George Hotz | Researching | multiGPU with HIP (or maybe without HIP) | HSA | HIP Graph | Part 1

Name: George Hotz | Researching | multiGPU with HIP (or maybe without HIP) | HSA | HIP Graph | Part 1
Uploaded: 2024-01-21T23:12:32Z
Channel: george hotz archive
Description: 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 cha...

george hotz archive · Beginner ·🧠 Large Language Models ·2y ago

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 torch…

Watch on YouTube ↗ (saves to browser)

Chapters (40)

intro

0:20 no warning, linkedin ban, child prodigy

2:25 torchrl

3:20 tinybox pre-order, AI computer, lambda labs

5:20 lambdalabs vs tinybox

12:20 7302p vs 7532 epyc

15:00 tinybox raspberry pi for ML

16:10 not buying Apple Vision Pro

16:45 fastdd github, Meta buying H100

18:10 twitch removing content warning, linkedin post

20:15 linkedin worst dating site

20:35 selling rolls royce, money

21:25 George not a good fit for twitter, culture war

23:00 Peter Thiel dinner party, e/acc

25:55 better processor and it got slower

27:35 drive faster than dev 0

30:00 boost frequency, perplexity

33:50 bios, ipmi, epyc boost, not boosting

44:40 btop, pcie 4 vs 5

46:25 direct democracy

47:40 boost speed

50:10 hip graph is not fast

52:20 ROCm 6.0, Llama-2-70b slow, single thread

52:55 single thread, multithread, multiprocess tinygrad

53:55 ggml, tinygrad long term goal, universal

55:00 event, block slow

56:40 GPU queue sync, multiprocess

58:35 writing your own GPU driver, userspace

58:54 AMD HIP, clone of CUDA

1:00:30 finding HIP graph code

1:02:50 spinlocks, multiprocessing, GPU driver

1:03:25 how do GPUs work?

1:06:00 prebuilding the queues, hip semaphores

1:09:20 rdna 3 instruction set

1:12:40 so much complexity, micro engine scheduler

1:14:40 Alex

1:19:55 reading the code to send packets

1:22:45 hate free stream

1:23:30 amd gpu scheduling

1:32:50 perplexity valu

Playlist

Uploads from george hotz archive · george hotz archive · 0 of 60

← Previous Next →

George Hotz | Researching | multiGPU with HIP (or maybe without HIP) | HSA | HIP Graph | Part 1

Chapters (40)

Playlist

Lesson complete!