george hotz archive ：George Hotz | Researching | multiGPU with HIP (or maybe without HIP) | HSA | HIP Graph | Part 1
Date of the stream 20 Jan 2024.
from $1250 buy https://comma.ai/shop/comma-3x & best ADAS system in the world https://openpilot.comma.ai
Live-stream chat added as Subtitles/CC – English (Twitch Chat) – at the bottom – Show Transcript
Sources:
– https://hsafoundation.com/wp-content/uploads/2021/02/HSA-PRM-1.2.pdf
Follow for notifications:
– https://twitch.tv/georgehotz
Support George:
– https://twitch.tv/subs/georgehotz
Pre-order tinybox:
– https://buy.stripe.com/5kAaGL6lk9uX9nW144 (https://tinygrad.org/)
Chapters:
00:00:00 intro
00:00:20 no warning, linkedin ban, child prodigy
00:02:25 torchrl
00:03:20 tinybox pre-order, AI computer, lambda labs
00:05:20 lambdalabs vs tinybox
00:12:20 7302p vs 7532 epyc
00:15:00 tinybox raspberry pi for ML
00:16:10 not buying Apple Vision Pro
00:16:45 fastdd github, Meta buying H100
00:18:10 twitch removing content warning, linkedin post
00:20:15 linkedin worst dating site
00:20:35 selling rolls royce, money
00:21:25 George not a good fit for twitter, culture war
00:23:00 Peter Thiel dinner party, e/acc
00:25:55 better processor and it got slower
00:27:35 drive faster than dev 0
00:30:00 boost frequency, perplexity
00:33:50 bios, ipmi, epyc boost, not boosting
00:44:40 btop, pcie 4 vs 5
00:46:25 direct democracy
00:47:40 boost speed
00:50:10 hip graph is not fast
00:52:20 ROCm 6.0, Llama-2-70b slow, single thread
00:52:55 single thread, multithread, multiprocess tinygrad
00:53:55 ggml, tinygrad long term goal, universal
00:55:00 event, block slow
00:56:40 GPU queue sync, multiprocess
00:58:35 writing your own GPU driver, userspace
00:58:54 AMD HIP, clone of CUDA
01:00:30 finding HIP graph code
01:02:50 spinlocks, multiprocessing, GPU driver
01:03:25 how do GPUs work?
01:06:00 prebuilding the queues, hip semaphores
01:09:20 rdna 3 instruction set
01:12:40 so much complexity, micro engine scheduler
01:14:40 Alex
01:19:55 reading the code to send packets
01:22:45 hate free stream
01:23:30 amd gpu scheduling
01:32:50 perplexity valuation, how to value a company
01:34:00 HSA queue
01:35:30 perplexity fast, GPT4 slow, anthropic
01:39:00 HSA level 0
01:41:30 HSA runtime book, anna’s archive
01:45:30 no copyright infringement intended
01:46:20 AQL packets
01:52:55 piano
01:53:20 AMD is for people who likes to get twice as much GPUs for their money
01:54:50 tinygrad pay per token API
01:58:30 replacing HIP support with HSA support
02:00:20 Nvdia vs AMD datacenter, customer GPU architecture
02:02:30 secret good version of openpilot joke
02:03:40 HIP does not use DMA engine
02:05:00 bit blit
02:14:20 rocm-bandwidth-test
02:17:00 hca kmt api amd
02:17:55 Alex
02:20:08 the hidden song
02:23:30 going on a journey
02:27:55 real completion events
02:31:30 hsa example of kernel dispatch
02:32:00 cool that AMD is so open
02:35:40 just using HSA, HSA rabbit hole, hsa foundation
02:37:50 the chapel language with 0 github stars
02:39:40 HSA Programmer’s Reference Manual
02:40:20 linkedin post
02:42:40 the weather people, if you could design a country, deep state
02:45:45 conservatism, progressivism quote
02:49:00 Alex
02:50:00 thinking from first principles, experiments hard
02:50:20 has anyone heard about HSA foundation
02:52:50 scientific computing people, OpenMP, OpenACC
02:55:20 AMD extensions
02:56:15 traveling salesman, 2^n algorithm, scientific computing funding
02:59:10 leslie greengard
02:59:40 deep learning revolution
03:01:00 tinygrad experiment, complexity dysfunction of governance
03:01:38 misunderstanding of how software is developed today
03:02:00 compression is intelligence
03:02:20 complexity management instead complexity reduction
03:02:40 spacex rocket landing controls genius
03:05:10 complex systems, twitter
03:06:10 software 0 cost to replication
03:07:40 twitter acquisition best political dollar ever spend
03:09:30 making the tinybox good
03:09:46 making money off OSS
03:10:10 pre-order tinyboxes
03:11:10 etched.com, tenstorrent.com
03:13:50 tenstorrent offering a card to George
03:14:15 respect to tenstorrent, intel tier
03:15:05 extropic.ai
03:16:40 science grants, fundamental research that needs to be done
03:17:40 bullish on perplexity
03:18:40 atomicsemi.com
03:19:30 ranking startups, tenstorrent open source
03:21:00 tinygrad factorization
03:23:30 hammer.lol, berkshirehathaway.com
03:24:40 stop using javascript
03:27:20 apple.com website, feross.org
03:30:10 lana_lux 5k viewers
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://github.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.