9 ms·

So for reference, this geohot is George Hotz who has a company Tiny Corp [0] in the space. Among a long list of things, he had a moment in the spotlight recentl

by roenxi 3y ago

So for reference, this geohot is George Hotz who has a company Tiny Corp [0] in the space. Among a long list of things, he had a moment in the spotlight recently for a long rant [1] where he "gave up on AMD" because their drivers sucked more than he expected. The situation is reasonably complex - AMD have some programmers working on ROCm who seemed to be operating at standard fare when that is really not what AMD needs right now (I personally, suspect there is/was a PM in a key position who didn't "get it", although I am not sure what it is either). As far as I know it got the attention of Lisa Su. I'm cheering him on, even though his complaints were a little melodramatic. My experience is the driver technically supports everything I could possibly want. The problem is if I spend an evening trying to do anything with OpenCL or ROCm the kernel hard-locks and I go to bed early. If the problem inside is what it looks like from the outside (repeating myself, a key manager somewhere just doesn't get the space) they really need some pressure from grumpy customers like George to realign their software development process. That context might be related to this particular case. I see a suspicious folder called "crash" in this repo. [0] https://tinygrad.org/ https://tinygrad.org/ [1] https://www.youtube.com/watch?v=Mr0rWJhv9jU https://www.youtube.com/watch?v=Mr0rWJhv9jU - as I recall he ran the demo suite and it crashed.

mobilio 3y ago
And follow-up of that video: https://geohot.github.io/blog/jekyll/update/2023/06/07/a-dive-into-amds-drivers.html https://geohot.github.io/blog/jekyll/update/2023/06/07/a-div...
echelon 3y ago
Anyone working on getting us out of the Nvidia monopoly on commerical AI should be praised. We need lots of competition in this space. I'm hoping Google spins off its TPU division and sells chips to everyone that wants them. I'm hoping AMD gets their act together and that they open up their drivers and stack. I'm hoping a lot of other hardware companies and compute stacks start making inroads on this. Nvidia is not just bilking the whole world on margin, but they're also a limiting reagent in the grand scheme of progress.
- sorenjan 3y ago
  You'll note that AMD (and Nvidia, but that's expected) is unfortunately missing from the members list for UXL, which seems to be based on Intel's oneAPI which in turn is based in part on Khrono's SYCL: > Our mission: > • Build a multi-architecture multi-vendor software ecosystem for all accelerators. > • Unify the heterogeneous compute ecosystem around open standards. > • Build on and expand open-source projects for accelerated computing. https://uxlfoundation.org/ https://uxlfoundation.org/ If AMD got their act together they would go all in on open standards together with other actors in the segment, but maybe they really want to see if their rocm can repeat Cuda's success and lock in? I just want accelerators to become commodities and interchangeable like CPUs, let me write my code once and deploy on whatever hardware the user have, preferably without having to compile multiple versions.
- adolph 3y ago
  > hoping Google spins off its TPU division There’s a lot of stuff Google could have spun off but they seem content to ignore things until they turn out the lights.
- ein0p 3y ago
  [flagged]
  - wmf 3y ago
    Making GPUs work properly requires years of effort so she needs a time machine to go back and start earlier.
    - ein0p 3y ago
      [flagged]
      - qotgalaxy 3y ago
        [dead]
      - threeseed 3y ago
renewiltord 3y ago
I thought he sort of gave up and is going to build his stuff on Nvidia in the end? It's wild. I'd want AMD to present some sort of alternative if only because you can't get any high-mem chips these days but they seem to have some massive organizational uselessness on this front.
- mindcrime 3y ago
  They changed their mind a few days later and announced they're going to do NVIDIA and AMD versions of their box.
  - renewiltord 3y ago
    Interesting news. Thank you for that update.
causality0 3y ago
AMD GPUs are fantastic values for money when it comes to running mainstream PC games. Once you step outside that narrow box they become infuriating. My 7900XTX is probably the last AMD GPU I ever buy. It's just horrible at all the other things I want to use it for, especially and gallingly, game console emulation. For the first few months I had it I actually kept my old GTX1070 plugged into my motherboard because the eight year old $400 GPU was so much better at half my use cases than the $900 AMD card.
MuffinFlavored 3y ago
> The problem is if I spend an evening trying to do anything with OpenCL or ROCm the kernel hard-locks and I go to bed early. In a race against NVIDIA where every comment on here is about just how much better CUDA is than any alternative, why doesn't literally what you just said "also" have the attention of Lisa Su?
- georgehotz 3y ago
  I have had multiple replies from Lisa Su on both Twitter and by e-mail. It doesn't help. AMD is structurally incapable of fixing these issues. They don't even have a 7900XTX in CI. Crashing the firmware is so trivial there's no way anyone fuzzed anything. They debug by the application, adding mitigations at many layers of the stack to make it work. While this strategy is fine for the 20 mainstream games that come out per year, unless you root cause issues and have a good CI, you'll never build a stable GPU for general compute.
  - MuffinFlavored 3y ago
    Genuine question for you then, why are you investing your time reversing their firmware if it sounds like you think they're failing/will fail? Why not just "adopt" the golden child/golden standard NVIDIA now for your usecase, reverse that (if needed) and move on? Genuinely curious what you gain from trying to make AMD work/investing in AMD on your end here. Is it that you see a value proposition through the "muck" of crashy-kernels that I'm missing?
    - georgehotz 3y ago
      This is my weekend hobby, I'm learning about GPUs. Besides, I'd like to see competition to NVIDIA. From a business side, tiny corp is selling boxes with both AMD and NVIDIA, we'll let the user choose how much they value money vs pain.
      - rudedogg 3y ago
        What are your thoughts on using Intel GPUs for deep learning?
ultra_nick 3y ago
NVidia is making boatloads of money because their driver works and they have a software library called CUDA that accelerates neural networks. Nobody expects AMD to match them, but George thought they could at least write a GPU driver. If AMD can get that driver out, then George could provide a competitor to CUDA (for neutral networks only). They'd both make boatloads of money. However, AMD was less capable than expected and their drivers were too buggy to run neural networks like those needed for the MLPerf benchmark. So now, it appears that AMD, Tinybox, and investors like me won't be making boatloads of money.
- DiabloD3 3y ago
  Slight mistake in your description: CUDA is an out of date API that was replaced by Khronos's own official compute APIs. Khronos is a standards consortium that Nvidia is a founding member of. Although the marketing department at Nvidia still pushes for greenfield CUDA codebases, no new code should be written in it, and they should opt for open source international standards only. Khronos APIs are implemented by over 120 vendors.
  - greenavocado 3y ago
    Show me the announcement deprecating CUDA
  - mshockwave 3y ago
    With my past experience in Khronos, NVIDIA is indeed a member and they sent decent guys to the meetings -- but only for strategic reasons, rather than "advocating for open standards" as you described. My experiences there actually told me the opposite that they will never drop CUDA. Objectively they also have incentive to do so: fighting with 100-ish companies to ratify something is always slower than rolling out an feature in an ecosystem you have total control of.
    - pixelpoet 3y ago
      Not just that, but they get to abuse their near monopoly to strongarm many companies into using Cuda. Can't say more unfortunately...
NicoJuicy 3y ago
That's not really on how I see it. He's trying to do everything on AMD's consumer based GPU's, while he's comparing it with NVidia's enterprise GPU's with the same expectations. That's just not how it works in real life. But if he succeeds, yeah, he could save tons of money and his hardware for ML could actually be completed.
- modeless 3y ago
  This is wrong for two reasons. One, he claims that AMD's enterprise GPUs are also unreliable. Two, he is comparing vs. 4090 which is a consumer GPU and CUDA is very reliable as it is on all of Nvidia's consumer hardware.
- dannyw 3y ago
  NVIDIA’s consumer GPUs are highly reliable and support the same CUDA capabilities as the enterprise products. This is something NVIDIA gets right. Meanwhile, ROCm isn’t even supported on certain lower end AMD consumer chips.
  - llm_trw 3y ago
    Rocm is only supported on the last generation of flagship graphics cards. Which is a shame because the Radeon VII is still better than anything else when it comes to 64bit float computation.
    - mnau 3y ago
      "Supported" I watched geohot streams and if he can go for an hour with a need to restart the computer due to crashed kernel driver/locked up card, it's a miracle (various computers, one or multi-card setup). And he is not trying to crash the driver, just making the most basic stuff work. I seriously think there is a deep HW bug that caused the difference in performance between RDNA3 presentation and reality and they had to do some horrible hacks to make it work. Either that or the card never seen a tester for compute. Which is strange, considering they added it as a "supported" option.
trws 3y ago
They do need to rework their development processes. It’s worth noting though that there are two (kinda three) different drivers with very different quality and different teams behind them. There’s the binary driver, on windows and the fglrx driver for Linux, that contains the full OpenGL stack and all the proprietary stuff. That’s been maintained and extended since the ATI days by a team primarily based in Markham. It’s garbage. It’s been garbage since the 90s. This is not new or surprising. It’s unstable, poorly maintained and just generally a bad time. It was also the reason OpenCL required an X session run as root with no access control and sometimes a stub dvi cable on the card for many years. It is a source of endless sadness and despair. The other driver is the HSA/ROCM driver on Linux, which can be used with the amdgpu upstream kernel driver and is itself upstream. The amdgpu driver is 2d only, but has been developed in the open, and stable, for a long, long time. It’s maintained by a completely separate set of teams mostly in (last I knew) Germany and Austin. The HSA driver has likewise been immensely more stable than the binary drivers since it was introduced. There are potential problems with it, but I’ve managed to recover every single one I’ve run into without a reboot, and I’ve been doing GPU compute with AMD, NVIDIA and intel since 2008. When it was the binary driver, we needed reboots every time an AMD GPU locked up. I’m not saying there’s no problem here, but he’s reporting problems I’ve never seen in thousands of hours of high-load compute both personally and for work with the driver one ought to use for ROCM. There are other problems, but HSA/amdgpu drivers have not been high on that list for me. Possibly because I was so used to how bad the other ones were I guess…
- rstat1 3y ago
  I believe your info is a bit out-of-date. FGLRX hasn't been maintained or actively developed for a quite a while and unless you're using super-duper-ancient HW no one actually uses it anymore. AMDGPU is the the current Linux driver, combined with Mesa, and certainly is not 2D only.
  - pezezin 3y ago
    Yup, FGLRX was deprecated from all major distros in 2016. Nobody has used it for ages.
  - 3y ago
andrewmcwatters 3y ago
It was particularly hard to listen to himself admit that he basically became unpaid QA for AMD. Like, dude, you’re making the same junior mistake people writing apps for the App Store make. It’s not your platform.
- llm_trw 3y ago
  If the AMD code makes it into the kernel, like their graphics drivers did, then we've won because it will just work on every kernel update. Compared to now where AMD drivers just don't work and NVidia drivers only work with the right blobs installed I'd be willing to take a paycut for 6 months to have 10 years of painless development. That said I jumped ship to NVidia since I need to work and not spend all my time trying to get servers to the point that they can work.
- threeseed 3y ago
  As a developer you're always building on top of something you don't control. At least with the App Store you make pretty good money. It's worse when you have companies who have APIs you need that you know they don't care about.
- modeless 3y ago
  His thesis is that owning a good software platform is a prerequisite for a successful AI hardware company. He is (or was) using AMD as a stepping stone to create and flesh out that software platform (tinygrad) before starting custom ASIC development, instead of going the other way around with a costly ASIC first and trying to develop a software platform after, or trying to push support for an ASIC into someone else's platform such as PyTorch. Not sure I agree but it's an interesting idea. It's also worth noting that tinygrad is bypassing as many layers of AMD's software platform as they can get away with (MIOpen, ROCm, and even the userspace GPU driver).
  - andrewmcwatters 3y ago
    Well, I think he's probably right, it's just a shame that he's going through this difficulty. I hope he ends up succeeding in his goals.
lannisterstark 3y ago
>I'm cheering him on, even though his complaints were a little melodramatic. I've really soured on him when he literally went "I'm going to fix twitter search, hire me elon" and then went to public saying "Hey, want an internship? Fix it, and then MIT license your code(so he could use it lol), and maybe we can talk - oh btw I have no real authority to give anyone an internship." He might be a brilliant programmer and problem solver, but that incident left a very sour "Grifter" vibes.
- quantumwoke 3y ago
  I soured on him a bit more when I watched some of his stream yesterday and he was ranting about diversity being a root of evil in Silicon Valley. Struck me that he hasn’t matured much since he was slinging mud at Sony and Apple.
  - UberFly 3y ago
    He seems to be the type who doesn't care who's "soured on him".
- hughesjj 3y ago
  Yeah, I was feeling similar vibes back when Lex Friedman interviewed him for some self driving stuff back in the day (...I've soured on lex too since he's not really an interviewer/journalist so much as a host/promoter and rarely gets to anything interesting in his talks imo). His work for jailbreaking/actually being able to own the devices we pay for back in the iPhone+PS3 days was and is still inspiring, but I think being so smart and encountering so much bullshit in his life kind of led him to the 'rockstar coder I know better than you regardless of domain' personality, and that leaves him both a bit grating and underquipped to drive consensus. He's still a great wrench to throw into the bullshit machine though, and the guy is smarter and more accomplished with computers than I'm ever likely to be.
  - infogulch 3y ago
    This is a great take. We need more people that are willing and able to jump in and wreck new bullshit machines as they crop up in diverse industries. They can even keep their boastful grandstanding attitude as compensation.
benreesman 3y ago
I’m going remind everyone that AMD and NVIDIA are atypically competition avoidant, but I don’t want to defend a generalization that broadcast let’s be specific: Hopper and MI300 are on paper just straight up competitive, both super scarce, high-margin cards that really love PyTorch, (which really loves NVIDIA, even TPU is, not the fist among equals). But all the MI300 is going to supercomputer and adjacent things, and all the Hopper is going to giant tech LLM type stuff, it’s not the same people bidding on those bins of those. Oh, and their respective CEOs are closely blood-related and in a trivially first name if not family gathering basis. The people root for George generally think this isn’t real capitalism. We think it’s a trend towards the failure to enforce anti-trust laws.
- dotnet00 3y ago
  >I’m going remind everyone that AMD and NVIDIA are atypically competition avoidant, but I don’t want to defend a generalization that broadcast let’s be specific: Hopper and MI300 are on paper just straight up competitive, both super scarce, high-margin cards that really love PyTorch, (which really loves NVIDIA, even TPU is, not the fist among equals). > But all the MI300 is going to supercomputer and adjacent things, and all the Hopper is going to giant tech LLM type stuff, it’s not the same people bidding on those bins of those. They clearly compete fiercely, every feature NVIDIA adds to its consumer GPUs gets an equivalent version by AMD. Previously AMD has introduced features that have caught on enough to force NVIDIA to implement them. AMD's offerings often force NVIDIA to discontinue certain products or to drop prices. The reason MI300 is going to supercomputing and Hopper to AI is the self fulfilling prophecy that AMD's software still sucks for AI, thus they have to target general supercomputing. >Oh, and their respective CEOs are closely blood-related and in a trivially first name if not family gathering basis. This is mostly just a meme, they're distant relatives, which is not too surprising considering they're both from Taiwan. Jensen's maternal uncle is Lisa's grandfather, with said maternal uncle being the eldest of 12 siblings, 18 years older than Jensen's mother. I barely know the names of the eldest of my aunts, who have a similar age gap, let alone have a close relationship with them, because that age gap and number of siblings means that even my mother barely knows them.
  - 3y ago
braiamp 3y ago
Common confusion. It isn't the kernel driver that has problems. Not their userland "drivers". It's their firmware inside the GPU's.

by roenxi 3y ago

mobilio 3y ago

And follow-up of that video: https://geohot.github.io/blog/jekyll/update/2023/06/07/a-dive-into-amds-drivers.html https://geohot.github.io/blog/jekyll/update/2023/06/07/a-div...

echelon 3y ago

Anyone working on getting us out of the Nvidia monopoly on commerical AI should be praised. We need lots of competition in this space. I'm hoping Google spins off its TPU division and sells chips to everyone that wants them. I'm hoping AMD gets their act together and that they open up their drivers and stack. I'm hoping a lot of other hardware companies and compute stacks start making inroads on this. Nvidia is not just bilking the whole world on margin, but they're also a limiting reagent in the grand scheme of progress.

sorenjan 3y ago
You'll note that AMD (and Nvidia, but that's expected) is unfortunately missing from the members list for UXL, which seems to be based on Intel's oneAPI which in turn is based in part on Khrono's SYCL: > Our mission: > • Build a multi-architecture multi-vendor software ecosystem for all accelerators. > • Unify the heterogeneous compute ecosystem around open standards. > • Build on and expand open-source projects for accelerated computing. https://uxlfoundation.org/ https://uxlfoundation.org/ If AMD got their act together they would go all in on open standards together with other actors in the segment, but maybe they really want to see if their rocm can repeat Cuda's success and lock in? I just want accelerators to become commodities and interchangeable like CPUs, let me write my code once and deploy on whatever hardware the user have, preferably without having to compile multiple versions.
adolph 3y ago
> hoping Google spins off its TPU division There’s a lot of stuff Google could have spun off but they seem content to ignore things until they turn out the lights.
ein0p 3y ago
[flagged]
- wmf 3y ago
  Making GPUs work properly requires years of effort so she needs a time machine to go back and start earlier.
  - ein0p 3y ago
    [flagged]
    - qotgalaxy 3y ago
      [dead]
    - threeseed 3y ago

renewiltord 3y ago

I thought he sort of gave up and is going to build his stuff on Nvidia in the end? It's wild. I'd want AMD to present some sort of alternative if only because you can't get any high-mem chips these days but they seem to have some massive organizational uselessness on this front.

mindcrime 3y ago
They changed their mind a few days later and announced they're going to do NVIDIA and AMD versions of their box.
- renewiltord 3y ago
  Interesting news. Thank you for that update.

causality0 3y ago

AMD GPUs are fantastic values for money when it comes to running mainstream PC games. Once you step outside that narrow box they become infuriating. My 7900XTX is probably the last AMD GPU I ever buy. It's just horrible at all the other things I want to use it for, especially and gallingly, game console emulation. For the first few months I had it I actually kept my old GTX1070 plugged into my motherboard because the eight year old $400 GPU was so much better at half my use cases than the $900 AMD card.

MuffinFlavored 3y ago

> The problem is if I spend an evening trying to do anything with OpenCL or ROCm the kernel hard-locks and I go to bed early. In a race against NVIDIA where every comment on here is about just how much better CUDA is than any alternative, why doesn't literally what you just said "also" have the attention of Lisa Su?

georgehotz 3y ago
I have had multiple replies from Lisa Su on both Twitter and by e-mail. It doesn't help. AMD is structurally incapable of fixing these issues. They don't even have a 7900XTX in CI. Crashing the firmware is so trivial there's no way anyone fuzzed anything. They debug by the application, adding mitigations at many layers of the stack to make it work. While this strategy is fine for the 20 mainstream games that come out per year, unless you root cause issues and have a good CI, you'll never build a stable GPU for general compute.
- MuffinFlavored 3y ago
  Genuine question for you then, why are you investing your time reversing their firmware if it sounds like you think they're failing/will fail? Why not just "adopt" the golden child/golden standard NVIDIA now for your usecase, reverse that (if needed) and move on? Genuinely curious what you gain from trying to make AMD work/investing in AMD on your end here. Is it that you see a value proposition through the "muck" of crashy-kernels that I'm missing?
  - georgehotz 3y ago
    This is my weekend hobby, I'm learning about GPUs. Besides, I'd like to see competition to NVIDIA. From a business side, tiny corp is selling boxes with both AMD and NVIDIA, we'll let the user choose how much they value money vs pain.
    - rudedogg 3y ago
      What are your thoughts on using Intel GPUs for deep learning?

ultra_nick 3y ago

NVidia is making boatloads of money because their driver works and they have a software library called CUDA that accelerates neural networks. Nobody expects AMD to match them, but George thought they could at least write a GPU driver. If AMD can get that driver out, then George could provide a competitor to CUDA (for neutral networks only). They'd both make boatloads of money. However, AMD was less capable than expected and their drivers were too buggy to run neural networks like those needed for the MLPerf benchmark. So now, it appears that AMD, Tinybox, and investors like me won't be making boatloads of money.

DiabloD3 3y ago
Slight mistake in your description: CUDA is an out of date API that was replaced by Khronos's own official compute APIs. Khronos is a standards consortium that Nvidia is a founding member of. Although the marketing department at Nvidia still pushes for greenfield CUDA codebases, no new code should be written in it, and they should opt for open source international standards only. Khronos APIs are implemented by over 120 vendors.
- greenavocado 3y ago
  Show me the announcement deprecating CUDA
- mshockwave 3y ago
  With my past experience in Khronos, NVIDIA is indeed a member and they sent decent guys to the meetings -- but only for strategic reasons, rather than "advocating for open standards" as you described. My experiences there actually told me the opposite that they will never drop CUDA. Objectively they also have incentive to do so: fighting with 100-ish companies to ratify something is always slower than rolling out an feature in an ecosystem you have total control of.
  - pixelpoet 3y ago
    Not just that, but they get to abuse their near monopoly to strongarm many companies into using Cuda. Can't say more unfortunately...

NicoJuicy 3y ago

That's not really on how I see it. He's trying to do everything on AMD's consumer based GPU's, while he's comparing it with NVidia's enterprise GPU's with the same expectations. That's just not how it works in real life. But if he succeeds, yeah, he could save tons of money and his hardware for ML could actually be completed.

modeless 3y ago
This is wrong for two reasons. One, he claims that AMD's enterprise GPUs are also unreliable. Two, he is comparing vs. 4090 which is a consumer GPU and CUDA is very reliable as it is on all of Nvidia's consumer hardware.
dannyw 3y ago
NVIDIA’s consumer GPUs are highly reliable and support the same CUDA capabilities as the enterprise products. This is something NVIDIA gets right. Meanwhile, ROCm isn’t even supported on certain lower end AMD consumer chips.
- llm_trw 3y ago
  Rocm is only supported on the last generation of flagship graphics cards. Which is a shame because the Radeon VII is still better than anything else when it comes to 64bit float computation.
  - mnau 3y ago
    "Supported" I watched geohot streams and if he can go for an hour with a need to restart the computer due to crashed kernel driver/locked up card, it's a miracle (various computers, one or multi-card setup). And he is not trying to crash the driver, just making the most basic stuff work. I seriously think there is a deep HW bug that caused the difference in performance between RDNA3 presentation and reality and they had to do some horrible hacks to make it work. Either that or the card never seen a tester for compute. Which is strange, considering they added it as a "supported" option.

trws 3y ago

They do need to rework their development processes. It’s worth noting though that there are two (kinda three) different drivers with very different quality and different teams behind them. There’s the binary driver, on windows and the fglrx driver for Linux, that contains the full OpenGL stack and all the proprietary stuff. That’s been maintained and extended since the ATI days by a team primarily based in Markham. It’s garbage. It’s been garbage since the 90s. This is not new or surprising. It’s unstable, poorly maintained and just generally a bad time. It was also the reason OpenCL required an X session run as root with no access control and sometimes a stub dvi cable on the card for many years. It is a source of endless sadness and despair. The other driver is the HSA/ROCM driver on Linux, which can be used with the amdgpu upstream kernel driver and is itself upstream. The amdgpu driver is 2d only, but has been developed in the open, and stable, for a long, long time. It’s maintained by a completely separate set of teams mostly in (last I knew) Germany and Austin. The HSA driver has likewise been immensely more stable than the binary drivers since it was introduced. There are potential problems with it, but I’ve managed to recover every single one I’ve run into without a reboot, and I’ve been doing GPU compute with AMD, NVIDIA and intel since 2008. When it was the binary driver, we needed reboots every time an AMD GPU locked up. I’m not saying there’s no problem here, but he’s reporting problems I’ve never seen in thousands of hours of high-load compute both personally and for work with the driver one ought to use for ROCM. There are other problems, but HSA/amdgpu drivers have not been high on that list for me. Possibly because I was so used to how bad the other ones were I guess…

rstat1 3y ago
I believe your info is a bit out-of-date. FGLRX hasn't been maintained or actively developed for a quite a while and unless you're using super-duper-ancient HW no one actually uses it anymore. AMDGPU is the the current Linux driver, combined with Mesa, and certainly is not 2D only.
- pezezin 3y ago
  Yup, FGLRX was deprecated from all major distros in 2016. Nobody has used it for ages.
- 3y ago

andrewmcwatters 3y ago

It was particularly hard to listen to himself admit that he basically became unpaid QA for AMD. Like, dude, you’re making the same junior mistake people writing apps for the App Store make. It’s not your platform.

llm_trw 3y ago
If the AMD code makes it into the kernel, like their graphics drivers did, then we've won because it will just work on every kernel update. Compared to now where AMD drivers just don't work and NVidia drivers only work with the right blobs installed I'd be willing to take a paycut for 6 months to have 10 years of painless development. That said I jumped ship to NVidia since I need to work and not spend all my time trying to get servers to the point that they can work.
threeseed 3y ago
As a developer you're always building on top of something you don't control. At least with the App Store you make pretty good money. It's worse when you have companies who have APIs you need that you know they don't care about.
modeless 3y ago
His thesis is that owning a good software platform is a prerequisite for a successful AI hardware company. He is (or was) using AMD as a stepping stone to create and flesh out that software platform (tinygrad) before starting custom ASIC development, instead of going the other way around with a costly ASIC first and trying to develop a software platform after, or trying to push support for an ASIC into someone else's platform such as PyTorch. Not sure I agree but it's an interesting idea. It's also worth noting that tinygrad is bypassing as many layers of AMD's software platform as they can get away with (MIOpen, ROCm, and even the userspace GPU driver).
- andrewmcwatters 3y ago
  Well, I think he's probably right, it's just a shame that he's going through this difficulty. I hope he ends up succeeding in his goals.

lannisterstark 3y ago

>I'm cheering him on, even though his complaints were a little melodramatic. I've really soured on him when he literally went "I'm going to fix twitter search, hire me elon" and then went to public saying "Hey, want an internship? Fix it, and then MIT license your code(so he could use it lol), and maybe we can talk - oh btw I have no real authority to give anyone an internship." He might be a brilliant programmer and problem solver, but that incident left a very sour "Grifter" vibes.

quantumwoke 3y ago
I soured on him a bit more when I watched some of his stream yesterday and he was ranting about diversity being a root of evil in Silicon Valley. Struck me that he hasn’t matured much since he was slinging mud at Sony and Apple.
- UberFly 3y ago
  He seems to be the type who doesn't care who's "soured on him".
hughesjj 3y ago
Yeah, I was feeling similar vibes back when Lex Friedman interviewed him for some self driving stuff back in the day (...I've soured on lex too since he's not really an interviewer/journalist so much as a host/promoter and rarely gets to anything interesting in his talks imo). His work for jailbreaking/actually being able to own the devices we pay for back in the iPhone+PS3 days was and is still inspiring, but I think being so smart and encountering so much bullshit in his life kind of led him to the 'rockstar coder I know better than you regardless of domain' personality, and that leaves him both a bit grating and underquipped to drive consensus. He's still a great wrench to throw into the bullshit machine though, and the guy is smarter and more accomplished with computers than I'm ever likely to be.
- infogulch 3y ago
  This is a great take. We need more people that are willing and able to jump in and wreck new bullshit machines as they crop up in diverse industries. They can even keep their boastful grandstanding attitude as compensation.

benreesman 3y ago

I’m going remind everyone that AMD and NVIDIA are atypically competition avoidant, but I don’t want to defend a generalization that broadcast let’s be specific: Hopper and MI300 are on paper just straight up competitive, both super scarce, high-margin cards that really love PyTorch, (which really loves NVIDIA, even TPU is, not the fist among equals). But all the MI300 is going to supercomputer and adjacent things, and all the Hopper is going to giant tech LLM type stuff, it’s not the same people bidding on those bins of those. Oh, and their respective CEOs are closely blood-related and in a trivially first name if not family gathering basis. The people root for George generally think this isn’t real capitalism. We think it’s a trend towards the failure to enforce anti-trust laws.

dotnet00 3y ago
>I’m going remind everyone that AMD and NVIDIA are atypically competition avoidant, but I don’t want to defend a generalization that broadcast let’s be specific: Hopper and MI300 are on paper just straight up competitive, both super scarce, high-margin cards that really love PyTorch, (which really loves NVIDIA, even TPU is, not the fist among equals). > But all the MI300 is going to supercomputer and adjacent things, and all the Hopper is going to giant tech LLM type stuff, it’s not the same people bidding on those bins of those. They clearly compete fiercely, every feature NVIDIA adds to its consumer GPUs gets an equivalent version by AMD. Previously AMD has introduced features that have caught on enough to force NVIDIA to implement them. AMD's offerings often force NVIDIA to discontinue certain products or to drop prices. The reason MI300 is going to supercomputing and Hopper to AI is the self fulfilling prophecy that AMD's software still sucks for AI, thus they have to target general supercomputing. >Oh, and their respective CEOs are closely blood-related and in a trivially first name if not family gathering basis. This is mostly just a meme, they're distant relatives, which is not too surprising considering they're both from Taiwan. Jensen's maternal uncle is Lisa's grandfather, with said maternal uncle being the eldest of 12 siblings, 18 years older than Jensen's mother. I barely know the names of the eldest of my aunts, who have a similar age gap, let alone have a close relationship with them, because that age gap and number of siblings means that even my mother barely knows them.
- 3y ago

braiamp 3y ago

Common confusion. It isn't the kernel driver that has problems. Not their userland "drivers". It's their firmware inside the GPU's.