Agreed. Economies of scale are about the only way this makes sense.

I was just working thru the use case for purchasing a 64gb Strix (vs the typical coding I might do using OR or Open Code Go). The ROI is somewhere in the order of 7-10 years for me. So, on a purely dollar cost basis, local inference is $$$. OTOH, from a sovereignty and “own your own code” basis…there’s something to be said for it.

Probably I will run a few tests using OR with a few outstanding tickets of mine, to kick the tyres of Qwen3.8-27B. I’ve tried Qwen3.6-35B and frankly it’s…not very good at code gen or patching.

To boot, code gen isn’t the only thing I use LLM for, so I suppose it depends on how bloody minded I want to be about future use case (I do actually have one).

Ah…that’s just the one-off costs. Remember…they still have to pay monstrous electric bills. Maybe in the order of $100K or more a month.

No dramas. This is probably my 3rd or 4th Lemmy account in last 2yrs; trying dbzer0.com to see what it’s like.

In the spirit of things, I ask GLM5.3 what that set up actually looks like. Ouch.

Fetched that page for real. Specs (from gpuworld’s DGX B200 datasheet page):

Dimensions: 17.5" W × 87.5" D × 42.9" H (cm: 44.5 × 222.3 × 109.0)
Weight: 275 lb (124.7 kg)
Rack height: 10U equivalent
Power: up to 14.3 kW max per system

In real-world terms:

Height: 10U = 10 rack units = 10 × 1.75" = 17.5" tall — about the height of a bar fridge standing on its side. But that's the chassis; you'll also want ~2U of clearance for airflow/management, so plan for ~1.5U of breathing room above it in a rack.
Depth: 87.5cm = just under a metre — so it'll stick out of a standard 1m-deep server rack unless you've got deep racks (or you let the back panel hang open). Standard 42U racks are typically 1.1–1.2m deep to handle this.
Weight: 275 lb / 125kg. That's a hefty man's weight — two people minimum to lift, ideally four or a rack lift. Standard datacentre floor loading is usually fine at this density but worth checking if you're stacking multiple in a rack.

Rack math for 7 units: 70U of just these systems — so a single 42U rack can’t hold them all (you’d need 2 racks), and at ~14kW each you’re looking at ~100kW total draw — that’s serious power/cooling, way beyond a standard office closet.

Short version: think “fridge-sized box, half a tonne with rails, eats 14kW” — not a pizza box.

Cute that they called it $0.3 / 1 M tokens.

GPU WORLD
US$559,950 per unit: https://gpuworld.ai/gpu/nvidia-blackwell-b-series-gpus

So, IRL…about $3.5 million dollary-doos capex. Given they probably have thousands of engineers on staff, yeah, I can see it being worth it.

Why should I give a flying fuck? I have 3 lenovo tiny boxes, a Rpi4 4gb I got pre 2020, a stack of old laptops and phones, a compute stick and 4 tablets. And if I want to tinker (and I do), ESP32 + CYD is hard to beat.

The Pi 5 does nothing for me that I can’t do faster or cheaper (or both) elsewhere. I could literally spend $20 on a WYSE thinclient to get something that’s 3-4x the pi while being comparable in power draw.

No…if I want a shiny, I’ll get a FPGA or a Commodore 77. If I want to thinker, I’ll get ESP32s. And if I want low power always on compute…well, I’m good thanks.

Oui

chop chop

High treason, I’m afraid. You’ll swing for that.

I don’t understand. So, if VLC was downloaded from Tucows or GitHub…who pays? Surely not VLC, because as I read it, this taxes “producers of digital data” but the measurement mechanism is ISPs counting bytes by sender IP.

So if a French user downloads VLC from Tucows, Sourceforge, or Akamai’s CDN, the bytes come from Tucows/Akamai’s IP, not VideoLAN’s. Under a strict sender-IP reading, Tucows pays, not VideoLAN.

Wat?

If in some crazy world this actually goes ahead as written, one of two things happen

1. Everyone on earth geoblocks French IPs
2. Everyone dumps everything onto mirrors and washes their hands of it.

Impossible to say. The token use difference between instant (nigh on useless,) and medium or high (what people anecdotally default to) could be significant.

Eg: if we consider GLM 5.3, the difference in token use between reasoning none and reasoning high is about 3x

https://artificialanalysis.ai/models/comparisons/glm-5-3-low-vs-glm-5-3

But without access to actual token consumption figures, comparing the old and new usage allowances is difficult. Suffice it to say it’s unlikely to be in your favour.

I just don’t get it. Like…who’s buying these things in quantities vast enough to charge that much? Oh, I’ve heard Open AI snapped them and Mac Minis up in bulk for “testing” (whatever that means)… but given the cost now… why / who is buying them in quantities that warrant these sort of prices? It’s not as if they’re rare per se. Hell, there might even be a glut relative to actual demand. Yet, those prices…

Human psychology is weird.

Meanwhile you made me look at prices. JFC.

Oh, your running some JEVs? How are they? I just can’t figure out a use for em.

Yes, concurrency is amazing. I can pull some tricks with CPU (OCR, TinyBERT, Klein etc) while the GPU runs inference…but it’s not a Strix.

Yeah, the BC-250 had / has real promise, esp with the core unlocks…but like P40s and P100s (server cards) people twigged on “this is cheap and fast”. I can’t get one locally for under $600 now. They use to be $150. By the time you fit it out properly, it’s nearly $1K. Not worth it.

The curse and blessing of unified memory!

I was considering getting a BC-250 (a cut down PS5) with 16GB unified memory, but at 400GB/s bandwidth.

When they were cheap, it almost made sense. But realistically, I can get similar speeds with a MoE thanks to cpu and system ram, and 16GB is just not enough for anything top shelf (no, Prism MLs Bonsai models don’t count. I don’t care what the benchmarks say - they’re lobotomised).

Still, given the chance to get a Strix on the cheap, I’d go for it. There’s something very nice about a low power box that can run a 120B model at speed.

A Strix. In 2026. Must be nice :)

It’s a two pronged attack for sure. Make local expensive, make cloud worse.

I think there’s still some wiggle room to be had in the 8 to 16GB VRAM range…so I’m hoping Qwen 4 series brings its A game.

Enjoy and good luck!

For a start, it’s lighter on system resource. By a lot.

It’s something you’ll probably download directly from HuggingFace. That is - people will de-censor models and you just download them.

Of course, you can grab the tools and de-censor your own. It’s a bit of a black art tho; I abliterated a Qwen3.5-2B once (I thought quite well, actually; minimal damage 0.0006)…only to have BADLY ruined its tool calling abilities.

Leave it to the known good quantizers over on HuggingFace.

I thought Google fixed the Gemma template especially for tool calling…or is this ANOTHER regression?