George Hotz | AMD PC Build | tinygrad: building the new tinygrad computer | EPYC 7662 RX 7900 | ROCm
Date of stream 30 Apr 2023.
from $1499 buy https://comma.ai/shop/comma-three
Live-stream chat added as Subtitles/CC – English (Twitch Chat) – three-dot menu icon – Show transcript
Source files:
– https://github.com/geohot/tinygrad
Follow for notifications:
– https://twitch.tv/georgehotz
Support George:
– https://twitch.tv/subs/georgehotz
Programming playlist:
– https://www.youtube.com/playlist?list=PLzFUMGbVxlQs5s-LNAyKgcq5SL28ZLLKC
Chapters:
00:00:00 intro
00:00:48 Alex
00:02:30 light setup
00:04:25 horizontal stream
00:05:15 ROMED8-2T/BCM EPYC 7662
00:05:50 Samsung RDIMM 8GB DDR4-3200
00:06:07 AMD ROCm, goal of tinycorp, bad Nvidia
00:06:35 MLPerf, competitor to Nvidia, AMD developer friendly
00:06:48 RDNA 3 open source,
00:07:15 bitcoin mining case, pcie x16 extenders
00:07:40 motherboard price, cheap gen2 epyc cpus
00:08:18 tinygrad discord, $6000 2019 CPU for $800
00:08:40 $38 8GB RAM sticks
00:09:25 no sponsors, it’s not about money
00:10:30 good mood George
00:10:45 220-T2-1600-X1 PSU
00:11:30 Noctua fans secret 12V fans runs on 5V
00:12:45 no twitch or comma technical support
00:13:00 airflow direction
00:13:30 missed installing the cpu
00:14:40 Noctua fan color
00:15:50 fully modular
00:16:50 overview of the stream
00:18:15 tinycorp headquarters
00:18:30 Alex’s favorite pokemon
00:18:50 pirate flag
00:19:30 work for tinycorp
00:19:55 tinygrad competitive with pytorch on AMD
00:20:05 $5000 computer, 5 more GPUs, multimachine mellanox
00:20:57 advantage by using AMD GPUs, nvidia no P2P support
00:21:20 sponsorship from AMD
00:22:00 feuds
00:22:35 $350 EVGA, Dell PSU $55
00:23:08 rubber feet
00:24:10 Alex’s favorite feud, jack in the box
00:24:55 gucci candle
00:25:20 always kind to service people
00:26:20 haircut, barber
00:28:05 Alex distracted by chat
00:28:40 first computer
00:29:30 Hiball Lemon Lime
00:30:40 good planes 787, Airbus A350, A380
00:31:35 video update, being famous, wealth, india
00:31:15 pc overview, 128 lanes of PCIe Gen 4
00:33:50 bad nvidia, price gouging
00:34:30 AMD P2P
00:35:00 computer on
00:36:10 quiet computer, hotswap 990 Pro 2TB
00:38:25 real facts George’s height
00:38:35 correct M2 slot
00:39:05 desk setup
00:40:28 not on drugs, Alex showing George’s stuff
00:45:00 PC starting up, connecting monitor, keyboard, mouse
00:50:50 finally boot screen on monitor
00:53:50 tarallini snack, looking for usb stick, maple almond butter filled
00:57:15 $100 paypal to George gmail for drug test to take on stream
00:59:10 computer for crypto
01:01:45 will George use HIP on AMD
01:04:05 Ubuntu 20.04, 22.04?
01:06:00 book shelf – Thank you Alex!
01:07:09 bootable Ubuntu 22.04.2 LTS on 870 QVO 8TB, Richard Stallman, GNU
01:14:30 booting from USB
01:18:00 wrong click, installing ubuntu
01:22:00 pc name, no ubuntu pro, meta stock loss
01:25:25 distorting the camera, ai regulation, bios/boot delay
01:31:00 Alex bringing pizza
01:33:00 update-pciids, lspci, uname -a, cpuinfo, meminfo
01:37:30 6 GPU cards, LLM on AMD, 1 pcie for network
01:40:00 ROCm vs Cuda, 7GB LLaMA, dma weights from nvme, data parallel
01:41:10 goal 13GB LLaMA float16
01:41:40 stream from macbook
01:43:35 securing the PC, disable ssh password, 50-cloud-init.conf, sudoers
01:53:30 APEI Generic Hardware Error
01:55:08 amdgpu-pro
01:59:00 will make amd drivers good, install amdgpu, amdgpu rocm
02:05:35 number of devices 0 opencl
02:10:30 PCI-E P2P 3090, researching ROCm
02:14:45 tmux config, rocm with open amdgpu
02:19:10 ROCclr, hsa-runtime64, AMD biggest fan, friends with George
02:31:40 libhsa-runtime-dev, AMD GPU user experience vs nvidia, clang
02:45:15 mi300, custom llvm needed to build rocm
02:48:10 GPT4 rant, openai lame
02:52:00 building clang, big htop, lag
02:57:00 need more ram, AMD uProf
03:05:00 make install, ld.lld doesn’t exist
03:19:45 hsakmt.h error
03:23:00 readme, ROCclr, amd_comgr, rpm package requires rpmbuild, dpkg see files
03:35:10 hsakmt error, X11/Xlib.h error, comgr not installed, ROCm-CompilerSupport
03:54:35 docker container, amd kernel module
04:03:10 amdgpu-dkms, build amdgpu driver module
04:05:22 repo.radeon.com
04:10:40 ubuntu 23.04, updating kernel
04:18:50 it was the kernel issue
04:22:30 add to group ubuntu
04:23:10 tinygrad, llama.py
04:25:30 it did not build, kernel upgrade worked
04:28:25 tinygrad needs only opencl
04:28:40 90% of a problem was the old kernel, chatbots inside llama, 10g net adapter
04:34:25 llama running on amd, optimization needed
Official George Hotz communication channels:
– https://geohot.com
– https://twitter.com/realGeorgeHotz
– https://instagram.com/georgehotz
– https://tinygrad.org
– https://geohot.github.io/blog
– https://twitch.tv/georgehotz
– https://github.com/geohot
– https://youtube.com/geohot
We archive George Hotz and comma.ai videos for fun.
Follow for notifications:
– https://twitter.com/geohotarchive
Thank you for reading and using the SHOW MORE button.
We hope you enjoy watching George’s videos as much as we do.
See you at the next video.