This is only a preview of the February 2024 issue of Practical Electronics. You can view 0 of the 72 pages in the full issue. Articles in this series:
Items relevant to "Active Mains Soft Starter":
Items relevant to "ADVANCED SMD TEST TWEEZERS":
Items relevant to "Active Subwoofer For Hi-Fi at Home":
Articles in this series:
|
Cheeky chiplets
Techno Talk
Max the Magnificent
Instead of a single humongous silicon chip, tomorrow’s devices may involve multiple chiplets in the
form of optical chiplets, libraries of chiplets, die-to-die (D2D) interconnect, and network on chip
(NoC) technology. A chiplet-based future is closer than you think.
I
t’s not so long ago that the
designers of digital integrated circuits (a.k.a. ‘silicon chips’ or just
‘chips’) employed the terms SSI, MSI,
LSI, VLSI, and ULSI to represent Small-,
Medium-, Large-, Very-Large-, and
Ultra-Large-Scale Integration, respectively. By one convention, the number
of primitive gates represented by these
terms were: SSI (1-12), MSI (13-99),
LSI (100-999), VLSI (1,000-999,999),
and ULSI (1,000,000 or more).
In order to ‘compare apples to apples,’ we also used to talk in terms of
the number of ‘equivalent gates’ on a
chip. Most designers – excluding those
working with emitter-coupled logic
(ECL) – adopted the convention that
an equivalent gate was represented by
a 2-input NAND gate, which equates
to four transistors in a CMOS chip.
I find it interesting that the huge number of gates and transistors that can
be implemented on today’s chips has
rendered such fine distinctions moot.
As a result, terms suggesting greater
than VLSI levels of integration are no
longer in widespread use.
How big? How many?
In my previous Techno Talk column
(PE, January 2024), we touched on the
topic of ‘technology nodes’ (a.k.a. ‘process technologies,’ ‘process nodes,’ or
just ‘nodes’), which refers to specific
semiconductor manufacturing processes. One way to think about this is that
the number associated with a process
node represents the size of the smallest
physical structure that can be created
in or on the surface of the chip. Thus,
a 1µm node can have features – like
the width of tracks, for example – of
one millionth of a meter in size.
The smaller the features, the greater
the number of transistors that can be
squeezed into the same area. For example, Apple’s latest and greatest A17
Pro processor, which is implemented
at the 3nm node (where 1nm is one
billionth of a meter), boasts over 19
billion transistors (my eyes are watering just thinking about this number).
The problem is that we are reaching
the maximum physical size of chips
8
that can be constructed using today’s
extreme ultraviolet (EUV) lithographic technologies. The practical limit
is around 25mm × 25mm = 625mm2
(roughly a square inch). It is possible
to create larger devices up to around
29mm × 29mm = 840mm2, but the yield
(good chips vs. bad chips) quickly falls
to unacceptable (unprofitable) levels.
The ‘problem’ is that we have an
insatiable desire for more and more
transistors. Artificial intelligence (AI)
and machine learning (ML) applications are sprouting like mushrooms.
Also, they are migrating out of the
cloud (in the form of data centers) to
the Edge, which is where the internet
meets the real world.
So, what’s the solution? I’m glad
you asked. The industry is currently
buzzing with excitement over the latest developments in…
Cheeky chiplets
The technical term for a semiconductor
chip is ‘die’ (plural ‘dice’). The solution to our capacity conundrum is to
implement a large design across multiple dice, mount them on a common
substrate (base), and present everything
in a single package. In this case, the dice
are commonly referred to as ‘chiplets’
or ‘tiles,’ the packaged product is referred to as a ‘multi-die system,’ and
the substrate is typically a silicon interposer, although organic substrates
(special printed circuit boards) are also
a possibility.
There are four main chiplet use cases.
One is to scale things up by mounting multiple homogeneous (identical)
chiplets (think multiple processing
units) on the substrate. Another is to
take a design that is too big to fit on a
single die, and to split that design into
two or more chiplets. A third use case
is to take functions like transceivers
that have already been proven at one
technology node and implement them
as chiplets that support the main die.
The fourth scenario is to disaggregate
everything into heterogeneous (dissimilar) dice, each being implemented at
the best node (in terms of cost, power,
and so on) for its function.
Today’s chiplet outliers
Until recently, the only companies using
chiplet technology have been outliers
like the folks at Intel who control every aspect of the fabrication process. In
the case of Intel’s Agilex 7 FPGAs, for
example, the main die is supported by
two to six transceiver (XCVR) chiplets.
Another Intel example is provided
by their Data Center GPU Max Series
of devices (these used to be known as
Ponte Vecchio GPUs). These bodacious
beauties boast 47 chiplets and 100+ billion transistors. ‘O-M-Goodness gracious
me,’ is all I can say.
Chiplets for the unwashed masses
The way digital logic designers create today’s system-on-chip (SoC) devices is as
a collection of ‘soft’ intellectual property
(IP) blocks. Many of these IP blocks are
provided by trusted third-party vendors
because there’s no point in reinventing
the wheel (in the form of a processor
core or a USB interface, for example).
The designers also create their own ‘secret sauce’ IPs that will differentiate this
device from its competitors. These IPs
are represented at the register transfer
level (RTL) of abstraction, and all the IP
descriptions are compiled (synthesised)
into the final gate-level chip design.
The dream is to be able to create tomorrow’s multi-die systems as a collection of
‘hard’ IP blocks in the form of chiplets.
In this case, some chiplets would come
from trusted third-party vendors while
others would be developed in-house.
If you’d asked me when this would
come to be six months ago, I would
have guessed it to be five or more years
in our future. Since then, I’ve been talking to companies like Ayar Labs (optical
interconnect chiplets), Arteris (network
on chip (NoC) technology), YorChip and
Zero ASIC (both offering libraries of
chiplets), and Eliyan (die-to-die interconnect). I’m also looking forward to the
forthcoming Chiplet Summit, which will
be held 6-8 February 2024 at the Santa
Clara Convention center. This will be
the place for chiplet companies to ‘see
and be seen.’ All I can say is that I now
think our chiplet-based future is a lot
closer than I’d previously thunk.
Practical Electronics | February | 2024
|