George Hotz | Programming | multiGPU with HIP (or maybe without HIP) | HSA_DISABLE_CACHE=1 | Part 2

Name: George Hotz | Programming | multiGPU with HIP (or maybe without HIP) | HSA_DISABLE_CACHE=1 | Part 2
Uploaded: 2024-01-22T20:55:17Z
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 · Intermediate ·📰 AI News & Updates ·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://rocm.docs.amd.com/projects/HIP/en/latest/user_guide/hip_porting_guide.html#threadfence-system 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 00:02:40 HIP P2P copy slow 00:…

Watch on YouTube ↗ (saves to browser)

Chapters (46)

intro

2:40 HIP P2P copy slow

15:20 hipDeviceEnablePeerAccess fast

17:30 synchronization mechanism

21:50 dumb questions = banned

23:20 no idea if this is working

25:40 hip c++ 11 atomics

29:10 cuda atomic sync

31:20 global cache flush hip

34:20 glc dlc

42:00 no idea if access is correct

43:00 hip atomic and system example

48:10 deep dive in spinlocks

49:00 hip cache flush

52:10 global_atomic_and_b32

56:30 hipDeviceSetCacheConfig

58:10 threadfence system

1:03:00 block

1:04:40 D0, D1

1:07:00 events are slow because syscalls

1:10:40 answering smart questions

1:11:25 bad athomics understanding

1:11:45 x86 vs arm difference

1:12:20 HIP event API slow

1:12:50 tinybox vs 8xA100

1:13:40 radeon pro benefit 48GB RAM

1:14:20 H100 price, MI300

1:15:00 buy gaming PC in $1000 range

1:16:25 6x7900 XTX

1:16:45 nvidia software support

1:17:25 tinybox power efficiency

1:18:20 lamma2 70b unquantized on tinybox

1:19:00 tinybox V2 same case, GPU refresh

1:20:00 6 tinyboxes a day factory

1:20:35 nvidia blackwell release timeline

1:26:10 never shipping anything in your life

1:32:30 removing lines

1:40:30 test_multitensor

1:48:20 deeper problems, good abstractions

1:50:55 does not work and it's slow

1:55:00 Hello nobody is talking

2:01:30 event 4

2:03:15 Lana_Lux raid!

2:03:50 introduction to new people, tinygrad

2:04:15 beautiful_mnist.py

2:05:30 tinygrad ope

Playlist

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

← Previous Next →

George Hotz | Programming | multiGPU with HIP (or maybe without HIP) | HSA_DISABLE_CACHE=1 | Part 2

Chapters (46)

Playlist

Lesson complete!