Author Topic: Analog In Memory Compute Startup For Mobile | Germany Dresden | Many Positions  (Read 7645 times)

0 Members and 1 Guest are viewing this topic.

Offline fhunstock-semronTopic starter

  • Newbie
  • Posts: 1
  • Country: de
  • My Work account
    • SEMRON AI Homepage

Hi I'm Fritz at SEMRON,

          ==== TLDR ====
  • remote on probation possible, but not preferred
  • main office: Dresden, Germany 🇩🇪 🇪🇺
  • working language: English🇬🇧
  • Semiconductor test (PCB design, test automation)
  • Digital Verification (SystemVerilog, Chisel)
  • Software (compilers, embedded)
  • Digital Design (RTL, Chisel)
  • Analog IC Design (Cadence, ultra low power low area ADCs DACs)
  • Machine learning (quantization, general research)

So if anything above even loosely fits you, please consider looking at our technology, and think if you want to build something amazing with it, we need all sorts of people.

Our core technology is a custom memcapacitive semiconductor device (CapRAM)

Our goal in one sentence is: Running (e.g.) GPT-3/4 (ChatGPT) on your phone, continuously (always on)

At the core of what we do is CapRAM, a new, but remarkably easy to manufacture with existing technologies, semiconductor device.

For the semiconductor physics aficionados, you can look at our Nature Electronics Paper.
But to make a long story short, CapRAM is a capacitor built from silicon electrodes, with a shielding layer in between the electrodes that has a (programmably) limited access to charge carriers that can counteract the electric field at the readout electrode.

So lets start with an Introduction to the basics of what is good to know in order to understand what makes our technology so cool.

Analog compute?
You may have heard of memristors, these are remarkable devices due to the fact that a single (and in turn potentially small) device can both hold a state, and serve a (comparatively) complex function in a circuit.
When you arrange these in a grid, and run a current that is proportional to the input vector of a matrix multiplication (with the previously programmed resistances of the memristors at the crosspoints corresponding to the matrix coefficients) you get the basic building blocks an analog computer. This effectively calculates a multiplication with every single memristor device at once.
That is very neat, and has seen some interest in academia and some startups around that basic concept for some decades now.
  However, as you might have noticed, analog compute using memristors is a bit like fusion power, everyone thinks it's neat, and it'll be dominating everything in just a few more years.
For a very long time now...

We are not doing analog compute using memristors, instead we build systems with memcapacitors, i.e. programmable capacitances.
  Why is this better? Remember 4kTR Noise? (the noise power a resistor produces) Well, Capacitive systems are only subject to kt/C noise, that means you need lower voltages for the same precision calculation, and can even just get more calculation precision with the same effort.

Also, this enables you to build systems with drastically less energy consumption, which directly leads into our next point:

3D Integration
Modern neural networks are.. very large.
Too large in fact to really take good advantage of many of the benefits of both In-Memory compute and analog compute at a system level. Often this is just because of the lack of the ability to scale the manufacturing techniques to billions of parameters at any reasonable density, diminishing the advantages of in-memory at the system level, or just blowing the footprint or power budget of any real application.
  But if you have a technology for which you can manufacture enough memory cells at a reasonable price, and then also use them with enough energy efficiency to actually be able to use them in a system, that just might be what brings all of this over the edge of product viability.

[An illustration from our Nature Electronics Paper about CapRAM]

Alll open positions

Company culture

As a startup we're naturally a very young team, but that is by no means a necessity. Our working language is 100% english, in fact our policy is that as long as anyone wo doesn't speak a language is at all in earshot, we have to default to english.

We want to build an environment where everyone can maximize their learning, and at all cost avoid anyone getting "stuck" in their personal development and learning.

[/size]Where you'll be working

We are located right in the historic city center of Dresden.

(the green is now built up with a historically reconstructed building, in which our offices are located)Photo by Photo Fitti

Photo by "Kolossos"

Photo by Rene Schwietzke
« Last Edit: March 15, 2024, 02:52:42 pm by fhunstock-semron »

Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo