That's the e-ink monopoly for you. Though you could probably use a slightly smaller b&w eink display with a second hand raspberry 4 to shave a lot from that number.

To be fair even if they could build a 10x larger e-ink panel factory it’s uncertain if they could shave costs by more than maybe 40% to 50% per panel.

And it’s very questionable whether there is genuinely 10x more latent annual demand for e-ink panels at that still not cheap price point.

It's great on public transport too.

"It's possible to just connect two computers together with Ethernet" - on the front page of hackernews. Incredible. Ethernet over USB4 interdomain protocol at least had some novelty, but this...

Can you please see https://news.ycombinator.com/item?id=49512349 and, in the future, share some of what you know instead of putting down others (including the community)?

HN has very little understanding of networking in general.

Can you please see https://news.ycombinator.com/item?id=49512349 and, in the future, share some of what you know instead of putting down others (including the community)?

It does seem to be a surprisingly common blindspot of programmers in general. I strongly recommend programmers learn it, it can save a lot of headaches in system architecture.

Any recommendations on a good path to learn this — curious if you have any personal, opinionated thoughts on this.

Buy some cheap computer, pop in a few NICs, setup a FreeBSD or OpenBSD gateway, including DNS and some other goodies.

And by cheap, any $10 relic will do.

You’ll find their manuals very informative.

I'll add looking up the ipv4 and ipv6 specs on www.rfc-editor.org .

Especially since most machines will just WORK with auto-assigned magic weird IPs and you'll actually be able to share files directly via the OS if you do this, no need for Linux netcat tricks at all.

Now I still remember when target disk mode over ethernet was being talked about.

Programmers have very little understanding of computers, in general.

They'll think hard about a beautiful unrealistic idealisation of the problem, and/or treat a software abstraction as if it's the underlying reality, e.g. (of the latter) thinking networking starts with HTTP and languages start with JavaScript.

Can you please see https://news.ycombinator.com/item?id=49512349 and, in the future, share some of what you know instead of putting down others?

Just read the contract from the other side! If focusing on idealistic software abstractions instead of the problem is bad, the implication is that focusing on the problem is good. If treating a software abstraction as the underlying reality is bad, the implication is that looking below it to the actual underlying reality is good.

Agreed, except IME this is "junior web developers" rather "(real) programmers" in general.

Mel wouldn't. But Mel isn't programmers.

HN has very little understanding of many things, especially if you are familiar with <specific thing>.

https://en.wiktionary.org/wiki/Gell-Mann_Amnesia_effect

Just to point out the usb-c on the original x4 is for charging only.

Also firmware updates (on the non-locked variants), no?

What other potential uses are there? Uploading books without WiFi is the only one that comes to mind.

As far as I can tell, firmware updates are only via WiFi.

What sort of pp/tg speed do you get on a Strix Halo?

This is the best I got, all with Unsloth's quantizations.

Laguna-S-2.1:UD-Q4_K_XL (no MTP) pp=186.4 t/s tg=27.8 t/s

Qwen3.6-35B:UD-Q4_K_XL (with MTP) pp=404.4 t/s tg=83.2 t/s

Qwen3.6-27B:UD-Q4_K_XL (recorded pre-MTP) pp=343 t/s tg=12.1 t/s

Laguna actually performed better than I remembered. I thought it was slower.

Have you benchmarked against full precision models for accuracy/ performance?

Not full precision. I've only benchmarked 27B across Q3-6 quants using lm-eval. I lack the hardware to bench 27B at BF16 but I might be able to do Q8_0. I haven't gotten around to doing 35B. I really should upload my collection of results to Github or somewhere.

Here's a summary of what I have for 27B. I used unsloth's UD-Q{3-6}_K_XL quants across 11 evals. The values are pretty linear between Q3 and Q6.

    Qwen3.6-27B     Q3    Q6
    ARC-Challenge   97.0  97.0
    BIG-Bench Hard  57.9  59.3
    GPQA Diamond    77.8  83.3
    GSM8K           92.4  92.6
    Hendrycks Math  35.5  38.9
    HumanEval       80.5  85.4
    HumanEval+      75.0  79.3
    IFEval          87.3  88.0
    MBPP            75.2  77.2
    MBPP+           88.4  88.9
    MMLU-Pro        83.1  83.5

What are pp/tg? I get 30t/s on 27B qwen.

pp is prompt processing how fast it processes the prompt. Tg is token generation how fast, it generates tokens.

okay, so prefill and decode would be the terms I was already familiar with.

Or you could get a $50 smartband and have so much more functionality and still recharge only once a week or two.

App support, stable ABI, uniform UI/UX, hardware vendor cooperation...

Tribalism is human nature, for better or (usually) worse.

As are mass hysteria dynamics.

The little print at the end says "Currently only available in India."

Reminds me a bit of HMD's KaiOS feature phones but this is apparently Android?

> The little print at the end says "Currently only available in India."

I think that was made obvious by the big thing on the marketing website scroll about checking cricket scores.

Because India is the only country that cares about cricket?

They're not even the only country in Asia where cricket's big.

I believe GP’s point is India is the place a company would highlight cricket scores above other marketing lines.

Cricket season is just starting up. I could see this ad in Bangladesh, Sri Lanka, or hell, even Australia. India's not special.

Do you have some examples? I love to explore such things.

No, it uses RTOS Touch. Looks like HMD/Nokia gave up on KaiOS a few years ago.

which makes it a new platform. any other phone uses rtos touch ?

Apparently some other HMD devices (HMD 105, HMD 110) use "RTOS". Most likely their proprietary feature phone OS, don't expect it will provide an option for user-installable apps.

Have you tried with a single CPU to get rid of the NUMA penalty? I understand this likely means halving the memory but I am interested in how much of a difference it makes

I have (192GB machine with two CPUs), pretty much does the trick. It just runs some small models used for embedding, etc. and has those on one CPU / memory node and all the Docker containers on the other one.c

I have a dual xeon also, same as OP: Ivy Bridge + 128GB DRAM, and was never really able to get decent LLM performance out of it. So I ended up biting the bullet and adding a "budget tier" A4000 20GB GPU. Too bad all my DRAM is wasted now--not sure if there is a way to take advantage of lots of DRAM once you move over to having inference happening on the GPU.

Have you tried putting the KV cache on the GPU and running inference from RAM? From what I gather, prompt processing is particularly painful using RAM alone.

I haven't, since the machine has almost no GPU to speak of. It's an old machine, although in theory it would make sense to pop an R9700 in there and load a bunch of models into RAM and swap them in and out.