George Hotz | Exploring | finding exploits in AMD's GPU firmware | Giving up on AMD for the tinybox

Date of the stream 21 Mar 2024.
from $1050 buy comma.ai/shop/comma-3x & best ADAS system in the world openpilot.comma.ai
Live-stream chat added as Subtitles/CC - English (Twitch Chat) - at the bottom - Show Transcript
Source:
- github.com/geohot/7900xtx
Follow for notifications:
- twitch.tv/georgehotz
Support George:
- twitch.tv/subs/georgehotz
Pre-order tinybox:
- buy.stripe.com/5kAaGL6lk9uX9nW144 (tinygrad.org/)
Chapters:
00:00:00 intro
00:00:40 blog post
00:02:00 open source driver, bugs
00:04:25 pre-orders, cheapETH, George jobs
00:06:30 no one use AMD? kernel, user space
00:08:25 not understanding, business platitudes, unpaid QA team
00:09:40 bugs are not in the driver, kernel driver does not talk to the hardware
00:12:45 GSP, MES
00:13:50 llvm amdgpu launching a kernel, firmware blobs
00:17:00 amd on mlperf
00:18:08 AMDGPU kernel driver change log, will never get better unless
00:19:05 Elon companies
00:19:20 people who succeed in companies do not care
00:21:15 believing wrong things, change, correcting
00:21:55 scaling comma.ai, throwing money at comma.ai, $1 mil per month
00:24:05 fundamentally correct, life wasted on wrong things
00:24:35 kernel dispatch, CP
00:26:08 technically wrong
00:27:49 deadlock, root causing
00:29:10 will never get better
00:30:40 props to AMD, fixing the issues
00:31:20 MES correct description
00:31:50 normal bug
00:32:40 oh god, appreciation for software development, horrifying solution, why is this a flag
00:33:40 drivers have been shit for 25 years
00:34:07 should be an if statement in the allocator not a flag
00:35:20 offer, downside risk, worries about bad PR
00:38:40 sony, suing someone, led by emotions
00:41:10 instinct, Nvidia generations, Apple monopoly
00:45:45 blackwell chip, RDNA 4, Dojo
00:48:50 tenstorrent, mlperf
00:50:45 turing completeness
00:51:00 Jim Keller
00:52:26 7900XTX firmware, Elon can execute
00:53:30 12xIntel Arc, Groq LPU scam demo, Nvidia king
00:55:00 fixing culture, qualcomm
00:57:05 working with George, Elon genius
00:58:44 JH7110, RK3588, qualcomm situation
01:00:25 oneplus folding phone money back
01:02:50 stress, clean ubuntu install
01:04:40 hsakmt debug level, business decisions
01:07:50 dynamic debug how to
01:10:10 hip graph vs cuda graph
01:13:40 loading static libs with ctypes, libdrm
01:20:20 moving on
01:23:55 hsa_queue_create
01:25:20 GPU user spaces very complex
01:27:00 exploits
01:33:50 building tiny boxes, open source timeline
01:37:20 user space crash
01:41:40 ctypes uint8_t pointer from address
01:48:30 random queue
01:55:10 fuzz things
01:55:35 teaching hacking, what is this thing
02:00:40 messages from the kernel, ubuntu rebuilding kernel
02:06:40 heven, are you a good person
02:09:10 unload, load, module,
02:09:25 hacking 50% setting up your env, knowing a lot, binary search by hand
02:18:20 gpu smu mes, phoronix are legit, gfx1100
02:26:05 architecture
02:26:30 kernel driver, minimal all the way to the hardware
02:29:00 RDNA3 firmware architecture
02:32:18 Lana_Lux raid!
02:33:40 recap of the stream, driver not opensource
02:34:30 firmware closed source, signed
02:36:20 talking with AMD, unpaid QA team
02:37:05 work with the product, ordering food
02:39:40 MES support
02:40:20 by monday it's over, hacking is addictive
02:44:20 it's not like qualcomm
02:46:20 what did AMD give, talking
02:52:15 Ghidra, food
03:11:44 verbose
03:24:30 size_bytes, ucode, assembler and disassembler
03:35:45 F32 disassembler, f32 microcode
03:43:10 technical info, Sam Altman
03:50:20 giving up, reverse engineering not sustainable
03:58:40 evil plot, un-paid QA team
04:02:00 break
04:08:50 amd gpu, psp, dcn
04:17:00 interoperability, nvidia for commercial purposes
04:22:50 sas, ta, sos, amd vcn, sdma
04:29:40 secretive amd, imu gc, this is dumb
04:33:40 muted
04:50:00 build one module in linux kernel tree, why reboot
05:37:20 if we corrupt it
05:47:50 amd psp function, amd gpu bootloader
05:53:00 signature check, amdgpu disable dma debug
06:14:25 intel, business development
06:15:25 shill for comma, toyota
06:20:00 security model
06:26:10 tenstorrent, rdna3 emulator
06:34:00 break
06:41:40 7900XTX vs 4090 compute, nvidia firmware, market forces
06:50:30 INVALID_ARGUMENT, docs
07:09:00 ROCm does not work, hard decision, crypto miner
07:15:25 nvidia bugs, docs to drive gpu
07:47:10 message to AMD, amd stock, gpu market share, nvidia, elon musk
08:04:10 boeing, arm ai chip, google, jensen huang, sam altman
Official George Hotz communication channels:
- geohot.com
- realGeorgeHotz
- georgehotz
- tinygrad.org
- geohot.github.io/blog
- github.com/geohot
We archive George Hotz and comma.ai videos for fun.
Follow for notifications:
- 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.

[@geohotarchive]

Read twitter.com/realGeorgeHotz/status/1770238969663594607 | github.com/geohot/7900xtx Bounties for tiny corp / tinygrad -> docs.google.com/spreadsheets/d/1WKHbT-7KOgjEawq5h5Ic1qUWzpfAzuD_J06N1JwOCGs/ kzbin.info/www/bejne/op-5gqaAf6uWmsk Hiring entire stack for tiny corp join if you are interested | kzbin.info/www/bejne/op-5gqaAf6uWmsk work major source of value in your life Pre-order tinybox buy.stripe.com/5kAaGL6lk9uX9nW144 more info on -> tinygrad.org | github.com/tinygrad/tinygrad comma 3X comma.ai/shop/comma-3x | best ADAS system in the world openpilot.comma.ai | from $999 comma.ai/shop/body the future of people Support George by subscribing twitch.tv/subs/georgehotz | Follow George on twitter.com/realGeorgeHotz to be up to date | Read George's geohot.github.io/blog/ Chapters: 00:00:00 intro 00:00:40 blog post 00:02:00 AMD open source driver, bugs 00:04:25 pre-orders, cheapETH, George changes jobs quickly 00:06:30 why no one use AMD? kernel space and user space 00:08:25 not understanding, business platitudes, unpaid QA team 00:09:40 bugs are not in the driver, kernel driver does not talk to the hardware 00:12:45 Nvidia GPU System Processor and AMD MicroEngine Scheduler (MES) 00:13:50 llvm amdgpu launching a kernel, firmware blobs 00:17:00 amd on mlperf 00:18:08 AMDGPU kernel driver change log, amd firmware will never get better unless 00:19:05 companies do not work how Elon companies work, they don't make good things 00:19:20 people who succeed in companies do not care 00:21:15 believing wrong things, change, correcting 00:21:55 better comma.ai, scaling comma.ai models, throwing money at comma.ai, $1 mil per month 00:24:05 fundamentally correct, life wasted on wrong things 00:24:35 kernel dispatch on AMD, AMD Command Processor (CP) 00:26:08 technically wrong 00:27:49 deadlock, root causing 00:29:10 why this firmware will never get better 00:30:40 props to AMD for releasing this, fixing the issues 00:31:20 MES correct description 00:31:50 normal bug 00:32:40 oh god, appreciation for software development, horrifying solution, why is this a flag 00:33:40 this why AMD drivers have been shit for 25 years 00:34:07 should be an if statement in the allocator not a flag 00:35:20 offer to AMD, downside risk, worries about bad PR 00:38:40 sony, suing someone, led by emotions 00:41:10 AMD instinct, Nvidia generations, Apple monopoly 00:45:45 Nvidia blackwell chip, RDNA 4, Dojo 00:48:50 tenstorrent software stack, mlperf 00:50:45 removal of turing completeness 00:51:00 Jim Keller 00:52:26 AMD will not open source 7900XTX firmware, Elon can execute 00:53:30 12xIntel Arc, Groq LPU scam demo, Nvidia king 00:55:00 fixing AMD culture hard 00:55:40 fixing qualcomm firing the whole sales division 00:57:05 don't want to work with George, Elon genius 00:58:44 JH7110, RK3588, comma solved the qualcomm situation 01:00:25 oneplus folding phone money back 01:02:50 AMD situation stressing George out, clean ubuntu install 01:04:40 hsakmt debug level, business decisions 01:07:50 amd dynamic debug how to 01:10:10 hip graph vs cuda graph 01:13:40 loading static libs with ctypes, libdrm 01:20:20 moving on from AMD 01:23:55 hsa_queue_create 01:25:20 GPU user spaces very complex 01:27:00 how many exploits with George find 01:33:50 building tiny boxes, AMD open source timeline 01:37:20 user space crash 01:41:40 ctypes uint8_t pointer from address 01:48:30 random queue 01:55:10 how do you fuzz things 01:55:35 teaching hacking, what is this thing 02:00:40 messages from the kernel, ubuntu rebuilding kernel 02:06:40 heven, are you a good person 02:09:10 unload, load, module, 02:09:25 hacking is 50% setting up your env, knowing a lot, binary search by hand 02:18:20 amd gpu smu mes, phoronix are legit, gfx1100 02:26:05 amd firmware architecture 02:26:30 kernel driver, tinygrad goal minimal all the way to the hardware 02:29:00 RDNA3 firmware architecture 02:32:18 Lana_Lux raid! 02:33:40 recap of the stream, driver not opensource 02:34:30 firmware closed source, signed 02:36:20 mistake was talking with AMD, unpaid QA team 02:37:05 work with the product, ordering food 02:39:40 MES support 02:40:20 if AMD open source anything by monday it's over, hacking is addictive 02:44:20 it's not like qualcomm 02:46:20 what did AMD give to George, talking at AMD 02:52:15 Ghidra, food 03:11:44 verbose 03:24:30 size_bytes, ucode, assembler and disassembler 03:35:45 F32 disassembler, f32 microcode 03:43:10 0 technical info from AMD, Sam Altman 03:50:20 giving up AMD GPUs, reverse engineering not sustainable 03:58:40 Evil AMD plot, un-paid QA team 04:02:00 break 04:08:50 amd gpu, psp, dcn 04:17:00 working on interoperability, nvidia consumer cards for commercial purposes 04:22:50 sas, ta, sos, amd vcn, sdma 04:29:40 secretive amd, imu gc, this is dumb 04:33:40 muted 04:50:00 build one module in linux kernel tree, why reboot 05:37:20 what would happen if we corrupt it 05:47:50 amd psp function, amd gpu bootloader 05:53:00 attacking signature check, amdgpu disable dma debug 06:14:25 talking to intel, no business development people 06:15:25 shill for comma, toyota pd0wm 06:20:00 better understanding of security model 06:26:10 tenstorrent people, rdna3 emulator tinygrad 06:34:00 break 06:41:40 7900XTX vs 4090 compute, nvidia nerfing the firmware, market forces 06:50:30 HSA_STATUS_ERROR_INVALID_ARGUMENT, AMD release docs 07:09:00 ROCm does not work, hard decision, let's fix all of this, crypto miner 07:15:25 nvidia bugs, docs to drive the gpu, amd_hip_bfloat16 07:47:10 message to AMD, amd imu gc, amd stock, gpu market share, nvidia, elon musk 08:04:10 boeing, british company, brazil, arm ai chip, google, jensen huang, sam altman

[@klekaelly]

AI summary of stream: George Hotz explores and critiques AMD's GPU firmware for its complexities and limitations in his KZbin video. He initially tries using AMD GPUs for his AI accelerator chip project, Tiny Grad, but faces numerous software issues and a lack of transparency from AMD. Hotz expresses frustration with AMD's business practices and the GPU's performance in machine learning. After extensive debugging and attempting to collaborate with AMD, he concludes that the firmware's opaque nature and AMD's inadequate response hinder progress. Disappointed by the lack of support and openness in improving the firmware, he decides to switch to Nvidia GPUs for his projects, abandoning AMD. Initial Expectations and Disappointments: - Hotz started with high hopes for using AMD GPUs due to their open-source driver claims. - Discovered that the AMD GPU drivers were not helpful for his needs as they contained opaque firmware blobs which complicated development. Technical Challenges and Debugging: - Hotz discusses the complexities of AMD's firmware, focusing on its layered nature and the opaque command processors. - He found that critical parts of the GPU functionality were handled by closed-source firmware, not the kernel or open-source sections, complicating his attempts to manipulate or optimize performance. AMD's Lack of Support and Transparency: - Hotz expresses significant frustration over AMD's lack of engagement with his project and their reluctance to open-source the firmware or provide detailed documentation. - Criticizes AMD for prioritizing public relations over substantive technical assistance. Business Practices and Market Position: - Compares AMD's approach unfavorably to Nvidia, pointing out that Nvidia, while also not perfect, seemed more straightforward in its business dealings and technical documentation. - Describes AMD's practices as a barrier to innovation and accuses them of not being supportive of small developers and companies. Decision to Abandon AMD: - Concludes that continuing with AMD is not feasible due to the unresolved issues and AMD's unhelpful stance. - Decides to switch to Nvidia GPUs for future projects due to better performance in machine learning tasks and more transparent business practices. Broader Implications and Reflections: - Hotz reflects on the difficulties of starting an AI chip company and the importance of robust software infrastructure. - Discusses the broader tech industry's issues, such as lack of transparency and the challenges small players face when dealing with large corporations.

[@alejandromedina1019]

90% of people who watches these videos are just here to get pumped by George's work ethic

[@GaminHasard]

indeed.

[@sirfranciis]

I don't understand anything but this guy motivates me to keep working

[@SandhyWaer]

INDEED with CAPITAL

[@avwie132]

“I wanna work at X!”, three days later “I don’t want to work there anymore”

[@truelies5431]

Truuuue

[@mikestaub]

I would not give up on AMD, just put it on the back burner. It's not as bad as it looks, they are likely just in reaction mode now against nvidia's enterprise strategy. They will put resources back to the linux code eventually. If you really want to align incentives and expedite these types of bug reports, I would try to get their Venture fund on your cap table. I'm sure they would like the narrative of Tinybox for their marketing team if it ends up working out. The vision of Tinybox is obviously the future. Eventually, AMD and the other players will realize this. Definitely don't fall for the trap of taping out your own chip.

[@biesman5]

Well said, hopefully George reads this comment.

[@andrewboldi47]

This is really good advice

[@mikestaub]

@@andrewboldi47 I started my career working on the AMD GPU driver teams, and I understand the culture there. The people are solid and care, but the company is always reacting to what nvidia, Intel, or qualcomm are doing. He is understandably frustrated, as he is trying to build something cool, in a market they are not prioritizing. The fact that he is struggling is actually alpha. If he can breakthrough the technical problems and form a partnership with key players in the org tinybox will be a serious threat to the incumbents.

[@shaedacode]

What do you mean by taping own chip and why is this a bad idea? Genuine question

[@mikestaub]

@@shaedacode tape out is the last milestone before the production of a new chip. Creating custom silicon is very complex and expensive, even today. I think it would dillute the strategy of Tiny in the medium term to focus on this aspect. Better to put all efforts on the software, distribution, and partnerships.

[@NicholasHadaway]

You are flying so close to the sun. Thank you.

[@hawtsauce4425]

Terry Davis burned his wings.

[@dooby1445]

@@hawtsauce4425The FEDs got to him :/

[@nobodynoone2500]

Lol, glad there are more people in the GPU game now, even if most of them are still missing the mark. Monoculture kills ecosystems.

[@ThisIsAGoodUserNameToo]

Hopefully they learn from failure.

[@melnick1985]

well said.

[@mateuszbugaj799]

I wish I could focus on one task for 8 hours straight

[@narrowbtw3970]

adderall

[@Sami-rn9lt]

Do something you like and add little complexity everyday

[@TheNitroPython]

To be fair, George rarely goes 8hrs without interruption lol

[@aintgonhappen]

You just need some hi ball energy :)

[@aminbusiness3139]

He doesn’t do adderall Don’t do that BS , it can F up your life

[@TheLaraJose]

Watching Emperor George code gives me an insane buff to just get going and do stuff. Thanks!!

[@faceofdead]

i don't understand anything about programming, but watching Geohot's stream is the most productive and exciting part of my day-to-day life. thank you

[@shadfurman]

If you're interested, it's easier than you might think. It feels foreign when you start, and you think you won't get it, but then a piece clicks in place and you unlock a new level "oh I get it." Those experiences never end. There's always another level. The competition isn't as good as you might think either. There are a lot of people who will be way better than you'll ever be, but there are not enough people to meet that demand. That's why software is so buggy, there just aren't enough decent programmers, so most of them are mediocre at best. Just keep working on it. It's incredibly fun. Also, read Clean Code and/or The Fundumentals of Programming, if you're going to do it professionally with other people. Also, programming isn't about the language. It's about learning to break plans and goals down into the simplest step by step functions and implement them efficiently. It's about checking and diagnosing problems, and good communication if you're on a team. The best way to learn is to start doing it, and keep doing it. There's almost no risk, it's not rocket surgery. Break things and have fun.

[@deadbody408]

4:33:39 - 4:50:22 muted with minor exceptions when changing songs

[@lachlanmacnee8704]

nice to listen to george in the background while coding

[@MrFujinko]

Every human is only equipped with one word processor (read Patrick Winston). Either you didn't understand anything from the stream or your code sucks because you were not paying attention to it. Why take the risk?

[@ToothPasteLOL]

@@MrFujinko do you not understand the concept of "in the background?" perhaps they just enjoy some background noise

[@dewijones92]

George. Full respect to you. You fight a good fight. I love you attitude and your devotion to open source. Love your streams

[@SirMo]

7:12:27 I've had a 79000xtx running koboldcpp ROCm fork for the past 6 months, not a single crash (and I'm running a LLM benchmarking utility on it with like 50+ different LLMs). Meanwhile I wrote a watchdog script to restart it for my Titan Xp because it randomly crashes running the same project. ROCm works at least for me. And it works pretty well (better than my CUDA GPU in fact). Granted I'm not pushing anything special. I'm just running .gguf models and it's not as fast as it could be, but it's been rock solid. It's sad, but I think the way you're going about this is pretty counter productive. We need competition in this space, and raking AMD over the coals gets us further away from it.

[@Wobbothe3rd]

Cope. AMD needs to be held to account for its flaws to even he competition. And your comparison is biased, comparing a 2016 Nvidia GPU to a 2023 AMD GPU. The 7900XTX is a waste of money, a glorified 3080.

[@SirMo]

@@Wobbothe3rd 3080 has 12GB of VRAM. It literally can't load most of the models I work with. 7900xtx is twice the GPU.

[@andychow5509]

"And it works pretty well (better than my CUDA GPU in fact)." Yeah... no way. As an owner of the RTX A6000 with 48 GB of RAM, I can guarantee that's not the case. On similar cards, ROCm is about 1/3rd the speed of CUDA on tensorflow/tensorflow-lite models (training and inference). Show me a benchmark with similar cards where ROCm is faster than CUDA.

[@SirMo]

@@andychow5509 my Titan XP crashes, while my 7900xtx doesn't.

[@eelcohoogendoorn8044]

Had the same thing happen to me 10 years ago. I was convinced AMD would put together a decent GPGPU stack as us boomers used to call it then. Nvidia was generating all this buzz and selling compute marketed cards for outrageous prices like 2k. How could AMD not hire a few software engineers and jump on that gravy train, I figured? Just two more weeks, right, before I could run a competitive FFT on an AMD card on my OS of choice? It took me far too long, probably more than a year, to realize it just wasnt going to happen. There just wasnt anyone working at that company, maybe at any level, that had a clue about the importance of software. Every year since, the glaringness of that has roughly doubled. Indeed these release notes illustrate the internal situation well. At this point, the only reasonable course of action is to stage a hostile takeover of AMD, fire every boomer that ever touched a software decision; then hire a new team with zero emotional attachment to their legacy garbage, perhaps take only a tiny bite out of nvidias market cap in being 10 years late to the party, but even so youd make a ridiculous profit on this manouver.

[@polomarco7053]

what is "legacy garbage", if you mind?

[@MrCarburettor]

That is a 10 year old strategy getting applied by all silicon vendors. There are many OS versions that they don't want to deal one by one, so they put all complexity into a dedicated small microcontrollers runs special firmware which are common on all HW and OSs. And they implement dumb wrappers around to communicate to that firmware, they called drivers, specific to OSs. These dumb wrappers can easily be open sourced to give the illusions, as they don't do much anyway. Goes without saying, good luck on getting those Firmware source codes.. GPU scheduler, power firmware, management firmware and this list going to get longer and longer. Intels GUC, HuC, DMC, ME AMDs MES, CP, IFWI and ME Nvidia, and Apple Silicon has very similar firmware's and microcontrollers.

[@jbmw16]

If you put a driver on the kernel you don't have to keep track of all kernel versions individually.

[@jbmw16]

So if you put more functionality on the kernel, less bugs and headaches. Hasn't this been said by kernel devs a lot?

[@MrCarburettor]

well there is a balance of bug free driver and new features you want to add, mostly for perf and power related. Features are your competitive advantage and you dont want to reveal much by adding those into kernel. And majority of new features are coming with a bug. Additionally many of these new hw blocks and firmwares aims to bypass kernel for better performance instead of keep adding ioctls. So...

[@damianbitel8073]

Noooo where are the time stamps :(

[@jaszjsz]

sounds like.. 10% luck 20% skill 15% concentrated power of will 5% pleasure 50% pain And 100% reason to remember binary code.

[@8Dbaybled8D]

at 34 mins, GeoHot reveals the oddity with the flag for 'no-queues'. It sounds as if an object, with options, gets handed for consumption by CP code. This is reminiscent of compiling without a linker or any helper modules where you have to set the flags to run, individually, and the compiler outputs a compressed (and hopefully) optimised object. Looking forward to the next 7:30hrs of this stream to find out

[@rainy5883]

For some reason I prefer to listen to this than music while doing work/gaming lol

[@slimesushi]

thank you george i enjoy your streams

[@brod515]

can anyone tell m where I can learn more about this? I have some experience with , C/C++ related to graphics but where can I learn more about gpu dispatching calls. I simply always assumes this was very low level and internal to manufacturers

[@ithilpalladium]

This isn't my field of expertise but I would start by reading about Vulkan since it's a low level graphics and compute library then read about other GPU compute libraries like CUDA. I think Intels Linux GPU drivers are open source so you can look at that code. Doing what he's doing and reverse engineering the proprietary kernel level drivers is the task of a mad man lol.

[@brod515]

@@ithilpalladium hi thanks. that's I'll see what I can get from the intel drivers. My issue is I'm not sure I'd just understand the code verbatim. like if you look at JPEG implementation you(or I) can't really understand it without first understand the way that implementation works prior to reading the code.

[@45whity]

I'm an average web-developer and don't understand sh1t in drivers neither ML, but watching George is simply inspiring.

[@cezar17negru]

Such a positive guy. Thank you Sir!

[@uzesi1463]

I haven't seen hotz code in a long time. You're my role model

[@dorboi]

You are the GOAT George , never stop doing what you do

[@tiberiumihairezus417]

George I just find out you are the best work companion!

[@commonlogic3646]

No chapters? 😮

[@geohotarchive]

This stream is huge so it takes a while to make them. They will be finished sometime tomorrow.

[@commonlogic3646]

@@geohotarchive i see thank you!

[@XEQUTE]

8 hourss , i didnt see before i read this comment @@commonlogic3646

[@bautistaetcheverry2251]

Does anyone know which color theme George uses for his terminal?

[@harshpatel105]

i would pay to watch real deal engineers code live like this. good work George!

[@RG-si1qz]

Was George playing VR games before this stream?

[@genkidama7385]

oh come on, just go and refactor amd's team. you are born for that

[@justins7796]

never gonna forget clicking 'Make it ra1n' in ole 2010

[@graham8316]

So crazy that amd is missing this opportunity. Does anyone know why?

[@sucim]

Lawyers most probably

[@SR-ti6jj]

Rich boomers that dgaf

[@gungrave10]

Probably because it wasn't a server grade gpu?

[@Yvngsick]

Does anyone know his current dotfiles setup? Especially .bash files

[@nickadams2361]

34:40 'yeah we are gunna have to unlock twitter' 😂

[@ebol08]

He said he spoke with Dr. Lisa Su herself on X. It looks like AMD is gone for now.

[@sjwright2]

4:15:27

[@eskii2]

Bummer, just bought a RX7600. Seems to work with games.

[@Altirix_]

JPR does those dGPU market share reports every quarter, the most recent Q4 23 had 80% 1% 19% for nvidia intel amd.

[@VladSuperKat]

Also amd chips that support 24GB or VRAM also support 48GB, "hint hint". Also if they do MCU they would support 48GB for each chip.

[@ТомашШубарић]

HAVE MY BABIES GEORGE, IVE GOT THE HOTZ FOR YOU

[@MrSenserus]

Can someone break down roughly what's going on here and what he's attempting at different times? I'm a first year CS student so I can understand some but not intuitively comprehend what he's doing.

[@meehai_]

the gpu has a bunch of software abstractizations: user space driver ('regular' app on your PC) that talks directly to the kernel driver (amdgpu) that talks directly to the firmware that operates the gpu hardware to do the actual low level operations (add, multiplies etc.). For this ml framework, he's trying to 'cut the middle man' as much as possible. Initially he started talking to the gpu driver directly (i.e. made his own user level driver) but then realised that in some situations, the low level layers hang (i.e. gpu crashes after 10 hours of ML training). He's trying basically to figure out at which level it crashes and learned that the kernel driver isn't the one 'driving' but actually the firmware which is closed source. The theory is that the bugs making the gpu crash and requiring a reboot are at the firmware level and tries to find ways to trigger it outside of the ML stack. He's doing various modifications in the kernel driver (that is open source) in the place where it talks to the firmware blobs. Around 5h he tinkers with the PSP component of the firmware (the security processor) which is the first 'layer' that loads the rest of the firmware stack. The issue is that since this is all built on top of each other, it's assumed that the hardware and firmware level are tested in all kind of scenarios and are reliable. A user-level set of instructions should never put the hardware down. Just imagine you have access to a server with 8 gpus (or 100 or w/e) and you putting one down takes the entire system down for all the others. The abstractions are built exactly for this, but in this case, the user level ML framework brings down the gpu due to firmware issues.

[@elCoronelCC]

Watched about 30 minutes and I would assume they do not know their stack. Someone (who probably is not working anymore because he got promoted) introduced the layered architecture, now new people need to learn what each layer is for and probably specialize on their layer instead of understanding problems end to end. Could see that leading to adding workarounds to pierce layers (a flag in the API), communication about actual problems getting blocked with platitudes and change logs being written without perfect clarity as no one understands the big picture.

[@tails4e]

I get the frustration with apparent firmware quality based on the release notes, but he said he already has llama working with the stack he has, so what is the main issue blocking him from proceeding as is? Does it absolutely require access to the firmware to make further progress?

[@geohotarchive]

@tails4e firmware crashes twitter.com/realGeorgeHotz/status/1771666351653781862

[@JoeStevens-x7k]

Selling my AMD stock with this revelation. AMD is lucy with the football every time.

[@Deadnature]

I wish George was my friend.

[@LarryTheRoleplayerTM]

That timeline annotation is some wild shit eh? High Functioning coreal lobe neuron link established. uploading derp

[@nokomoko8832]

the movement feels very floaty probably because of the new animations that i didn't like from the start

[@gitgudsec]

i love you george fr

[@iGavid_Doggins]

This stuff is so beyond me but so interesting at the same time...

[@ijustyap]

right i've been watching for hours lmao

[@channel11121]

I really hope the experience with AMD doesn't stop him working with Intel.

[@LokiCDK]

Apparently I need to pay more attention to your youtube :D

[@TimmacTR]

If you want to witness how a 200 iq mind works, take a seat.

[@dakota4714]

Change your smoke alarm battery please!

[@gickygackers]

George is a brotha

[@Walczyk]

youre a hero my man

[@seraphin01]

Good job George! I have both amd and Nvidia but to see amd just being freaking blockheads and keeping their crappy software proprietary instead of opening up like back in the days and having one, if not the, best dev in the world pimping it up to get closer to Nvidia is mind blogging to me.. I hope they see reason or something at some point before it's too late.. Who know maybe you'd have better luck with Intel even.. And if their next Gen gpu delivers the goods it might be a good time to switch blue

[@omprakashnain3508]

I procrastinate a lot, any ideas??

[@geohotarchive]

Just do it!

[@RESPECT-mf6cy]

Unsubscribe it :)

[@mufarodarlingtonkaseke9429]

I feel caught by your comment, I have an exam in a few days😅

[@mateuszbugaj799]

I use an app called Simple Time Tracker to measure how much time I'm spending on work during the day and just try to increase it little by little. It makes me motivated to get back to work when I see how little I actually worked in a given day. It's also important to have a somewhat clear vision on what to do so you could spend a little time each day to prepare a list of things that you should do on the next day. The tasks shouldn't be long so divide them appropriately. I also start to procrastinate when I'm tired so it's important to be well rested, get around 8h of sleep, get like 15 min of rest for each 1h of work.

[@send_love]

You're not procrastinating - you're simply guiding yourself into a different direction. No such thing as "not doing" something. Obviously a part of you feels the need to derail your plans and make the experiences happen that happen in your life. The easiest way out of this situation is to acknowledge what gives you joy, what gives you pleasure and to appreciate more what you actually do as opposed to what you would like to or think should do. Thus appreciation of the life you are having right now is your best bet to then pivot from there onto more. Okay that's one idea. Here's another: Start documenting what you are doing. Make a list with 3 items (never more!) for each day. At the end of each day make room for 5 minutes. In those 5 minutes you make a small drawing of what you did, and what you would like to do tomorrow. Do this drawing by hand. Doesn't have to be beautiful at all. Just wing it. Another idea: Write down your goals in big letters and print them out. Put them up in your room. Formulate goals like this: "My life is great, because I have accomplished ___ and I feel great about it." Even if you haven't reached your goal yet, this will influence you to do it. Another idea: Try self-hypnosis. Search KZbin for 'hypnosis procrastination' - and listen to it in private without distractions. Another idea: "Emotional Freedom Technique" - search KZbin for 'Brad Yates EFT procrastination' - and do whatever he does 1:1 again in private without distraction, drink enough water before and after the exercise. Yes, if you noticed that this technique could work for literally anything - it does - the trick is to do it for about 1h - 2h and your mind will come up with a new paradigm to your situation by doing the tapping exercise. Last idea: Think about all the people in your life that you could enrich by living your best self. Make actual commitments to help other people - weather you know them or not - realize that by having more goals reached, you can enrich the world and people around you. Start making small gifts here and there that will fuel your motivation to do the work. __ OK. Don't do any of this now. Don't do it tomorrow either. In fact, never even think of doing any of this - not now, not tomorrow, not ever.

[@foreignconta]

Lol. I can't finish it today. 😂

[@astroram4613]

time stamps please

[@creatednever9168]

Don’t know what blud’s doing but makes me wanna learn coding if that’s what he’s doing

[@2mj0lk]

MES FW change summary said it all, they opened a can of interns for this shit and there isn't even a senior engineer reading the code apparently because that shit is garbage.

[@gliceriagumuercindodocapit1459]

"i dont know where the cp is"

[@whynot01]

stupid question, but which OS is george using? I assume some Linux version, but why does the GUI look like MacOS?

[@pourya7294]

Because it is macos, the linux is just a terminal, has sshed into his pc

[@whynot01]

@@pourya7294makes sense, thx

[@bosnbruce5837]

Justo build your own RISC-V system

[@JoaquinValerio-nk8he]

I'm in love with him

[@jynx3383]

04:15:27 Bro laughs like a real life super villain😅

[@jinxscript]

here before the timestamps

[@tslomt]

i really would like to know how it feels like or how it is to be so fucking smart.

[@seebradrun]

It's a rare vid today #vodsquad

[@TheNitroPython]

No time stamps? Someone with an OpenAI api key? Lol

[@MARKXHWANG]

Not sure what are you doing, just buy a 4080 super with same price

[@MARKXHWANG]

Lol

[@MARKXHWANG]

It is not a culture problem, it is nobody knows this shit. The guy wrote this are already fired

[@MARKXHWANG]

It is spugatt pilled on top of spuggutty for 20 years

[@nobu1730]

this why nvidia can get away with charging such premium for their cards

[@s1lli]

mommy lisa pls...

[@donovanvanderlinde3478]

We not mad Adopt an anti AMD position And move forward

[@DavidWoodMusic]

Seeing this dude use VSCode and a Mac is wild.

[@danielromeo99]

Why

[@innerbloomset]

Why, do you expected vi and arch?

[@DavidWoodMusic]

@@innerbloomsetno

[@zaphhood4745]

Too long and complex for me, but I bet it's great stuff to people with better attention spans and base skillz.

[@oooogaabooogaa]

geohot therapy sesh

[@stranger_b1]

Your main interest and AMD's are divergent. You want to make computing accessible to everyone and decentralized while they want it to be centralized. Why would they help you on this ? It's surprising that someone that seems that smart believes they could open source their driver to be possible

[@stranger_b1]

Also, reusing old stuff is not compatible with capitalism. Companies want to "invent" new products to sell them, to make profit, to pay investors and employees. This is critical in Capitalist economies cycles. It's like doctors who wanted to use "Chloroquine" to cure Covid. They were blocked by Big Pharma because, they wanted to sell something new and generate money. So I am again surprised how you did not consider all of this in your approach with AMD, this is so candid.

[@gungrave10]

Did Nvidia open source their driver?

[@thevanderion]

Why is he saying at some point that everyone is leaving IDA for ghidra? IDA is so much better to work with..

[@ksk31337]

he is hacking the linux!

[@phoneaccount6907]

I think he might be using vitamin K and opioids for therapy.

[@weich1q2w]

Welcome the 99.99% of viewers whom also don’t understand even half of this

[@parabolic_33]

OUR GOAL IS TO BUILD AI ACCELERATOR CHIPS

[@TravisMilne24]

thinking that elon is the end all decision maker is crazy. he definitely isn't lol. that's why all the dumb shit that he says hardly ever comes to fruition. he's making claims he can't cash, he doesn't want to seem ignorant or not in power, so he says yes or no when presented with the questions publicly. its obvious he then has to go to the doers and coordinate with them to recalibrate the possibilities of the things that he just agreed to. How many broken "promises" has he shared publicly over the last decade?

[@SOLTYGAMINGSTV]

i lodt my mom i dont have a lot in life an i only have a small job streaming video games an stuff an i got a video card that has checkerboard artifacting that extremely fast but goes away when the card isstressed or worked in used in gaming or a workload it doesn't show the but it shows the artifacting only on Chrome mostly in 2D processing if you Google or black and gray checkerboard artifacting flicker you'll see it if i under clock or over clock it dont stop it or any thick idk if this is driver or what win i put it in a other pc an downlode driver 24hrs go by an it pops yp i ran a program called mem test but it not the real mem test its ran on my deask top not in boot abl glash dirve

[@eastsidechump92]

Type better. I hope you're just young or english isn't your first language

[@tylerrose2226]

so you can't touch sony products anymore huh

[@stranger_b1]

you could ask help from amernime

[@carlo-sacchi]

I can't focus for 8 minutes...

[@motbus3]

Is twitter search fixed?

[@snarkyboojum]

Go with Intel - please don't go with Nvidia, even if it means you're not as successful.

[@pets9921]

what makes you choose intel?

[@gungrave10]

Competition.

[@threesixn1ne]

What would you blame yourself over? Over not building your own chip startup at 4 years old? I'm not even a software guy and I see these crutches and cringe.

[@christianjoachimgruber2773]

😂

[@ps3301]

Only buy amd gpu for games. U want to run llm only on Nvidia. That is why Nvidia is worth 2 trillion and amd isn't.

[@alainportant6412]

What is this kid drinking ?

[@DavidFregoli]

not this hairstyle

[@tmaey89]

You are insane man! Could you please work on a Anti-Cheat system which helps detecting DMA cheats for multiplayer-FPS games? (CS2, Valorant, etc.) There is currently waves of Cheater flooding those communities, with even live-streamers showing and explaining their cheats in such games. It is ridiculous! Those communities need help from real experts as the game devs are way out of their league.

[@GG18767gg]

totally waste of time. Devs of that games are not interested to invest in anti cheats, so why Hotz should?

[@deliciouspops]

furrysday stream

[@RFG-np6bl]

George, get colonoscopy.