[HN Gopher] AMD Instinct MI300X Accelerators
       ___________________________________________________________________
        
       AMD Instinct MI300X Accelerators
        
       Author : bx376
       Score  : 22 points
       Date   : 2023-12-25 05:22 UTC (17 hours ago)
        
 (HTM) web link (www.amd.com)
 (TXT) w3m dump (www.amd.com)
        
       | frognumber wrote:
       | I stopped supporting AMD when I bought an AMD card, and they
       | discontinued support a few months later. If AMD sends me an
       | MI300X for free, I'll make sure my software is compatible.
       | Otherwise, if you want to run my stuff, you need Nvidia (or,
       | hopefully soon, Intel).
       | 
       | (Note: I know they never will. This story, multiplied by the
       | number of people building stuff now, is why I'm bearish on AMD
       | and bullish on competitors. This would be a very nice card --
       | 192GB is sexy -- but I just get headaches thinking about
       | compatibility, software, and drivers issues. To quote Steve
       | Balmer, what AMD is missing is: "Developers, developers,
       | developers." That's not a small hole to dig out of.)
        
         | vlakreeh wrote:
         | What card could you have bought that lost support months later?
         | I can't think of any new card you could have bought that'd lose
         | support that quickly as far back as 2016
        
           | KeplerBoy wrote:
           | Not GP, but they dropped support for SKUs which were still
           | sold at that point. The MI50 would be a fun accelerator to
           | play with and easily worth 2000$ if it were made by Nvidia.
           | Since they are made by AMD and not supported, you can grab
           | them for 200$.
           | 
           | https://github.com/ROCm/ROCm/issues/2308
        
             | simfree wrote:
             | Support still exists, but the card is in maintenance mode.
             | This doesn't mean that your models that work on an MI50
             | today are suddenly going to stop working, or the models ran
             | on my MI25 will cease suddenly.
        
               | frognumber wrote:
               | That's only true if you don't mind sharing your systems
               | with strangers in Russia, Asia, and Africa.
               | 
               | "Bug fixes / critical security patches will continue to
               | be supported for the gfx906 GPUs till Q2 2024"
               | 
               | Worse, a lot of tooling will break before then. If you
               | need a security patch for anything built on top of ROCm
               | or CUDA, you often can't have an ancient version.
               | 
               | That's not to mention that I'd like to do new development
               | and prototyping.
               | 
               | I don't do business with AMD for the same reason I don't
               | do business with Google. They're not a reliable partner.
        
               | nixgeek wrote:
               | That the only bad actors are in Russia, Asia and Africa
               | is becoming quite the trope.
               | 
               | Some very spectacular compromises by Lapsus$ aren't
               | originating from any of the above, but instead from the
               | West, specifically the UK:
               | 
               | https://www.bbc.com/news/technology-67663128
        
               | frognumber wrote:
               | I think you're missing the point. Bad actors are equally
               | distributed around the world.
               | 
               | The difference, however, is that, in this case, as the
               | article points out, the bad actor is in custody.
               | 
               | My local law enforcement, lawyer, and judicial system are
               | able to help out with the bad actors here and keep the
               | problem somewhat contained.
               | 
               | For a bad actor in Russia or India, you quite literally
               | have no recourse. There are literally Youtube videos
               | (Mark Rober did a special recently) showing out-of-
               | jurisdiction scammer and hacker shops occupying entire
               | office buildings, and there is almost nothing we can do
               | about it.
               | 
               | I'm not even going to go on a diatribe about corruption
               | and the quality of law enforcement, since one can have
               | different opinions. What I will point out is that US+EU
               | have extensive treaties which allow my local law
               | enforcement agency to cooperate closely with law
               | enforcement agencies between US, UK, and EU.
               | 
               | For a bad actor in Iran or North Korea? You're not even
               | getting diplomatic contact.
               | 
               | Footnote: I mention India since they're a democracy and
               | by all standards, try to act responsibly on the global
               | stage. They just don't happen to be in-network for
               | various US/EU-centric international conventions.
        
               | simfree wrote:
               | The ROCm tooling only supports Ubuntu, Fedora and a few
               | other enterprise distro, but it is on track to make it
               | into the main Debian archive at which point the Debian
               | maintainers will backport fixes as needed.
               | 
               | The tooling for older MI25 cards has yet to break, please
               | avoid creating FUD on hardware you don't own.
        
               | KeplerBoy wrote:
               | You are right, but I'm not willing to bother with that.
               | From what I've read rocm is a mess to work with even on
               | supported hardware, so yeah..
        
               | simfree wrote:
               | Install Ubuntu, install ROCm, run your model. It's really
               | not that complicated, even with ancient MI25 cards.
        
           | frognumber wrote:
           | https://github.com/ROCm/ROCm/issues/1353
           | 
           | Bought in 2020. Stopped working in 2020. Not the latest, but
           | in-production, advertised ROCm-capable, and what I could find
           | during the Great GPU Shortage of 2020.
        
             | justinclift wrote:
             | Wow, not even an industry-standard-for-many-decades
             | deprecation notice up front, to give people a heads up
             | before hand.
             | 
             | That's pretty fucked. :(
        
               | frognumber wrote:
               | It's even worse.
               | 
               | Order of operations:
               | 
               | 1. It just silently stopped working, started crashing,
               | and I had no idea what was going on or why. Just sort of
               | intermittent f-age where some older versions would work
               | better, and newer ones would work worse. I never got it
               | reliably working. In most cases, at some point, the
               | system would become unresponsive, and then crash hard.
               | 
               | 2. AMD removed it from the supported list (with no
               | notice).
               | 
               | 3. The github ticket above was filed, which explained
               | what was going on.
               | 
               | Lots of time wasted debugging. This interacted with a
               | half-dozen other AMD bugs and issues (such as learning
               | the GPU only worked for compute headless; I needed to
               | drive my monitor with a different card).
               | 
               | Human time is more expensive than equipment, so the total
               | cost of this stuff was astronomical.
               | 
               | Yes, notice is industry-standard, but at the very least,
               | when support broke / was later removed, the driver could
               | try warning me "We no longer support this card, do you
               | REALLY want to proceed?" rather than letting me know by
               | hard-crashing my system.
        
         | JonChesterfield wrote:
         | At least some of this is "support" meaning different things to
         | different people.
         | 
         | AMD means something like "your engineers can raise jira tickets
         | related to the supercomputer you bought". The general idea
         | outside of that context seems to be "support" means "can run
         | stuff on the GPU". There's some analogy to games dev - if call
         | of duty doesn't run on an amdgpu, you also don't get to
         | complain to AMD about it, but the games studio certainly will
         | do. The games studio is the entity with the contact details.
         | 
         | ROCm is fairly likely to run on unsupported consumer cards if
         | you build the libraries from source or use them from someone
         | who did, which is just about viable from linux distributions
         | now. You don't get to raise official jira tickets about it when
         | it falls over because it's "unsupported". You'll have a
         | moderately bad time if you ignore the docs recommendations
         | about the kernel version for a given release, or if the
         | difficult task of building the libraries from source goes awry.
         | 
         | The established packaging idea of only shipping machine code
         | libraries for a subset of the GPUs you sell, and giving very
         | limited information on how to make those libraries exist for
         | other GPUs, is obviously making essentially everyone very
         | pissed off with AMD. Their engineers are aware of that and
         | disagree about what to do about it.
         | 
         | I personally think that'll resolve itself once Debian packaging
         | is generally usable. People will install ROCm from their distro
         | instead of from the official builds which don't have machine
         | code for gaming GPUs anyway. That will give rise to new
         | annoyances but at least the applications will try to run on
         | whatever GPU you've got locally.
        
           | frognumber wrote:
           | "Support," to me, means not dropping support in newer
           | versions of ROCm. See the current compatibility matrix:
           | 
           | https://rocm.docs.amd.com/projects/radeon/en/latest/docs/com.
           | ..
           | 
           | I bought my (in-production, advertised ROCm-capable) card
           | during the great GPU shortage of 2020, and literally anything
           | new stopped working a few months later when AMD decided to
           | update ROCm to drop support for GFX8 GPUs:
           | 
           | https://github.com/ROCm/ROCm/issues/1353
           | 
           | All the packages I was using (like Spacy) required later ROCm
           | almost immediately. There was literally no way to make them
           | work. Running older packages is a non-starter if you want
           | things like, say, your machines to not be compromised.
           | 
           | > if call of duty doesn't run on an amdgpu, you also don't
           | get to complain to AMD about it, but the games studio
           | certainly will do
           | 
           | If "Call of Duty" is advertised on the box, and doesn't run,
           | yes, this is false advertising, and "consumers" get to
           | complain to the graphics card maker or AMD. In this case,
           | small claims court would give me a refund if I sued AMD; an
           | advertised ROCm-compatible GPU stopped being ROCm-compatible
           | within the warranty period. It's just not worth complaining
           | or suing over a few hundred bucks.
           | 
           | I didn't need much performance, and this was purely for dev
           | -- to make sure I bought this purely so I could have
           | compatibility with AMD. AMD decided to not have compatibility
           | with me, and I said f--k AMD.
           | 
           | The result is that if you want to run my data pipeline, it
           | won't work on AMD. A major industry will hopefully
           | standardize on it, and guess who is locked out of selling any
           | GPUs to that industry if that happens?
        
             | JonChesterfield wrote:
             | That seems to be a different thing again. You had at least
             | one rocm version that claimed to work with your GPU and
             | then later versions dropped it. Your options then are to
             | stay with the rocm version that claimed to support it or to
             | build a newer rocm yourself, or get it from distros etc,
             | where at some point in the future it'll probably stop
             | working entirely.
             | 
             | I don't know what plays into which GPUs are listed as
             | supported on which versions of ROCm. I'm sure my 6800GT was
             | marked "supported" on some version and isn't on the current
             | one. I think my radeon vii was likewise "supported" on some
             | of them and isn't any more. I used a 5700 XT for a year or
             | so which was never on the blessed list. I obsessively build
             | everything from source instead of using the release
             | binaries though.
        
               | frognumber wrote:
               | Neither of those would have worked, but that's another
               | post. However, I had a third option, which I took: I
               | bought a working NVidia card, put my AMD card in my
               | attic, swore off AMD GPUs and ROCm forever*, and began
               | warn other people against ever relying on AMD.
               | 
               | That's the option I took.
               | 
               | Now, my software is tied to CUDA. Others are building
               | more on top if it. If people want to run my software (or
               | by proxy, and software which uses my software as a
               | platform), they have more-or-less the same option if they
               | accidentally bought AMD.
               | 
               | So far, I haven't heard from any users who accidentally
               | bought AMD.
               | 
               | Right now, I'm just waiting for Arc A770 support to
               | improve, so I can also support Intel. That will be the
               | cheapest way to run my tools, since you get 16GB for
               | $300.
               | 
               | I'm not the only one who took this path. I'm just more
               | vocal about it.
               | 
               | * Some exclusions apply.
        
               | sangnoir wrote:
               | Good for you. AMD is focussing on larger customers who
               | write their own software (that runs on supercomputer and
               | in datacenters). Targeting a high-margin, limited scope
               | market before prosumers/consumers is decent strategy for
               | a company that doesn't have Nvidia's deep pockets.
               | 
               | It is frustrating as a consumer - dealing with ROCm on an
               | consumer card was a massive pain for the period I
               | bothered to put in the effort. However, I remain bullish
               | on AMD, I think their current GPUs are their Bulldozer-
               | equivalent; I'm waiting patiently for the AMD GPU
               | Zen-1-moment
        
               | frognumber wrote:
               | What you described would have been a fine strategy.
               | That's not AMD's strategy.
               | 
               | To implement your strategy, you either don't ship mass
               | market GPGPU cards, or you stick a big "beta / prototype"
               | label on ROCm and make sure people use it only _caveat
               | emptor_. I 've bought plenty of half-baked things because
               | I wanted to tinker with them or help a company along. AMD
               | cards are very competitive for gaming too, which really
               | brings volume much more so than GPGPU, so there's no
               | reason to oversell.
               | 
               | AMD's problem is that they're claiming these things work,
               | and they don't. I was fortunate to learn that lesson
               | spending a few hundred bucks. I know people who learned
               | that lesson spending thousands. I wouldn't want to figure
               | out on a data center or a supercomputer whether or not
               | the same lesson applies. If a company is willing to
               | cheaply torch relationships as it figures out a strategy,
               | ship half-baked ill-thought-out software, and make bold,
               | unsubstantiated claims, I'll pick another vendor if at
               | all competitive.
               | 
               | AMD is losing mind share, credibility, and branding. That
               | will hurt if they ever have a product ready for the
               | broader market, and it will also hurt (although not as
               | much) with the strategy you described.
        
               | flamedoge wrote:
               | AMD hardware is amazing. AMD software needs catching up
               | to do. It's pretty amazing how Nvidia's strategy of
               | making compute tools available on most consumer grade
               | cards is paying off latent dividends -- consumers learn
               | on their own hw and become developers/researchers using
               | same stack, something AMD needs to realize. The result is
               | Nvidia is a trillion dollar company and AMD only 1/5 of
               | that.
        
       ___________________________________________________________________
       (page generated 2023-12-25 23:02 UTC)