We make tinygrad; sell tinybox for the GPU middle class. Our mission is to commoditize the petaflop.
Announcement at 2PM PST today jointly with comma. Everyone have their wallets ready? I think it's the nicest GPU dock ever made. hardware specs: * ASM2464PD based * USB2/3/4 + PCIe 4.0 x4 * fully
I'm sure you've heard about the AI arms race in hiring. Automated resume spammers talking to automated screening systems and everyone having their time wasted. For a while, we had a bounty program
Once tinygrad is 1.0 we will begin to roll out our enterprise hardware strategy. Are you sick of managing hundreds of machines to train a model? Isn't what you really want one huge GPU? Imagine a
The USB4 Mac GPU stuff is ready for you to try. You have to disable SIP and install the DEXT in "extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension.xcodeproj" then tinygrad can see your AMD or NVIDIA
We've been moving away from the ShapeTracker and the legacy scheduler to one that scales to flash attention and beyond. It's called RANGEIFY. To understand RANGEIFY=1, a few prereq concepts help. A
These are amazing machines hardware wise. 2x EPYC Turin 9575F processors, 3 TB of DDR5 RAM, 2.3 TB of GPU RAM, and a whopping 12 power cables. People take it for granted, but AMD makes these CPUs too
As summarized from all the commits. Performance & Optimization In 2024, tinygrad made significant strides in performance. A major shift from LazyBuffer to a UOp system led to an overhauled scheduler.