CEO/Chief Architect @theArchality: solved the AI memory wall. 30+yrs systems & silicon. Fmr AMD, 2x founder. BMAC - SRAM/DRAM-less MATMUL.
Canada
Joined December 2024
- Tweets663
- Following58
- Followers183
- Likes436
Liam Coffey retweeted
Published another version of oj-hn.com today. Fixed some performance issues and upgraded all the dependencies. If you read HackerNews in a browser, this really makes the experience a little sweeter, without getting in the way. Free, open source, no tracking or weirdness.
Love and Thoughts to all that are going through this right now. All of Canada is with you.
It is with profound sadness that we confirm that a fire incident at CFB Wainwright on October 17 resulted in fatalities and injuries. Our immediate priority is supporting those affected and notifying next of kin.
canada.ca/en/department-nati…
ALT Text on image: News release
Liam Coffey retweeted
Woah! We just had two full bare metal 8x MI300x machines come back to us. First time in many months since this has been a thing. If you want both of them, you get the full 8x400G as well (reach out to us). Perfect for a training run! Otherwise, 8 hours minimum, ~$217 and you get a real supercomputer.
admin.hotaisle.app
Liam Coffey retweeted
Announcing: dgx-spark-networking
cabled four DGX Sparks into a ring with no switch, taught NCCL which way is clockwise, and shipped an all-reduce that finishes before NCCL has picked an algorithm.
10 KiB all-reduce: NCCL 91 µs. ours 17 µs. one kernel. replays in a CUDA graph. bit-identical on every rank.
the pun is dead (RIP RoCEnante), the horse is now called oneshot, the repo is public and Apache-2.0: github.com/christopherowen/d…
next: letting the GPU ring the NIC doorbell itself. first attempt rebooted a node. we will be back.
Liam Coffey retweeted
Here is a full run of all the steps in a 'hello world' 555 board.
youtube.com/playlist?list=PL…
Will large CSPs start locking up multi-year exclusive supply agreements with NVIDIA?
Will they build datacenters on spec for each new hardware generation, then sell older datacenters as complete units to neoclouds?
When will the build-out scramble become a sustainable process with a predictable pattern?
Why don’t we see more analysis addressing questions like these?
Even 5% of future capacity would not be a small piece of the pie for neoclouds. And do you know who the dark horse is in all of this? @HotAisle.
Ask your favorite LLM to go through all of their posts. You will see a lot of genuine innovation that doesn’t usually get talked about.
Look at what they have done with what they have. It’s only a matter of time before someone with real depth in business understands just how valuable the operating template they have built is.
Someone should conduct a thorough analysis of each neocloud to see who has built a replicable, scalable business. @damnang2 you up for it?
With A100s still going strong at @CoreWeave six years after launch, the Monopoly pieces are changing from hotels to datacenters, indeed.
Liam Coffey retweeted
Yesterday, during an expert call with a Samsung HBM packaging engineer, I asked whether Samsung might have an advantage because of its turnkey strategy.
His answer was
“It is a turnkey strategy, but it kind of feels like Kimbap Cheonguk. You can eat much anything there, but…”
I was honestly impressed by the analogy.
Then I immediately started wondering how on earth I was going to translate “Kimbap Cheonguk” in a way that my international readers could actually understand 😂
Can you imagine if @SpaceX @spacexai or @OpenAI acquired @Cloudflare just how much training data and capabilities they’d get overnight? Unlimited power.
I am surprised @elonmusk hasn’t done this yet just for the advantages of the starlink network alone.
This isn’t the first attempt and won’t be the last.
I worked on silicon-level security to solve for these types of solutions from 2016. We deemed it unviable because it ultimately boiled down to DRM-like capabilities. An actor who found the right vulnerability could violate the contract while appearing compliant, potentially pledging the same compute across multiple exchanges without detection. It was originally part of a range of use cases in confidential computing that a group of us came up with back then.
The harder problem isn’t creating a market for future compute. It’s proving, continuously and across independent markets, that the underlying capacity exists, remains unencumbered, and hasn’t been pledged elsewhere.
We’re already seeing adjacent forms of this problem. A lot of money is being made through nefarious routing arbitrage of models over APIs, en-masse distillation, and secondary markets for call-and-response transcripts.
Do the people that financed this know this already? If so, what could possibly be their justification for this financing decision? It’s not just a problem you can decide to defer to a later round; there’s silicon design needed across all computing.
How do you assure that the underlying capacity exists and is uniquely available to satisfy an obligation?
There were plenty of attempts at GPU compute futures and future-financing structures during the run-up to and peak of Ethereum mining. The fundamental problem was proving uniqueness of the underlying capacity. Proof of work at least provided observable evidence that the hardware existed and was doing work, but that still didn’t establish that the same future capacity hadn’t been sold or pledged elsewhere.
Many of these structures fell apart even before the market declined. As uncertainty increased, particularly around regulation and the ability to liquidate the underlying assets or contracts, financing became increasingly difficult.
A cash-settled hedge can establish a future price. It does not establish that the underlying compute exists, is unencumbered, or is uniquely available to satisfy the obligation.
I have serious doubts about this. It looks like some of the same structural problems are being rediscovered with fresh eyes and fresh money.
Liam Coffey retweeted
Nothing particularly new here, just always satisfying to watch the silk screen carving and solder mask removal.
I am not ok with this @venturetwins if you’ve created this. ‘Gloves off with Bernie’ doesn’t rub out the integrity line you crossed, regardless how this guy bumbles his way through and is being mislead.
Liam Coffey retweeted
Introducing Atlas:
The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D.
Model the world, move the camera, and simulate space & time.
Liam Coffey retweeted
The new Atlas world model is literally black magic. This shouldn't be possible from just 3 cameras 🤯
Liam Coffey retweeted
Business has improved significantly over the last twelve months, and I want to thank every customer who has believed in us.
When I say we’re consistently at capacity, I mean it, and now I have the graph to prove it.
We did this without spending a single dollar on marketing or sales. No parties. No merch. No wasted investor capital. Just products people want to use and a platform they can trust.
What’s interesting is that this chart was built entirely from an AI prompt. It pulls real-time customer assignment data from our custom NetBox plugin and lives inside the internal financial dashboard I’ve been blabbering about.
The speed at which we can build for our business is unreal. When a 10x developer learns how to build with AI-assisted coding, they are now 100x. When you have a foundation to build on, you grow.
No other NeoCloud operates with this level of transparency nor has access to data this detailed. That puts us far ahead of the industry. We have the data, we just prompt to expose it.
Once we deploy AMD MI355X and execute on what we’re planning on building over the next 5-10 years, I believe we’ll be unstoppable, and easily capable of becoming a $20B+ company.
❤️
