Back in 2008 Fujitsu has one of the best performing 10Gbps Switches. We were building 40 Gbps packet sniffers at Google (4x 10 Gbps NICs) and needed switches that could do things like mirror traffic across ports at line rate. Fujitsu was way ahead of the pack. I always wondered what held them back from building a meaningful networking business in the US.
I fail to see how they will take a significant market share or even break even on this venture.
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.
Fun fact, Fujitsu used to run their own fabs in Japan until they were sold to UMC and now they're another UMC fab although they are stuck at like the 20nm node. So this is likely made at JASM.
>The Phase 2 factory was under construction as of January 2025, next to the Phase 1 facility, and is expected to be completed by 2027.[3] Initially planned for manufacturing semiconductors using a 6 nm process, 3 nm was incorporated into production plans for 2028 in 2026
Yes the lack of transparency on ARM architecture and the obfuscation of the RAM capacity with just the slot information is annoying. Not sure if it is deliberate to sell this at a premium to "mainframe" customers or whether its just their marketing folk not being on the ball. Fascinated by the fact Japan has a 2nm process node but I suspect that too is marketing. Perhaps the 2nm portion of the chip is done at TSMC and the rest of the package done in Japan or something like that. It's not like there's a plethora of 2nm Fabs around the planet to choose from.
It looks to me like it is an AI optimized HPC CPU and while HPC CPUs are actually quite fast, they are not particularly well suited for transformer based neural networks.
It's weird that they don't give out any of the interesting numbers like number of memory channels, how much SRAM they have (CPUs tend to have more of it) or what their expected performance is going to be.
They actually do not in that article. I had to look elsewhere to find that out. Weird, one would expect that a big announcement like this would probably need to be vetted against licensing agreements with Arm that would probably want that fact to be prominently highlighted in press releases.
Yeah, searched the page for common architectures (x86,arm and Sparc since it was Fujitsu) since it'd seem silly to use a totally novel architecture but didn't find anything until noticing SVE2.
Thanks kinda how Japan works. Inventing something new is hard, but taking something that's a cool idea and making it actually usable is well within their wheelhouse.
Not mentioning the architecture is rather strange. Being Fujitsu I'd expect SPARC, but apparently not. Maybe we're reaching a point where some circles either don't care or just defaults to ARM?
Obviously they can just ignore the license in the future and continue development out of a local branch, but it's also a bit disingenuous to speak of "sovereign infrastructure" and then use licensed processor design.
According to a Feb press release, they're fabbed by TSMC:
> For FUJITSU-MONAKA, the 2nm
semiconductors will tape out next year. Due to this, Rapidus would be unable to meet the
deadline, so TSMC is handling the manufacturing.
At the very top of the article:
"Achieving world-class AI inference performance through Japan-developed 2nm 3D-stacked CPU and server integrated, developed, and manufactured in Japan"
"Japan-developed CPU..." plus the server itself (parts, power supplies, case, etc.) integrated, developed and manufactured in Japan. It sounds like they're weasel-wording it a bit, but I don't think the CPU itself is manufactured in Japan, only "developed."
Every developed country is going to push for their own homemade chips. Japan used to make Sparc CPU's back in the day. Probably many others I'm not aware of.
The two made-in-Japan, premium grade, Fujitsu laptops I had to work with were by far the worse pieces of electronic I have ever used. Both exhibited identical defects: fans running continuously at maximum speed, and the batteries would completely die within hours of the devices being powered off. Windows or Linux. And don't talk about that tiny unresponsive trackpad to me. Never again.
The Fujitsu Lifebooks were legendary for their robustness back in the day, I'm genuinely sad you had that experience.
The only fujitsu laptops I used was back in 2011- it was a budget version for on-call and it was fine..
Sounds like the fans running 100% probably contributed significantly to the issue here. If I had to guess it was likely that the C-States were disabled in BIOS somehow so the CPU clock was running at full tilt the whole time, maybe combined with a bad thermal paste job.
I'm not here to defend Fujitsu but I've had really bad experiences with basically every major laptop brand (Dell: majorly bad coil whine and especially faulty soldered RAM, HP: keys vanishing from the keyboard and very weak hinges that break all the time, Apple: The GPU unsoldering itself and the butterfly keyboard shenanigans).
I don't really know any brand with a flawless track record sadly.
Before anyone brings up Thinkpads, they're trading on a reputation that hasn't been true for over a decade. If you have fond memories of a thinkpad it's most likely that you had it from before 2016, or it's "fine", but certainly not great.
I do recall that Fujitsu had a very... unique approach to exposing the Setup option in their firmware. It was just a UEFI boot entry. If that entry was removed, you couldn't get back into the firmware setup; you had to find the name of the setup EFI file and then run that from the EFI console.
What year? The ones I had were the best. Very expensive, but service at the office & at home, very sturdy, very long battery life. That was somewhere early 2000s.
Reaching further back, the Sylistic line was mostly quite good/durable, and I still use my Stylstic ST-4110 when I need to use an old scanner or control a CNC machine on my back deck when cutting tropical hardwoods.
Really miss the transflective display and wish that there were newer devices with such technology
One problem we’re going to have with AI hardware is coming up with a standard set of specifications that are comparable. I don’t really care about the CPU GHz and the memory bandwidth, at least not directly. What I really want to know is how many tokens per second this will deliver, but that also depends on the model. We need a standard metric for that. Perhaps we agree on a specific open weight model (e.g. GLM 5.3 Flash or Qwen vWhatever) and then measure TPS on the hardware of interest.
I always heard of GB/s as most important number ... AI told me their 8800 MT/s on 8 Byte, 12 DDR5-Channels means 845 GB/s. A Nvidia RTX 4090 has 1008 GB/s. Nvidia B200 has 8000 GB/s.
Is this the right way of looking at it?
You have to load the model weights into VRAM over PCI-E (from RAM). So the (PCI-E) bandwidth strongly affects time to first token.
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)
> So the (PCI-E) bandwidth strongly affects time to first token
On dedicated inference hardware I'd expect model weights to never leave the RAM, and you'd probably load it on startup before even starting to serve requests
I've been calling for high-multicore CPUs (at least 100 cores) with local memories for a quarter century now. No winners so far.
A Pentium 4 hit 3.8 GHz longer than that ago in 2004, so that's not special. And the estimated price will be $7,000-10,000, which isn't special either since that's about 10x more than it should be.
Hot take: GPUs disrupted the CPU industry to such a degree that CPUs never recovered, and like the k-shaped economy, the current status quo only serves a small fraction of customers. We can and should do better, but sadly we won't. Still, it's good that Fujitsu did this, for the competition if nothing else.
Unless they are used for multi-processing in classical UNIX fashion, or proper microkernels, most applications will hardly take advantage of them.
Managed languages runtimes are probably the ones that would be better equipped to take advantage of them, for distributed JIT, GC and asynchronous code.
The Connection Machine style with StarLisp.
Very few devs can write optimal multi-threaded code that explores the single digit count of cores on their laptops or phones already.
I think what's most notable about this for me is the reduced environmental impact. The specifications list liquid-cooling as optional and it highlights the CPU as the main processor instead of a power hungry GPU. My takeaway is that the next generation of AI HPC's will focus more heavily on sustainable operations.
Their software is so poor that it caused probably the biggest miscarriage of justice in history. Well that and Fujitsu staff colluding with post office executives to put innocent people in prison rather than admit there was a problem
The 'Fujitsu' that made the shitty software was a British consulting company that was acquired by Fujitsu [1] to farm public contracts. This is made by the Japanese parent company.
You're right, but the scandal is so upsetting that I personally cannot ever look at Fujitsu the same way even. Also, I think you're neglecting that a large part of why it became so bad was that Fujitsu claimed there was no issue with their software. Even if it were built by a company they acquired, the response to the developing situation was something that the parent company should have stepped in to oversee.
https://en.wikipedia.org/wiki/British_Post_Office_scandal
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.
>next-generation CPU, FUJITSU-MONAKA [1], designed and developed in Japan
>Fujitsu MONAKA Server, powered by the FUJITSU-MONAKA CPU, enhances sovereign capabilities through domestic manufacturing,
https://en.wikipedia.org/wiki/Japan_Advanced_Semiconductor_M...
>The Phase 2 factory was under construction as of January 2025, next to the Phase 1 facility, and is expected to be completed by 2027.[3] Initially planned for manufacturing semiconductors using a 6 nm process, 3 nm was incorporated into production plans for 2028 in 2026
It's weird that they don't give out any of the interesting numbers like number of memory channels, how much SRAM they have (CPUs tend to have more of it) or what their expected performance is going to be.
What are the GPU capabilities?
Seems they have decided they don’t want to create their own architecture, but take the popular architecture and build on it.
See: - Blue LEDs - Quartz Watches - Lithium-ion Batteries - Bidets
Edit: it’s funny because this article from 2021 mentioned Fujisu precisely https://siliconangle.com/2021/03/30/arm-unveils-armv9-archit...
Obviously they can just ignore the license in the future and continue development out of a local branch, but it's also a bit disingenuous to speak of "sovereign infrastructure" and then use licensed processor design.
Also odd how big an emphasis they put on AI inference when they don't build the GPU?
> For FUJITSU-MONAKA, the 2nm semiconductors will tape out next year. Due to this, Rapidus would be unable to meet the deadline, so TSMC is handling the manufacturing.
Source: https://global.fujitsu/en-global/pr/news/2026/02/12-01?utm_s...
According to Rapidus webpage, they are explicitly starting mass production of their 2nm in 2027.
Source: https://www.rapidus.inc/en/iim/?utm_source=chatgpt.com
GPU in AI world is essentially a set of specific matrix calculations that this CPU supports on hardware-level. Thought memory speed seems low.
- Japan-developed 2nm 3D-stacked CPU
- server integrated, developed, and manufactured in Japan
The only fujitsu laptops I used was back in 2011- it was a budget version for on-call and it was fine..
Sounds like the fans running 100% probably contributed significantly to the issue here. If I had to guess it was likely that the C-States were disabled in BIOS somehow so the CPU clock was running at full tilt the whole time, maybe combined with a bad thermal paste job.
I'm not here to defend Fujitsu but I've had really bad experiences with basically every major laptop brand (Dell: majorly bad coil whine and especially faulty soldered RAM, HP: keys vanishing from the keyboard and very weak hinges that break all the time, Apple: The GPU unsoldering itself and the butterfly keyboard shenanigans).
I don't really know any brand with a flawless track record sadly.
Before anyone brings up Thinkpads, they're trading on a reputation that hasn't been true for over a decade. If you have fond memories of a thinkpad it's most likely that you had it from before 2016, or it's "fine", but certainly not great.
Really miss the transflective display and wish that there were newer devices with such technology
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 it on startup before even starting to serve requests
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 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.
A Pentium 4 hit 3.8 GHz longer than that ago in 2004, so that's not special. And the estimated price will be $7,000-10,000, which isn't special either since that's about 10x more than it should be.
Hot take: GPUs disrupted the CPU industry to such a degree that CPUs never recovered, and like the k-shaped economy, the current status quo only serves a small fraction of customers. We can and should do better, but sadly we won't. Still, it's good that Fujitsu did this, for the competition if nothing else.
Managed languages runtimes are probably the ones that would be better equipped to take advantage of them, for distributed JIT, GC and asynchronous code.
The Connection Machine style with StarLisp.
Very few devs can write optimal multi-threaded code that explores the single digit count of cores on their laptops or phones already.
Their software is so poor that it caused probably the biggest miscarriage of justice in history. Well that and Fujitsu staff colluding with post office executives to put innocent people in prison rather than admit there was a problem
[1] https://en.wikipedia.org/wiki/International_Computers_Limite...
https://en.wikipedia.org/wiki/British_Post_Office_scandal