Tenstorrent QuietBox tested: A high-performance RISC-V AI workstation trapped in a software blackhole

Tenstorrent probably isn't the first name that springs to mind when it comes to AI infrastructure. But unlike the litany of AI chip startups vying for VC funding and a slice of Nvidia's pie, Tenstorrent's chips actually exist outside the lab. If you're feeling a little non-conformist and want to color outside the lines, …

[trevorde]

Fine but can it run Crysis?

[Dizzy Dwarf]

Can it run Doom?

[RAMChYLD]

It's RISC-V

If you enjoy playing your Doom through a remote X Window, then yes it can run Doom. Probably more instances of Doom than you can shake a stick at too.

[Anonymous Coward]

No but would hope it has a copy of Zork/Lander on it. That's the only game I remember from the Risc machines of old

[Anonymous Coward]

Every BBC Micro had to include Grannies Garden and that teletext simulator thing. Every Archimedes had to have Lander/The Crystal Maze. It was the law (maybe).

Meanwhile, Windows got lumped with Minesweeper.

Look how miserable America is now! I'd suggest that the world would be a very different place if they'd grown up playing Grannies Garden/The Crystal Maze.

[druck]

Zork/Lander? You mean Zarch?

[Anonymous Coward]

Thanks for a useful article.

This is a very nice review that shows gives some indication of what the 'Quietbox' can do and the potential further down ther line.

I would be interested in how Tenstorrent takes onboard your comments and improves the final shipping version.

For what it is $12k does not seem too dear BUT there is too much unused potential as it stands !!!

Worth keeping an eye on in my opinion.

:)

[DJV]

I wonder...

...how many of those I will find under the Christmas tree this year...

At $12k each, probably slightly less than one!

[fromxyzzy]

It's not an SGI O2, but workstations are back on the menu.

[Anonymous Coward]

Respect

This should give an opportunity to relatively regular Joes and Janes to test-out dataflow computing in a workstation format rather than the more common racks (see also SambaNova RDUs, Groq LPUs, Cerebras WSE-3/CS-3, d-matrix DIMC, NextSilicon Maverick-2, ...). Seems to me that the architecture can help optimize both the sequential and parallel components of the computational graph together, in ways that disjoint CPU+GPU archs struggle with ...

In due time, we should have such computational elements (eg. baby RISC-Vs) density-distributed throughout HBM stacks, providing 3-D Hybrid distributed in-memory compute elements to further flexibly uplift compute perf by breaking out of the Von Neumann single file chaingang penitenciary execution assembly line conveyor belt (computational death row) and into the newfound parallel liberty of impedance-matched data-race-free (DRF) weak memory model compute where algorithms interact in frictionless harmony with their hardware environments, rather than constantly trying to beat them imperatively into rather unnatural submission, square-pegging the round holes and all, with every screw and bolt a nail to the rigid single-purpose hammer and suchlikes, imho.

It's like a party with many people in it, all interacting profusely, rather than just one ... much more enjoyable! ;)