Technological tidbits
Techno Talk
Since I offer freelance technical consulting and writing services (“I
write for food”), I end up chatting with a lot of companies about the
latest and greatest technological offerings that are coming soon.
A
s I may have mentioned before
on occasion, I am incredibly lucky.
I’m not referring to my awesome
looks, ready wit, or internationallycommented-on sense of fashion (“you
aren’t going out dressed like that, are
you?”, says my mother back in the UK).
I mean that I get to chat with a cornucopia of companies on the cutting edge
of technology, spanning myriad applications and industries. Let me share a
few examples.
Feel the power
Just a couple of minutes ago, I was
chatting with the folks at AmberSemi
(https://ambersi.com). This small, fabless
(and fabulous!) semiconductor company is poised to upend the power supply
market. I’m talking about the AC-to-DC
power supplies that plug in the wall and
power our electronic doodads, gizmos,
thingamabobs and what have you.
A common architecture for low-power
AC-to-DC converters employs 1950s-era
electromagnetic components like bridge
rectifiers, transformers and high-voltage
bulk capacitors. The folks at AmberSemi
have replaced all of this with a solidstate semiconductor implementation
that is much smaller and lighter, more
efficient, more reliable, and cheaper.
Be the power
Speaking of power, a few days ago
I had a call with the guys and gals at
Circuit Mind (https://circuitmind.io).
When they first came to my attention
a year or so ago, they told me about
their deterministic AI that helps electronic teams go “from architecture to
schematic in 60 seconds”.
You can use their interface to capture
a high-level block diagram (microcontroller, temperature sensor, LCD display
etc), annotate each block with any lowerlevel requirements (8-bit, 1°C accuracy,
640×480 resolution etc), while leaving
anything you aren’t sure about or don’t
care about unsaid. You use sliders to
specify trade-offs like cost, area, power
and component availability.
When you click the “Go” button, their
system selects components, generates the
circuit diagram(s) and provides a bill-ofmaterials (BOM), amongst other things.
In our most recent conversation, they
informed me that they’ve just added a
Practical Electronics | May | 2025
new power supply design module. This
performs all the tasks discussed above for
the power supply portion of the design.
People who can design power supplies (well, good power supplies) are
few and far between, so this capability will be worth its weight in gold for
many teams developing embedded systems and the “things” part of the IoT.
Losing count
Reticles, also known as photomasks,
are critical components in the production of silicon chips. The high-end
(size-wise) for today’s biggest and baddest silicon chips is typically around
625mm2 (25 × 25mm). The absolute
limit is around 800mm2 (28 × 28mm).
Recently, I was chatting with the lads
and lasses at Synopsys (https://synopsys.
com), a company that makes, among other
things, widely used IC design software.
They were excited to give me their predictions for chiplets and multi-die systems
in 2025. As part of this, we touched on
the new Blackwell graphics processing
unit (GPU) from Nvidia.
There are two main flavours: the
B100 and the B200. The latter consists
of two 800mm2 silicon dies joined by
a 10 terabytes-per-second (TBps) chipto-chip interconnect, forming a unified
single GPU. In the same package, the
B200 also contains 12 high-bandwidth
memory (HBM) stacks, each of which
contains multiple DRAM dies.
Excluding the HBMs, the B200 boasts
208 billion transistors (104 billion for
each main die). How much power does
this bodacious beauty consume? I’m
scared to think, but…
Strike while the die is hot
A few days ago, I had an interesting conversation with the chaps and
chapesses at Saras Micro Devices
(https://sarasmicro.com). We started
by looking at an image of an H100-based
accelerator card that plugs into a server
motherboard in a data centre, where the
H100 (Hopper) GPU is the 2022/2023
predecessor to the Blackwell.
The H100 is a big package that occupies about ¼ of the accelerator card’s
real estate. The remaining ¾ of the
card is taken up with power conversion and delivery components. 48V DC
coming in is converted to 12V, which is
Max the Magnificent
subsequently stepped down to the pointof-load voltages. I’m told we are looking at
1000W <at> 1V, which equates to 1000A.
In the not-so-distant future, when core
voltages drop it 0.7V, we’ll be looking
at 1500A! Think about this; 1500A on a
single, relatively small accelerator card,
with 2000A heading our way. This is
known as ‘horizontal power delivery’,
because the power travels horizontally
across the board from the power conversion components to the main processor.
This is where the folks at Saras leap
onto the stage with a fanfare of flugelhorns (once heard, never forgotten). Their
goal is to remove the inductors and capacitors that consume so much valuable
real estate, and to embed these components in the board or package substrates.
This will pave the way for ‘vertical
power delivery’, with much shorter
tracks that will dramatically reduce
IR voltage drops and associated I2R
thermal losses.
Memories are made of this
There are lots of different semiconductor memory technologies, the
main contenders being flash, which is
relatively slow but non-volatile (it remembers its contents when power is
removed from the system); dynamic
RAM (DRAM), which is cheap, has high
capacity, and consumes little power;
and static RAM (SRAM), which is expensive, has a low capacity, and is fast,
but consumes lots of power.
When designing custom silicon chips
like system-on-chip (SoC) devices, onchip memory like caches and blocks of
memory are usually made of SRAM to
provide the highest possible performance.
Other memory technologies, such as
magnetoresistive RAM (MRAM), have
attractive qualities, like non-volatility
and low power consumption, but they
are significantly slower than SRAM.
I was recently conversing with the
folks at Numem (https://numem.com).
They have created an MRAM-based
memory technology called NuRAM with
a ‘SmartMem’ subsystem that results in
MRAM with SRAM-like performance.
The scary thing is that I talked with
all the companies mentioned here in
just the past couple of weeks. Who
knows what we’ll learn in the days
and weeks to come?
PE
3
