Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/intel-says-it%E2%80%99s-on-course-to-regain-chip-production-leadership.17167/page-2
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Intel Says It’s on Course to Regain Chip Production Leadership

Don’t many 2D and 2.5D mcm designs not have a parent die they stack onto? Unless the parent die is just the first die that is placed on the package.
 
Why do we even call them chiplets? As far as I know IBM just called each one a die that was part of an MCM design.

It is a marketing term just like IoT. There is even a Chiplet Summit so marketing works:


I'm surprised www.chiplet.com is still available.
 
It is a marketing term just like IoT. There is even a Chiplet Summit so marketing works:


I'm surprised www.chiplet.com is still available.
Is there a naming convention standardization body or committee?
 
Is there a naming convention standardization body or committee?
Not for marketing terms. Even when industry standards bodies name things, many people often choose to use terms they make up. For example, cabled Ethernet does not have a switch specification. Ethernet "switches", as the industry calls them, are actually called "bridges" in IEEE 802.1 WG specifications. Ethernet links are correspondingly called "segments", and what the industry calls switches "bridge" between segments. (The Ethernet specs still support CSMA/CD buses.) All of this legacy and ancient terminology does not please marketing people much, so they call Ethernet components whatever they want.

 
Yeah it seems weird, given semi folks don't use other marketing names to refer to all other implementations of said technology. For example trigate (finfet), Hyperthreading (SMT), Infinity cache (L3 cache on a GPU), InFo (fan out), X-cube (3D-stacking), RibbonFet or MBCFET (HNS).
 
Yeah it seems weird, given semi folks don't use other marketing names to refer to all other implementations of said technology. For example trigate (finfet), Hyperthreading (SMT), Infinity cache (L3 cache on a GPU), InFo (fan out), X-cube (3D-stacking), RibbonFet or MBCFET (HNS).
That's because there isn't any marketing necessary for any of that esoteric technology. No one outside of semiconductor of computer system geeks care one bit about any of those terms. Intel tried to market SMT, and got all wrapped around the axle trying to explain the difference between a CPU hardware thread (SMT) and a CPU core.
 
That's because there isn't any marketing necessary for any of that esoteric technology. No one outside of semiconductor of computer system geeks care one bit about any of those terms. Intel tried to market SMT, and got all wrapped around the axle trying to explain the difference between a CPU hardware thread (SMT) and a CPU core.
In this case how is chiplet any different? To the end consumer it literally doesn't matter how many dies their device has nor how they are connected.
 
In this case how is chiplet any different? To the end consumer it literally doesn't matter how many dies their device has nor how they are connected.
Because chiplet technology is being marketed within the industry, not to consumers. In client CPUs, SMT was being marketed to consumers.

The most interesting intra-industry marketing going on right now is for CXL, in my opinion. Many, many companies working on products, lots of excitement among data center operators (especially the big cloud companies), much implementation complexity in hardware and software (and a 900+ page specification)... it should be interesting to see how the implementations perform.
 
That's because there isn't any marketing necessary for any of that esoteric technology. No one outside of semiconductor of computer system geeks care one bit about any of those terms.
So why do industry folks accept this industry marketing term, and not others. I don’t think that every fab’s extra DCTO opportunities will just be called finflex. Nobody will ever call HNS MBCFET or ribbonFET. None of these terms are targeted at consumers either. My only guess is people thought that “disaggregated die”, “the x die in our soc”, or “x componet of our mcm design” isn’t as fast as “our x chiplet”. Whereas something like finFET is just as fast as trigate.
 
So why do industry folks accept this industry marketing term, and not others.
I can only guess. I have noticed over the many years I was in the business that some folks get a kick out of defining their own terms and having the industry use them. Also, some companies really don't like other companies, and refuse to use that company's terminology.
 
Not for marketing terms. Even when industry standards bodies name things, many people often choose to use terms they make up. For example, cabled Ethernet does not have a switch specification. Ethernet "switches", as the industry calls them, are actually called "bridges" in IEEE 802.1 WG specifications. Ethernet links are correspondingly called "segments", and what the industry calls switches "bridge" between segments. (The Ethernet specs still support CSMA/CD buses.) All of this legacy and ancient terminology does not please marketing people much, so they call Ethernet components whatever they want.

Thanks that also explains why articles about Ethernet are so confusing to read!

All the optical stuff now becoming predominant will likely spur more marketing innovations, if it hasn't already.
 
In my opinion, always use Webster's dictionary. My company policy is that the receiver of information must regurgitate back in their own words. Communication is key.

IMO, the industry is changing so fast, you should word things generically. The exact implementation gets negotiated with the customer. As I see it, interposer, MCM, hybrid, etc is all the same thing.

900 page spec... are you kidding me? Did Congress write it?
 
900 page spec... are you kidding me? Did Congress write it?
In a way, a Congress-like entity did write a lot of it. While Intel wrote the first version of the spec, like they did with PCI, PCIe, UCIe, and NVMe. Intel likes to get the industry behind them by forming consortiums, and based on the model they created for PCI. There's a board of directors, which consists of representatives from the industry leaders negotiated by Intel, and they approve working group directions, proposed specifications, etc. Then there are voting member companies, and there can be many tens of these. Personnel from the voting members can become part of the specification working groups, and within working groups decisions are made by voting, and it's on a "one company, one vote" basis. This structure creates a highly political environment, where companies negotiate with other companies in a you-vote-for-mine-and-I'll-vote-for-yours arrangement. So you end up with a feature-heavy specification, where the features are often based on work one company has done internally, and may have already filed invention disclosures. Or sometimes even have a design and test silicon for. The patents aren't a problem per se, because all of these consortiums force companies to license their spec-related IP on a reasonable and non-discriminatory (RND) basis. But this political stuff leads to fat specs with sometimes bloated and odd features.

CXL is also interesting in that Intel conceived it for creating a CPU-centric coherent memory domain for use with accelerator devices in a modest radius, and the industry took the original design and now focuses on a distributed datacenter memory fabric. You know, like Congress would with earmarks and riders on a bill. :)
 
Back
Top