Nils Cremer

CPUs suck. We're building a new general-purpose chip that scales to thousands of cores while being more energy-efficient. We're hiring hardware design engineers, consider joining us tendrils.co/jobs What we do differently ...

@shipfr8 home

We took over a former technical university. This is Hogwarts in real life. For people who want to work on something too early, too weird, too ambitious.

FR8 is the single one place to be if you want to create the impossible. What the team has managed to put together in the past year is just unreal and what will happen in the next decade is unthinkable

nils is probably the smartest friend i know :) would not miss the chance to work with him to build the future of chips

I had the luck to briefly talk with Nils at the SPH last year. During our conversation I remember thinking “wow” this guy is actually going to change the world! This is a one in a generation company.

@osmarks1 The hardware has some similarities but we have a sane programming model

@nilscmr You've invented Graphcore?

CPUs suck. We're building a new general-purpose chip that scales to thousands of cores while being more energy-efficient. We're hiring hardware design engineers, consider joining us tendrils.co/jobs What we do differently ...

@Paralell_Shazli Going across CPUs is very expensive

@nilscmr How is this different than Hadoop cluster based cpus?

@StefanZeiger No the opposite, lazy evaluation kills parallelism

@nilscmr I'm wondering how you're scheduling lazy evaluation, or dealing with the lack thereof. Avoiding strict evaluation of duplicators seems to be essential for a practical implementation.

@GrigoryEvko We actually support affine type systems :) Have a look at the Fork trait of Vine

@nilscmr Okay but rust has affine types not linear and supporting only linear typed programs is not what reads from this advertisement 🙃

@EmphyrioLives DAGs are static while interaction nets are dynamic. The hardware has some vague similarities with TRIPS. The biggest difference is that they relied on a smart compiler and unfortunately compilers are not smart.

@nilscmr how is an interaction net different from a DAG? And how is your design different from research projects 20+ years ago like TRIPS?

@itsclivetime You still have to decide whether you have cache coherency across your NUMA slices. If yes then you again hit scalability issues, if not your accesses become incredibly slow unless you have very embarrassingly parallel workloads.

@nilscmr core clusters with local caches and their own NUMA slices would probably work imo, if we're talking about massively parallel workloads where throughput is all that matters. not a showstopper if cross-cluster access is way slower than contemporary cpus, just needs to be functional

so happy someone’s finally doing this. most cpu area is branch prediction, icache, etc. to support insane 5-6 IPC speeds. what if you used that area for more cores instead? wide & slow is the future of cpus, networks, memory, and more

@itsclivetime A big advantage is that we can manage memory in hardware without relying on cache coherency because of the linearity of interaction nets (it's the same reason Rust requires neither a garbage collector nor manual memory management)

@itsclivetime Solving memory becomes very hard then. You'll probably not be able to scale cache coherency to that level so one has to go with software-managed scratchpad memories which are hard to program. You can only do that with a few applications (that's what Tenstorrent is doing)

@botir33751732 When's the next one?

@nilscmr Demo at London systems?

@gatelevelanon Fun seeing you reason through this, I'd love to chat!

@nilscmr Okay. I am beginning to believe..

The team at Tendrils reminds me daily that elegance still exists, and standards are something we want to stand for Truly a once in a lifetime opportunity to work with some of the most humble and capable teams I've ever interacted with.

nils, hugo and t6 are ones of the smartest and most humble people that i know in fact, whenever i go outside and speak of technical excellence at @shipfr8, i bring them up i’ve only met 2 teams in my life that are on this level of excellence this is one of them please read the blog on interaction nets, please read the papers behind it albeit the tagline sounds they are doing GPUs but it’s something different, the nuance lies in “general-purpose”

if you understand what he’s saying i highly recommend joining his company nils is great :)

These guys are one of the most cracked teams I’ve ever met working on one of the most interesting problems. Would 1000% recommend checking out there stuff.

@nilscmr and the Tendrils team will outcompete AMD on every level. If you’re bored of your big tech 9-5 you should seriously consider joining them and make your life’s work

Generational company in the making.