SoylentNews is people
SoylentNews
Intel Targeting Zettascale (1000 Exaflops) by 2027?
[Intel CEO Pat Gelsinger] showed a chart tracking the semiconductor giant progressing along a trend line to 1 trillion transistors per device by 2030. "Today we are predicting that we will maintain or even go faster than Moore's law for the next decade,"[*] Gelsinger said.
[...] In a Q&A session after his keynote, Gelsinger revealed that achieving zettascale computing using Intel technology "in 2027 is a huge internal initiative."
Intel Aims For Zettaflops By 2027, Pushes Aurora Above 2 Exaflops
"But to me, the other thing that's really exciting in the space is our Zetta Initiative, where we have said we are going to be the first to zettascale by a wide margin," Gelsinger told The Next Platform. "And we are laying out as part of the Zetta Initiative what we have to do in the processor, in the fabric, in the interconnect, and in the memory architecture — what we have to do for the accelerators, and the software architecture to do it. So, zettascale in 2027 is a huge internal initiative that is going to bring many of our technologies together. 1,000X in five years? That's pretty phenomenal."
[...] If you built a zettaflops Aurora machine today, assuming all of the information that we have is correct, it would take 411.5X as many nodes to do the job. So, that would be somewhere around 3.7 million nodes with 7.4 million CPUs and 22.2 million GPUs burning a mind-sizzling 24.7 gigawatts. Yes, gigawatts. Clearly, we are going to need some serious Moore's Law effects in transistors and packaging.
If Intel doubled compute density every year for both its CPU and GPU components, it would still take somewhere around 116,000 nodes to do the zettaflops trick. And if it could keep the node power constant — good heavens, that is a big IF — it would still be 772 megawatts. Lowering the power and the node count while driving up performance by a factor of 411.5X on the node and system level ... tilt.
And here we were thinking the next five years were going to be boring. Apparently, we are going to witness technical advances so great they will qualify as magic. We look forward to seeing how this Zetta Initiative unfolds. You got our attention, Pat.
Intel CEO Pat Gelsinger Says Moore's Law is Back
Intel CEO Pat Gelsinger says Moore's Law is back:
Moore's Law, the gauge of steady processor progress from Intel co-founder Gordon Moore, has taken a beating in recent years. But it's making a comeback, Intel Chief Executive Pat Gelsinger said Wednesday.
"Moore's law is alive and well," Gelsinger said at the company's online Innovation Day event. "Today we are predicting that we will maintain or even go faster than Moore's law for the next decade."
[...] But miniaturization has faltered as research and manufacturing grows ever more expensive. Chip elements are reaching atomic scales and power consumption problems limit the clock speeds that keep chip processing steps marching in lockstep.
As a result, people use Moore's Law these days often to refer to progress in performance and power consumption as well as the ability to pack more transistors more densely on a chip.
Gelsinger, though, was referring to the traditional definition referring to the number of transistors on a processor -- albeit a processor that could consist of several slices of silicon built into a single package. "We expect to even bend the curve faster than a doubling every two years," he said.
Success will mean Intel just catches up to rivals, a moment Gelsinger has pledged will happen in 2024.
(Score: 0) by Anonymous Coward on Friday October 29 2021, @01:11PM
(Score: 4, Insightful) by Anonymous Coward on Friday October 29 2021, @01:28PM (4 children)
This is all meaningless cheerleading to pump up their flagging stock.
(Score: 4, Interesting) by Freeman on Friday October 29 2021, @02:16PM (3 children)
Sure, it's kind of a PR stunt. Doesn't mean it's not relevant. Sure, I like the bare naked hardware specs as much as the next nerd, but speculative future technology is also interesting. Certainly more interesting than reading about how Fusion Reactors are soon to be realized in the next 25 to 30 years. Again, after the same claim was made 25 to 30 years ago.
Joshua 1:9 "Be strong and of a good courage; be not afraid, neither be thou dismayed: for the Lord thy God is with thee"
(Score: 2) by mrchew1982 on Saturday October 30 2021, @02:12AM (2 children)
But there's no explanation of *how* they plan to do this... The "how" would be news for nerds, some empty boast that we're going to "x" is just pr.
(Score: 2) by takyon on Saturday October 30 2021, @02:40AM
The "how" is mostly going to involve 3D stacking and packaging. They did already outline some steps [soylentnews.org] that would result in "50x gate density" from chips made on the current "Intel 7" process node.
Getting to 1000x in 5 years would almost certainly involve a 3DSoC approach that puts layers of memory (gigabytes) nanometers away from CPU or GPU cores. Even with gains in other areas, there's probably a 100x to account for in there and that's the only way I can imagine it being done. Unless they are lying.
Zettascale implies the relevant products would be an evolution of the Ponte Vecchio GPUs [soylentnews.org] for high performance computing that are in development.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0, Disagree) by Anonymous Coward on Saturday October 30 2021, @07:13PM
So is increasing clock speeds. Increasing clock speeds, theoughput, etc doesn't increase the number of transistors in a cpu.
It's a race between software bloat and hardware improvements, and bloat has won. Get over it.
(Score: 1, Informative) by Anonymous Coward on Friday October 29 2021, @02:02PM (4 children)
and carry a vaporware stick?
So Intel went from being focused and good in both tick and tock, to giving up on both, to proclaiming they will do magic in both.
Like a big ship going nicely full ahead, then idle and reverse, now the wheelhouse calling for a full emergency speed.
The rudder seems ok, but the throttle control loop seems bonkers. Still needs refinement to get going again.
(Suggest verifying the internal and external feedback paths are providing a ground truth.)
The cheerleading for making something great is an improvement over giving up.
But only if it doesn't cause one to forget to actually make the parts and the software base that knows how to use it?
(Score: 3, Interesting) by takyon on Friday October 29 2021, @02:35PM (3 children)
Zettaflops by 2027 is a big claim to make when it's already 2021 and exaflops isn't around.
It can only mean one of three things. 1. They're going to use 3D packaging akin to 3DSoC and that will boost performance by orders of magnitude. 2. They will massage the numbers, like measuring INT4 performance. Or 3. Pat Gelsinger was doing cocaine before his presentation.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by JoeMerchant on Friday October 29 2021, @05:19PM (2 children)
I think it means that the next level tech is sufficiently within Intel's sights that they feel it's in their interests to make this announcement and maybe gauge market feedback from the possibilities.
If it's not a cocaine fuelled fantasy, then we probably will see a return of Moore limited progress. I have always thought that the Moore progress curve was an economically imposed ceiling: maximize profit by drawing out the upgrade cycle on a long stream of sales compelling improvements, keeping the technical progress conservative and predictable careful to not improve so quickly that the market is left wondering what to do with your overcapable new generation of top end products.
🌻🌻 [google.com]
(Score: 1) by shrewdsheep on Friday October 29 2021, @06:29PM (1 child)
I think you are spot on here. Exponential trends are not sustainable. For the first time in history, cost per transistor are rising which is another barrier to further progress.
(Score: 2) by JoeMerchant on Friday October 29 2021, @09:12PM
Could go either way... if they've got a way to do 3D layering and keep the heat acceptable that could open a new world of density: 2 layers next year, 4 layers in 2024, 8 layers in 2026, 16 layers in 2028.
🌻🌻 [google.com]
(Score: 4, Interesting) by richtopia on Friday October 29 2021, @03:08PM (2 children)
https://semiwiki.com/semiconductor-manufacturers/304196-intel-betting-the-farm-shrinking-business-margins-down-for-few-years-but-aggressively-investing-40b-43b-a-year-and-more-with-subsidies/ [semiwiki.com]
To summarize: Intel is recognizing they are behind and are spending like there is no tomorrow in order to catch up and surpass TSMC. The author recognizes how difficult it will be and there is a significant chance Intel won't succeed. However, this risky play is better than the death by cost cutting so many high tech companies have succumbed to. IBM and GE are examples in the article.
(Score: 2) by takyon on Friday October 29 2021, @03:18PM
They could still have node problems like they had with the node formerly known as "10nm", but they should already be in the planning stages for products that will be out in 2027. R&Done?
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 1, Funny) by Anonymous Coward on Friday October 29 2021, @03:28PM
Should've used the successful hype the software industry uses and said Moore's Law is dead, replaced by Moore's Law 2.0 !!! It's 10x as good as the one it replaces !!!
Speaking of which, why hasn't Tim O'Reilly jumped on this meme already? Is he still alive? I know, dating myself...
(Score: 0) by Anonymous Coward on Friday October 29 2021, @03:31PM
"Success will mean Intel just catches up to rivals, a moment Gelsinger has pledged will happen in 2024."
I feel like a "pledge" dumping ground reading the news lately.
So like the rest, sending these to >/dev/null
(Score: 2, Interesting) by Anonymous Coward on Friday October 29 2021, @03:50PM (7 children)
if intel fails, computing will be something for the rich only ... and even they will have to pay thru the nose.
everybody else will be using mobile dumb (but marketed as "smart") phone-home terminals with food-like expiration dates ...
then again, all backend, today, from factory, logistics, stockmarket, inventory, etc etc, the "animal brain of the world economy" is still x86 ...
nobody wants a, for example, warehousing databse to run on a chip and OS welded together that "expires" in 3 years ... then again maybe this will give each factory/industry a leg up that has a in-house design team (like a IT engineering sub-department) that can design a custom chip and set it off to taiwan to print on their chip printers and receive their chip in a fortnight? so it's not about how good your product is but how good (and cheap) you can manage it?
then again, if you're in transport you're prolly not stupid and got stocks in some exxon or shell or whatnot so the dividends offset your cost, or sumething?
(Score: 1, Informative) by Anonymous Coward on Friday October 29 2021, @04:32PM
There is AMD, open hardware, and more.
(Score: 3, Interesting) by JoeMerchant on Friday October 29 2021, @05:22PM (5 children)
I could be happy with Raspberry Pi level compute capability from here to eternity. I would probably have a farm of 100 of them on my home network if they were all I could buy, but at $35 a piece that farm would still be cheaper than my 1991 $5k desktop dev machine.
🌻🌻 [google.com]
(Score: 4, Insightful) by sgleysti on Saturday October 30 2021, @02:30AM (4 children)
I'm with you on Raspberry Pi being sufficient.
What I'd really love way more than increases in hardware performance is software that is more responsive, less resource intensive, and less buggy...
(Score: 0) by Anonymous Coward on Saturday October 30 2021, @03:28AM
Better software could be achieved by "treading water" or even moving "backwards" in hardware-land.
Q: Why does software have bugs?
A: Programming computers is complicated.
Q: Why is programming computers complicated?
A: Computers are complicated
Q: Why are computers complicated
A: It pays more like that
(Score: 3, Insightful) by JoeMerchant on Saturday October 30 2021, @02:47PM (2 children)
The software stack is definitely bloated, and I'm not sure there's a viable cure for that... I've been using the Qt API since 2006, and I'm hopeful that their foray into embedded systems will help keep the algorithms tight on the desktop, but anytime I "step into" in the debugger, every access to an object in a container seems to hop through 4+ layers before reaching the data I'm after. I guess the message is that "containers are for rapid development, bare metal C arrays are for speed."
With Qt and similar, you can at least ditch the OS and fast boot a Pi or similar system, though I must confess: I have never actually done this myself, the convenience of a Linux desktop / CLI is just too alluring and apparently I find it an acceptable trade for long boot times.
I'd really like to see the Pi 4 level performance remain constant and the price drop (which, effectively, it is dropping through currency inflation), the only thing I feel the other Pis lack is RAM, and that's only lacking when you're running a bloated browser (and aren't they all?)
🌻🌻 [google.com]
(Score: 0) by Anonymous Coward on Sunday October 31 2021, @01:29AM (1 child)
Pi 4 performance is unusally slow for using Cortex-A72, owing to either the weak, proprietary GPU or unoptimized hardware acceleration. Low-end x86 is a better choice for almost any Linux user.
(Score: 2) by JoeMerchant on Sunday October 31 2021, @02:07AM
Oh, yeah, my "standard" go-to desktop system for the past 10+ years has been Ubuntu on whatever the latest Core i5 NUC is with 2x whatever the "standard" RAM is for the year. Pis are for playtime. I have a Pi Zero on the boat that serves video to a little screen and bluetooth to the speakers, the Kodi menus are painfully slow, but the playback is flawless. One of the kids has a similar setup in their bedroom to play their movies on a dumb-screen in their room, and the other has a Pi 4 desktop he used to use for web browsing, youtube, minecraft, etc. until I got him a Chinese smartphone that is his go-to for everything now. Then there are scattered Pis around the place doing various headless stuff. The living room Kodi/Netflix i5 NUC is running ZoneAlarm "in the background" watching the outside cameras, and there's a Pi Zero connected to a speaker that plays bird calls when ZoneAlarm detects significant motion, another one is running a continuous MP3 shuffle player that I intend to some-day install in a car or two to replace the radio.
Pi definitely isn't a performance optimized design, I'd say it's a cost optimized design - and that's O.K. I do hope they continue to keep focus on low power consumption, and low price, first and worry about getting better performance out of their chips after the first two are taken care of. I've got two Pi 2 Ws on the way, one to replace the boat media center, and another just to play around with.
🌻🌻 [google.com]
(Score: 2) by Immerman on Friday October 29 2021, @07:34PM
If it's "back", wouldn't that imply it went somewhere?
Moore observed that the number of transistors in a single dense integrated circuit double about every 2 years, later revised to 18 months. Nothing about density or speed, just number. And so far as I've been aware that's pretty much continued unabated.
I suppose they're talking about the implications early on that later overshadowed that simple statement - basically that performance would increase at the same pace. Which for a long time meant clock speed, and then increasingly depended on ever larger and more clever pipelining, the incorporation of more powerful instructions such as vector operators, and eventually building an ever-increasing number of often-slower CPUs into the same die.
I'm not sure how supercomputer-oriented chips are especially relevant though, as the observation appears to have been geared primarily toward consumer hardware. There've been specialty supercomputing chips that consume the entire usable area of a wafer for what, a decade or more at this point? And as the reduction of feature size has slowed to a crawl, for Moore's law to continue in the supercomputing realm would require exponential growth in either wafer size or layers.