SoylentNews is people
SoylentNews
Intel on Thursday announced that it would delay mass production of its 10 nm processors from 2018 to 2019 due to yield issues. The company has claimed to be shipping some of its 10 nm chips in small volumes right now, but due to cost reasons the firm does not intend to initiate their high-volume manufacturing (HVM) at this time. Intel executives also stated that they are confident of their product roadmap and intend to launch Whiskey Lake and Cascade Lake products later this year.
[...] Intel blames a very high transistor density and consequent heavy use of multipatterning for low yields. Brian Krzanich has said that in certain cases the company needs to use quad (4x), penta (5x), or hexa (6x) patterning for select features as they need to expose the wafer up to six times to "draw" one feature. This not only lengthens Intel's manufacturing cycle (which by definition rises costs) and the number of masks it uses, but also has an effect on yields.
Intel's 10 nm fabrication technology relies solely on deep ultraviolet (DUV) lithography with lasers operating on a 193 nm wavelength at this time. The company's 7 nm manufacturing process will use extreme ultraviolet (EUV) lithography with laser wavelength of 13.5 nm for select layers, eliminating use of extreme multipatterning for certain metal layers. As it appears, right now Intel executives do not consider EUV technology ready for prime time in 2019, so the company's engineers have to polish off the last DUV-only process (again) rather than jump straight to 7 nm.
The delay means another generation of "14nm" products:
Intel does not elaborate whether it intends to ship (in volume) its 10 nm CPUs in the first half or the second half of 2019, but only says that the company’s engineers know the source of the yield problems and are working hard to fix them. As a result, it is pretty safe to assume that the actual ramp of Intel’s 10 nm production will begin towards the second half of next year.
In a bid to stay competitive before its 10-nm CPUs ship in the H2 2019 – H1 2020 (production ramp takes time, bigger processors will launch later than smaller parts), Intel plans to release another generation of products made using its 14 nm process tech. This generation of chips includes Whiskey Lake products for client PCs and Cascade Lake for the datacenter, and both are scheduled for release later this year.
(Score: 0) by Anonymous Coward on Sunday April 29 2018, @01:29AM (15 children)
I just upgraded from Sandy Bridge to Coffee Lake. The speed boost (# cores) x (frequency) is nice, 90 %, but it took 6 long years to happen.
(Score: -1, Flamebait) by Ethanol-fueled on Sunday April 29 2018, @01:54AM (1 child)
I just upgraded from insulting Jews and darker-skinned minorities to insulting the stupid naming conventions of the tech industry.
* Angrily glares at Ubuntu
(Score: -1, Troll) by Anonymous Coward on Sunday April 29 2018, @03:11AM
Since Murdock was murdered by jews, shouldn't it be named Deadian Linux?
(Score: 3, Informative) by takyon on Sunday April 29 2018, @03:08AM (6 children)
Top of the line AMD and Intel CPUs now have 16 and 18 cores respectively. A big and very recent jump in value.
Intel Core i7-2700K was $332 in October 2011. ~10% inflation [bls.gov] since then puts that at about $366. There were some more expensive enthusiast chips, the 3930K at $583 ($643), and 3960X and 3970X at $999 ($1102).
You can get the 6-core Intel Core i7-8700K for under $360, or 8 cores from AMD for as low as $260 (Ryzen 7 1700). 12-core Threadripper for about $670. At $1100, you can get the Intel Core i9-7920X with 12 cores, or 16-core AMD Threadripper for just $880.
It would be nice if the price per core were lower. 6 cores are only now taking the place of quad-core on the Intel side (with Coffee Lake, launched a few months ago), while AMD does 8 cores comfortably at that price point. Maybe those will get bumped up if AMD includes more 50% more cores in the Zen 2 CCX and releases a 12-core chip for under $400. Area reductions at "7nm" and below should make it possible (or physically necessary) to boost core counts.
Your equation doesn't take into account IPC increases. Although IPC increases since Sandy Bridge were often minuscule, those have added up from Sandy Bridge → Ivy Bridge → Haswell → Broadwell → Skylake → Kaby Lake → Coffee Lake. Apparently, that factor is up to +35% IPC between Sandy Bridge and Kaby Lake [anandtech.com] (others put it closer to 20% [hardocp.com]).
If you want to squeeze the last bit of performance out, you could consider AVX2 instructions (it's unclear when AVX-512 will go mainstream now that Cannon Lake appears to be delayed).
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Sunday April 29 2018, @03:37AM
AVX-256 and AVX-512 are bullshit. When a CPU throttles its clock rate to run vector instructions or electrically powers off the top half of its registers to avoid melting itself, that's when you know overeager designers promised "ours goes wider up 512" with no regard for how much performance will suffer when the extra wide-ass registers actually get used.
Benchmark with AVX-128 before assuming anything wider is better.
(Score: 0) by Anonymous Coward on Sunday April 29 2018, @05:27AM (4 children)
It was a consumer chip, 2600k and now 8700k (on sale at Microcenter for $300). It's always easier to pull the trigger when the deal feels good. Odd that Intel doesn't advertise the all-core turbo frequency of 4.3 GHz, since that's when the processor is processing the most. Thanks for the IPC info, but according to the ratio of benchmark results [cpubenchmark.net], the IPC gain is only +3%. The benchmark ratio is 1.899, while (6*4.3)/(4*3.5)=1.843. Even if you use base speeds (3.7 vs 3.4 GHz), which I don't think is correct, the IPC gain is only 17%.
(Score: 2) by takyon on Sunday April 29 2018, @05:45AM (3 children)
Passmark gives a single thread rating for both chips:
2709 / 1942 = ~1.395
1.395 / (3.7 / 3.4) = ~1.28
There's your 20-35% IPC improvement, I suppose.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Sunday April 29 2018, @06:54AM (2 children)
But the CPUs generally run at turbo speeds, not base. Then the IPC gain for single threaded performance is only 13%. Who buys a fast CPU and throttles it down?
(Score: 2) by takyon on Sunday April 29 2018, @01:07PM
https://www.makeuseof.com/tag/forced-induction-intel-turbo-boost-works-technology-explained/ [makeuseof.com]
Turbo is not a complete indicator of how fast the CPU is going to be, and it's not a "you can overclock it to this" number. It will depend on the cooling setup, it could ramp up to the maximum clock for less than a second, etc. You are also unlikely to reach the maximum turbo with more than 1 core active. However, I think Intel publishes more granular turbo clocks for its newer CPUs, showing what 1, 2, 3, 4... cores can boost up to.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by requerdanos on Sunday April 29 2018, @06:39PM
The person who puts it in a box in a hot server room to which he has to drive if there is a failure?
(Score: 2) by bzipitidoo on Sunday April 29 2018, @04:32AM (5 children)
I'm more interested in the huge cut in power usage. Compared to the computer I purchased in 2008, I got about 10% increase in performance, and the big one, a reduction in power use of 80%, And it's in a much, much smaller box than those wastes of space tower cases.
(Score: 0) by Anonymous Coward on Sunday April 29 2018, @05:26AM (1 child)
It's astounding how you trend followers became so obsessed with low power rigs after being obsessed with giant space heating towers with giant power supplies.
If you truly cared about performance per watt then you would ditch x86 entirely and use ARM64 everywhere. Desktop, laptop, phone, cloud. Y u only ARM in yur phone, brah?
I await the day when every forum I read is an echo chamber of trendy bros who agree the inevitable migration to ARM64 everywhere was obviously obvious. See you then.
(Score: 2) by bzipitidoo on Sunday April 29 2018, @12:29PM
Oh, I've tried ARM, tried the Raspberry Pis. They're great for devices with small screens. Runs a GUI okay at a 1080p resolution, but a video playing in a fullscreen browser window can be a strain for it. Depends a lot whether the video is in a format it can decode in hardware. X86 has the muscle to do software decoding in realtime, the ARM not so much. If playing with a new video codec such as AV1, which is not yet implemented in hardware anywhere, going to want that power. Another problem is less X86 vs ARM, it's X86 GPU (Nvidia GEForce or AMD Radeon or even Intel's integrated HD Graphics) vs ARM GPU (Mali).
Anyway, AV1 in hardware is another reason you might want to wait a couple of years if you don't need an upgrade sooner.
(Score: 2) by Ramze on Sunday April 29 2018, @01:07PM
I'm not so much interested in the lower power aspect as the corollaries -- quieter performance, less heat, and as you already mentioned, smaller form factors.
I love that I can set up a fan-less system with an SSD as a multimedia center connected to a 60" TV and not have it heat up the room, drown out the audio with fan noises and hard drive clicks or take up a large area on the TV stand or one that needs a lot of ventilation.
(Score: 2) by toddestan on Sunday April 29 2018, @05:36PM
I don't understand the people who make such a big deal out of power usage, but leave their stuff running 24/7 for no reason whatsoever.
Also, while idle power is much lower, the power usage at max load hasn't really changed much. Want to actually use that new processor at 100% for extended periods of time without having it throttle? Well, better get one of those big "waste of space" tower cases to make sure it can dissipate all that heat.
(Score: 2) by requerdanos on Sunday April 29 2018, @06:42PM
I don't mind that, don't get me wrong, if the speed doesn't suffer.
But all things being equal, I'd rather have a *fast* computer that dims the lights on the whole block when it powers up than a dainty light no-power-using CPU that anxiously weighs the power costs on mother earth before performing any calculation.
I want to be able to *set* it to do that, sure. But if that's baked in and I can't change it, then it is serving a master and owner who is not me, even if I ostensibly bought the thing.