We make tinygrad; sell tinybox for the GPU middle class. Our mission is to commoditize the petaflop.
San Diego / Hong Kong
Joined June 2023
- Tweets5.4K
- Following212
- Followers85K
- Likes990
For 10-411, should we skip multithreaded CPU and jump straight to GPUs, or are CPU threads a good intro to how GPU distributes global axis across CU/SMs? First 6 weeks are CPU no matter what, but should week 7 be a multithreaded CPU aside?
73%Yes, multithreaded CPU
27%No, GPU a week sooner
271 votes • 15 hours 2
13
4,846
Batch 2 of the 4xMI300X machines is live if you know where to find it. Max 2 per customer. Will ship latest Q2, but probably also in Q1. Refurbishing is going better than expected. We used the Batch 1 money to buy parts for 80 machines.
Ever sorted and binned 1,000 GPUs? These boxes are probably the most compute power you can feasibly get in a 120V home for a while. Batch 2 will be it until after Batch 1 ships.
3
39
4,551
Is this just brutal, or is there a way to simplify it? mixed-radix transform the LLMs call it. Assume you are checking the whole %x36, no tricks using the reduce.
I'm starting to see this as the real bottleneck. Anyone have any ideas? We need to commoditize the means of petaflop production, like making babies but they are computers.
The real thing they are trying to push with the spy hoax narrative is remote tracking and control over chips. Today it's China, tomorrow it's you running a model that's not aligned with them in your garage. What control freak losers!
Three bipartisan bills are key. The AI Overwatch Act will codify the ban on the export of advanced AI chips to China, ensuring that the CCP will never get a hold of the best chips. It also gives Congress the ability to overturn chip export licenses, adding a crucial mechanism to stop unwarranted licenses. The Chip Security Act will give visibility into the physical location of chips and deter smuggling. The Remote Access Security Act will clarify that the U.S. government has the authority to crack down on China’s remote access of chips, preventing overseas AI server farms like Megaspeed.
It may have sounded crazy in 2023, but here we are. And make no mistake, that's exactly the purpose of this psychological operation.
To the massive dismay of those looking to control compute, it's way too late. CPUs and GPUs are everywhere. Models will get smarter, algorithms will get better, and bottlenecks we used to believe in will be overcome. You will lose and the Internet will remember what you did.
your agent can't read 10 million lines and spend 30 minutes rebuilding large c projects, but it can read 26k and wait no time modding pure python. it will optimize *everything* from the model runner to the kernel to the driver, so of course it's the fastest. tinygrad will win.
I know it's not as exciting as the other launch, but we're offering the $50k tinybox red whale with 14xR9700 and 448 GB GPU RAM. This "tall" 2 PSU case is what the 4xMI300X machine comes in, and it's now available in the configurator for custom builds.
AMD really cooked here. Kimi K3.1 Pro on our 8xMI350X (default SGLang recipe) is a whole different experience at 300 tok/s (up from like 60 with K3), the speed makes it often a better choice than Astra/Opus.
223.3 GB/s drive read speed in tinybox red kimi. Custom drivers for the GPUs and NVMe, in tinygrad master. 16 TB of capacity, all at that speed.
This is an old 6x7900XTX tinybox red running Qwen3.8-Flash-Next. We have some of these around; going to offer free tokens at the freegpt party to stress our inference service in a week or two.
Batch 1 of the tinybox red 4xMI300X is sold out! Sign up for the mailing list linked from our site to be notified when batch 2 drops.
We did a ton of work to make this happen, and huge thanks to AMD for supporting us. If you are serious about local inference, you do not want to miss this. tinycorp.myshopify.com/produ…
