iiinova — where your digital self keeps evolving

> Each Mac reads its own shard locally at 1.2 TB/s, while only relatively small intermediate tensors/activations/collectives cross Thunderbolt. some references you might want to peek at: * github.com/ml-explore/m... * github.com/ml-explore/m...

Table comparing M5 Ultra Mac Studio configurations for local AI/LLM workloads. Columns: configuration, cost, physical RAM, raw local memory-bandwidth sum, estimated effective single-model bandwidth equivalent, and independent compute. 1×96 GiB: $6,799, 96 GiB, 1.2 TB/s raw and effective, 1× compute. 1×256 GiB: $10,799, 256 GiB, 1.2 TB/s raw and effective, 1×. 2×96 GiB: $13,598, 192 GiB, 2.4 TB/s raw, ~2.1–2.3 TB/s effective, 2×. Estimated 1×512 GiB: ~$17,199, 512 GiB, 1.2 TB/s raw and effective, 1×. 3×96 GiB: $20,397, 288 GiB, 3.6 TB/s raw, ~2.8–3.1 TB/s effective, 3×. 2×256 GiB: $21,598, 512 GiB, 2.4 TB/s raw, ~2.1–2.3 TB/s effective, 2×. 4×96 GiB: $27,196, 384 GiB, 4.8 TB/s raw, ~3.4–3.7 TB/s effective, 4×.

Comparison table for one 512 GiB M5 Ultra Mac Studio versus two 256 GiB M5 Ultra Mac Studios. The estimated single 512 GiB system costs about $17.2K, while two 256 GiB systems cost $21,598. Both provide 512 GiB total RAM, though the dual-machine setup uses distributed memory. The single Mac has 1.2 TB/s raw aggregate bandwidth and about 1.2 TB/s effective single-model bandwidth, while the two-Mac setup has 2.4 TB/s raw aggregate bandwidth and an estimated ~2.2 TB/s tensor-parallel effective bandwidth equivalent. The dual system also offers about 2× concurrent independent throughput and 2× compute, but requires RDMA and has higher software complexity.

thats a beautiful sight that I would not want to be close to either.. haha

plus, an RTX 6000, or even 5090, or even a couple used 3090s for cheap, STILL require a big PSU, Mobo, mem, CPU.. and will cost you much more in electricity.. and will heat up a room all I need to do to heat half my house is query my 5090 running qwen3.8, and I have a decent water cooling setup

020

you're right.. & CUDA is way more polished than MLX right now.. but M5 Ultra Studio's can compete with RTX6000 tok/watt is better on Macs.. and software is easier these days.. MLX will improve and IMO: nvidia is ripe for competitors.. and I think Apple is better positioned than AMD for consumers

220

Yea i think the m5 ultras are going to be just a hair out of your budget for 256gb, but will be a beast… I just had chatgpt run some simulations on cluster configs on the various macs.. and 2 256gb will be better than the 512 unless they release the 512gb for less than ~$16k

130

A single mac studio is decent and im sure will be getting much faster

Two 3090s do decently.. but depends on how you plan to use the server. Single user? Continuous? Single sessions? Parallelized? Do you have solar and cooling?

210

SgtPooki @sgtpooki.com · 25/08/2026

I built one (not just an llm server though) last year for just a hair over that.. this year, it would be much more difficult.. But i would bet on some used nvidia gpus.

110

of course it was.. I love everything @norman.life does.

030

SgtPooki @sgtpooki.com · 25/08/2026

@ipfs.tech would not be where it is without the folks behind @ipshipyard.com! they had some high quality team-members, I hope folks looking for quality grab them up

070

SgtPooki @sgtpooki.com · 25/08/2026

idk who redesigned @ipfs.tech landing page, but it's beautiful and the anatomy of a CID is really cool. kudos.. The whole page feels modern, exciting, and informative.

141

GitHub - FilOzone/infura-rescue: Download all pinned IPFS data from an Infura account to CAR files before the Aug 15 shutdown, preserving CIDs

The README is written so you can hand it to an agent: set your creds as env vars, paste the prompt from the repo, done. Aug 15 is a hard deadline.. run the rescue today even if you migrate later. github.com/FilOzone/inf...

github.com

Download all pinned IPFS data from an Infura account to CAR files before the Aug 15 shutdown, preserving CIDs - FilOzone/infura-rescue

000

Why @filecoin.io ? This exact scenario. Your data sits with storage providers directly.. there's no storefront whose shutdown takes your data with it. Providers prove onchain that they still hold it, and you add resilience by adding replicas.

100

Once it's on your disk there's no deadline. Want it back online? ipfs2foc migrates the whole set to @filecoin.io: small items packed into ~1 GiB units, stored as 2 copies, ~$5/TB/mo.

100

The output is one CAR file per pin.. the standard @ipfs.tech archive format, byte-identical on re-runs, so your CIDs (and every existing link) survive the move. This is not a download-and-re-add, which could change your CIDs and the URLs folks use to access them.

100

GitHub - FilOzone/infura-rescue: Download all pinned IPFS data from an Infura account to CAR files before the Aug 15 shutdown, preserving CIDs

Infura shuts down its #ipfs service on Aug 15, dedicated gateways included. Anything pinned there that you haven't moved stops being served. One command downloads every pin with CIDs intact: `npx infura-rescue --project-id <ID> --project-secret <SECRET>` Guide: github.com/FilOzone/infura-rescue

github.com

Download all pinned IPFS data from an Infura account to CAR files before the Aug 15 shutdown, preserving CIDs - FilOzone/infura-rescue

110

Reposted by SgtPooki

ProbeLab @probelab.io · 27/07/2026

1/5 New dashboards are live for @filecoin.io Onchain Cloud (FOC). Track approved storage providers, approval criteria, and raw performance — all in one place with detailed description of what's what. Brought to you by ProbeLab! 💪 Check it out at: probelab.io/filecoin/foc/

probelab.io

Filecoin | ProbeLab Analytics

Decentralized storage marketplace with comprehensive network analytics.

142

SgtPooki @sgtpooki.com · 15/07/2026

So jealous there are no cool conferences like this close to me!

011

Reposted by SgtPooki

Daniel Norman @norman.life · 01/07/2026

We shipped content addressing to millions of users, then spent ~10 years pretending the spec already existed. 😅 Well, billions of CIDs later, it's finally real: the official CID spec is live on the @ipfs.tech specs. specs.ipfs.tech/cid/ Thanks @lidel.org and @robin.berjon.com

specs.ipfs.tech

CID (Content IDentifier)

Self-describing content-addressed identifiers for distributed systems

28413

SgtPooki @sgtpooki.com · 18/06/2026

I align with the moat definition above. I’ve also never got VC funding, so take that with a grain of salt. To me a moat is protection of your value prop. If your value prop is quality service with credible exit and user sovereignty… that can still be a moat. Moats are not the bad guy; “they” are.

SgtPooki @sgtpooki.com · 17/06/2026

Do i need to do something differently?

100

SgtPooki @sgtpooki.com · 17/06/2026

Send a DM to whom!?

100

Reposted by SgtPooki

IPFS @ipfs.tech · 17/06/2026

Tomorrow's CID Congress is shaping up to be a good one! Retrieval protocol discussions, a new Matadisco implementation for digital collections data, and a new mobile-first, developer-friendly IPFS node called Meshkit. We have room for 1-2 more demos, send a DM if interested. luma.com/smdplgg8

luma.com

CID Congress · Luma

CID Congress is a series of working sessions towards making CIDs & content addressing easier to use: more deterministic, better compatibility, leaner…

141

Reposted by SgtPooki

Michael D. Green, PhD @michaeldgreen.phd · 28/05/2026

President of MIT not mincing words today in @statnews.com www.statnews.com/2026/05/27/s...

So let me say it as clearly as I can: Without basic scientific research, supported by the kind of farsighted public investment that allows large-scale, undirected, curiosity-driven inquiry, the scientific pipeline will run dry.
In daily life, people may not feel the effects right away, or even in 10 years. But we will feel it. And when someone we love needs therapies that could have emerged but didn't or when other countries now
investing in science can launch new science-based industries or run their societies on vast resources of fusion energy or reap the benefits of quantum computing power or advanced medical breakthroughs, America will wish it sustained its leadership in scientific research here and now.
Sally Kornbluth is the president of MIT.

447188346434

Reposted by SgtPooki

IPFS @ipfs.tech · 20/04/2026

Manifest-declared protocol handlers just landed (behind a flag) in Chrome 146 🎉 This is how ipfs:// or atproto:// can become first-class citizens of the browser! The web that's coming is built on keys, hashes, and verifiable data. Browsers need to meet it there. ipfsfoundation.org/protocol-han...

ipfsfoundation.org

Protocol Handler Registration in Chrome 146 and the Browser as OS

Manifest-declared protocol handlers in Chrome 146? The web that's coming is built on keys, hashes, and verifiable data. Browsers need to meet it there.

14310

SgtPooki @sgtpooki.com · 12/02/2026

Them: "I tried it and it didn't work" Us: "tried what? can you give us more info? we're happy to help but we need more context" Them: "I hate project XYZ, they're unsupportive and things don't work"

000

SgtPooki @sgtpooki.com · 22/01/2026

The blog post is so dense but such a good read

030

SgtPooki @sgtpooki.com · 05/01/2026

Pretty good quick read

IPFS @ipfs.tech · 11/12/2025

Reposted by SgtPooki

We're drafting our 2026 plans to make browsers better for the decentralized web and would love your input. What matters to you? What are we missing? 🔮 discuss.ipfs.tech/t/browsers-s...

discuss.ipfs.tech

Browsers & Standards Work 2026: Call for Community Input

TL;DR: IPFS is planning the next phase of work to make browsers better for the decentralized web – not just IPFS, but any protocol that needs lightweight crypto, peer-to-peer connections, or native br...

0172

SgtPooki @sgtpooki.com · 29/11/2025

This is a massive improvement. Upgrade your Kubo nodes.

060

SgtPooki @sgtpooki.com · 28/11/2025

It should have been a short film. 15min storyline milked and beaten to death with a crappy ending.

020

SgtPooki @sgtpooki.com · 28/11/2025

The movie didnt tell a story, it subjected viewers to a few moments of interesting conflict that it replayed from multiple views with characters it continued to leave more underdeveloped each time, and then it left scared. No name because it failed so bad and I don’t want to feed it.

100

SgtPooki @sgtpooki.com · 21/11/2025

This is super cool daniel :) nice work

SgtPooki @sgtpooki.com · 29/10/2025

This one is worth pondeaning

SgtPooki @sgtpooki.com · 18/10/2025

Ripgrep is a standard tool for me. Its great

000

SgtPooki @sgtpooki.com · 17/10/2025

Have you heard about filecoin-pin yet? docs.filecoin.io/builder-cook... If you are using #ipfs @ipfs.tech @filecoin.global @ipshipyard.com, you should check it out. It's in alpha, but it's pretty exciting stuff!

docs.filecoin.io

Filecoin Pin | Filecoin Docs

Pin IPFS content to Filecoin using familiar IPFS tools and workflows.

020

SgtPooki @sgtpooki.com · 26/09/2025

I love seeing projects like these

SgtPooki @sgtpooki.com · 10/09/2025

Hack incoming. It's rare to get a heads-up!

030

SgtPooki @sgtpooki.com · 09/09/2025

Im still not comfortable enough to post all my derangement on here.. i’ve got walls to talk to