News Tachyum: Prodigy production starts in 2024

[Admin]

Tachyum re-emphasizes plans to start production of Prodigy processors in 2024, claims it is well positioned to capitalize on the rise of AI, HPC, and datacenter demand.

Tachyum: Prodigy production starts in 2024 : Read more

 

[gg83]

What are our thoughts? Is this just another FPGA? Could this be a real chalenger, or is there a requirement for a robust software ecosystem like CUDA or whatever?

 

[Blastomonas]

I suspect that this will not arrive this year or next.

 

[Findecanor]

I see buzzwords.

They have posted YouTube videos supposedly running emulation of other ISAs within QEMU, which is an common open source program, on top of FPGAs.

Their success in the marketplace will depend on many factors, but most of all I think about being cost-effective.

As an assembly geek, I am curious about their ISA, but extremely little about it has been made public.
I'm a bit disappointed that all things point to the current iteration of Prodigy being a (boring) traditional OoO RISC architecture, with no special features announced other than the emulation potential and support for ridiculously low-precision matrix multiplication.
I would like to see security features, like what ARM, Intel and AMD have added to their line-ups in recent years and which are in development for RISC-V.

 

[bit_user]

does not forget to mention 'failed memory and interface IP blocks from key vendor'

You have to sell more than just excuses. Every business has setbacks. Failure to plan for & accommodate these suggests an unrealistic business plan.

Meanwhile, the company admitted that it still has not taped out its processor.

Oof. It'll probably be obsolete, by the time they do.

Launching on a 5 mm node already isn't great. AMD's MI300 is made on N5, for the compute-heavy portions. Nvidia's H100 is made on the N5-derived "4N" node. So, they'll be arriving a late to that party.

Tachyum managed to integrate a baseboard management controller (BMC) and unified extensible firmware interface (UEFI) platforms into the Prodigy FPGA prototype system, the company said.

Creating a new server platform from a clean sheet is a ton of work. Just ask Ventana, whose first generation CPU never achieved volume production. Much of what they're touting about their second gen product seem like baseline features customers would expect of a proper, server-grade CPU.

Imagination and Ventana to Build a RISC-V CPU-GPU Platform

New RISC-V platform could address client applications.

www.tomshardware.com

 

[bit_user]

What are our thoughts? Is this just another FPGA?

There's a very thorough writeup, here:

Tachyum’s Revised Prodigy Architecture

The world of semiconductor startups looking to break into the market is huge, either with some large AI training chip, or some super fast small inference device, or perhaps a HPC focused design for…

chipsandcheese.com

Could this be a real chalenger,

IMO, they're making too many of the same mistakes as their predecessors. History is littered with the bones of over-ambitious HPC startups.

It's just really hard to beat the big industry players, especially when you're trying to do it with an all-around uber chip. If they'd just focused on one niche, like AI, that's achievable (but hard enough). By trying to be the best at everything, they're setting themselves up for failure. They'll be going up against Granite Rapids and whatever AMD's Zen 5-based EPYC is called... not to mention a new generation of server GPUs. It's a fool's errand, IMO.

I think their only hope is government/EU subsidies. I'm pretty sure that's the only reason they're still around. Even then, it's not as if they aren't facing competition from other EU-based efforts.

 

[bit_user]

As an assembly geek, I am curious about their ISA, but extremely little about it has been made public.

They'll have to upstream their GCC patches, before this thing ships. Most likely, they won't even be able to upstream their Linux kernel patches without taking that step, because it's almost guaranteed to contain some assembly language portions.

I'm a bit disappointed that all things point to the current iteration of Prodigy being a (boring) traditional OoO RISC architecture, with no special features announced other than the emulation potential and support for ridiculously low-precision matrix multiplication.

Totally agree.