Posted by my123 2 days ago
like:
- home.kpmg - global.honda
and probably more I forgot about
but as you can see, they're pretty terrible for replacing .com domains
I know that for GPU's the model weights have to be transferred over the bus initially, but that only has to occur once for inference use cases, so is the Fujitsu system more about training scenarios? Or is the focus more about efficiency, as these are ARM-based cores with AI additions?
It's an overcrowded market. Far better would be to focus on semiconductor supply chain, which Japan already supplies some elements, to sell to fabs.
Because the bottleneck is the fabs. If a new 2 nm fab came online today, it would immediately sell all its capacity to 2030 no problem, without Fujitsu trying this gambit.
It's not 'sovereign', the architecture is not designed in Japan, the silicon is not fabbed in Japan. This whole thing is sideways.
The text says "next-generation CPU, FUJITSU-MONAKA, designed and developed in Japan".
Doesn't matter. You have to start somewhere and this is a good start.
So there's a good chance that it's still an optimized half-SPARC inside.
You have been convinced by fake news, please recheck this claim, its false.
> wealthy people
*rich people
> All they know how to do is tax people
Oh god you're one of those. People who complain about taxes have historically been on the wrong side of arguments.
Oh you're one of those. The arrogance is astounding. You never been a country which overly taxes it's citizens and squanders the gains on bureaucracy and monetary black holes? Grow up, your ideological blind spots are making you look like a fool.
And here's more technical information about this generation: https://news.ycombinator.com/item?id=49443040
> CapeTheory on Dec 14, 2024
> Fujitsu seems to be weirdly stubborn about global commercialisation of their CPUs. I know of at least a couple of HPC customers who wanted to build systems based on A64FX and the answer was basically "...nah we're good".
Kudos to Japan
Fujitsu MONAKA Server will be
made broadly available from
November 2026 to data center
operators, enterprises, and
the academic and HPC sectors
in Japan and Europe, as well
as to the defense sector,
contributing to national
security.
There is a fairly direct link between the two numbers. You can predict the latter from former reasonably well
You have to run inference on the GPU by reading and writing to VRAM. So TFLOPS of the compute matters, and bandwidth to the VRAM (Always integrated with the GPU, rarely a bottleneck), and this strongly affects tokens/s
If you're doing training workloads or offloading to system RAM, it gets more complicated. (And mostly bound up trying to feed compute on time)
(Edits for clarity.)
On dedicated inference hardware I'd expect model weights to never leave the RAM, and you'd probably load them on startup before even starting to serve requests
I do find it odd that the 2U model has less storage capacity than the 1U on their chart. That doesn't make much sense to me.