Intel has really taken a backseat for the desktop CPU market and focused primarily on the mobile market
I got a bad feeling about
I got a bad feeling about this.
The fact Intel said this means…….they are very very confident their product already beaten AMD Zen before it even out……
I got this feeling when intel
I got this feeling when intel blocked non-k overclocking.
Or they’re simply stuck with
Or they’re simply stuck with the laws of physics and can’t iterate as fast as they used to.
Intel has an 18 core Xeon
Intel has an 18 core Xeon they could rebrand to desktop anytime they want. Or they could use a 6 core, or the 8 core, or the 10 core, 12 core, 14 core, or 16 core. They literally just need to change the laser etcher template from Xeon to Core iwhatever.
From Tick-Tock to
From Tick-Tock to Tic-Tac-Toe! The new rhythm by which they devise their devices…
Higher and higher clock
Higher and higher clock speeds met their end to multiple cores. Now that shrinking process is getting down to the number of atoms you can count I predict adding more cores will also come to an end. Well an end as we know it. I think we will be seeing more of the system on chip and/or cores evolving into processing units where you no longer count a CPU by its cores.
Oh, *now* it’s official.
Oh, *now* it’s official. Sounds like they’re still pretending that Broadwell doesn’t exist.
This is quite big. Do you
This is quite big. Do you guys plan to investigate as to why exactly Intel took this decision? I doubt its anything to do with the limits of their own technology.
It is pretty obvious that
It is pretty obvious that there are yield issues at these smaller process nodes, so I expect it definitely has to do with the limits of their own technology. In fact it has to do with some basic physical limitations. They are now trying to make 14 nm features with 193 nm light. It is amazing that they managed to stretch it this far. Billions of dollars are being poured into research, but I don’t think they have a clear path to a solution yet. It is pretty basic science that exponential functions can not be maintained indefinitely, so CMOS scaling will come to an end. They still may be able to make improvements to the process tech without going smaller.
How do you know how expensive
How do you know how expensive that specific research is? How do you know it is poured?
We must be aware how
We must be aware how complicated is to switch from 10nm to, let’s say, 7nm. Switching from 7 to 5 will be much more complicated. At this sizes we start talking of few atom distances and nobody knows what is waiting there.
On other side chip must be verified and this is very complicated and takes ages to simulate. Even a small change will require weeks to verify.
But shrinking also makes new problems on software side. Today most of software is not optimized and situation will get worst.