Announcement

Collapse
No announcement yet.

AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #71
    Originally posted by Panix View Post
    $1000+ cards for only gaming? Why not just buy some console with AMD gpu hardware - a $1000+ gpu that can't do anything else is outrageous. Nvidia's tech - not just CUDA - is superior in everything - so, I am not sure why the fanboys rage about it - oh yeah, the open source. Yeah, good point there -except, for when you need productivity software and need to use the closed driver - and then you are limited to a few enterprise distros. Right?
    you really sould read the news for example this: FSR3 outperforms DLSS on a Nvidia 4090.... this means of course Nvidia tech is not superior.



    "except, for when you need productivity software and need to use the closed driver - and then you are limited to a few enterprise distros."

    can you tell me the name of the productivity software who use the closed driver ??? because the AMD.com driver just installs the open-source driver on the enterprise distros.

    also you do not unterstand the staple amd kernel driver API AMDGPU because this stable kernel API makes it possible to use RADV and AMDVLK and also CLosed source openCL and opensource openCL and the closed source FGLRX driver all at the same time.

    this means you can run one application with RADV the next application with AMDVLK and you can run a old closed source openCL driver runtime at the same time... all connect to the same kernel driver AMDGPU
    Phantom circuit Sequence Reducer Dyslexia

    Comment


    • #72
      Originally posted by qarium View Post

      you really sould read the news for example this: FSR3 outperforms DLSS on a Nvidia 4090.... this means of course Nvidia tech is not superior.



      "except, for when you need productivity software and need to use the closed driver - and then you are limited to a few enterprise distros."

      can you tell me the name of the productivity software who use the closed driver ??? because the AMD.com driver just installs the open-source driver on the enterprise distros.

      also you do not unterstand the staple amd kernel driver API AMDGPU because this stable kernel API makes it possible to use RADV and AMDVLK and also CLosed source openCL and opensource openCL and the closed source FGLRX driver all at the same time.

      this means you can run one application with RADV the next application with AMDVLK and you can run a old closed source openCL driver runtime at the same time... all connect to the same kernel driver AMDGPU
      AMD === Like a magic

      Comment


      • #73
        Originally posted by Panix View Post
        You're the one who's trolling - 7900 XTX cards are around the same price or more expensive than the 4080 Super and 4080s already didn't sell very many, apparently - but, the same can be said about 7900 XTX cards - in fact, they increased price for most of their shelf life. There was a temporary discount but then they all increased in price.
        You're being absolutely disingenuous.
        In both cases they were above their respected MSRP when it came to AIBs, whether or not certain cards entered each others price territory in the after market is irrelevant,
        do I really need to start digging up articles comparing the price? The cheapest RX 7900XTX still fell way below the cheapest RTX 4080.

        Originally posted by Panix View Post
        Perhaps, you're living in the Wizard of Oz land where it's just a fairy tale. A gpu series that has almost no features that beat Nvidia - plus, lacking in most features outside of gaming should be WAY CHEAPER - no one should be praising AMD or saying how great it was that they were $100 cheaper than their competitor card! LOL! You must be crazy.
        No you live in a fairy tale were all everyone cares about is Ray Tracing, DLSS and CUDA. What are you, Sophisticles alt?
        There are non gaming cases where the RX 7900XTX beats the 4080 in performance, there are cases the extra VRAM is the difference in working software or not (wait a minute I already linked this in my previous comment, almost as if you are selectively ignoring points that undermine your argument).
        For some people that extra gaming performance alone can be worth the difference since they don't bother with any of the other stuff, that is just how capitalism works.

        ZLUDA (you know what the article of this comment section is about) has proven that the large performance difference in some software is software related.
        Would it be ideal if AMD could make it so ROCM/HIP can achieve that level (or even beyond as ZLUDA couldn't utilize Optix) of performance natively?
        Absolutely, doesn't take away however that in certain uses cases the AMD cards already perform very well for their pricepoint and that the hardware has some potential for future performance gains in other places.

        And before you go build the brainless strawman (or scarecrow I guess) on what I just said Oz-boy, no that obviously does not mean you should buy AMD on the premise of better future performance, if your software relies on CUDA or for any other reason runs undeniably better on Nvidia hardware you obviously go Nvidia. It's just that it is clear that there are scenario's where AMD can outperform Nvidia and maybe you or someone else could find yourself to be in those scenarios and have no care of need of performance outside of those.

        Originally posted by Panix View Post
        $1000+ cards for only gaming? Why not just buy some console with AMD gpu hardware - a $1000+ gpu that can't do anything else is outrageous. Nvidia's tech - not just CUDA - is superior in everything - so, I am not sure why the fanboys rage about it - oh yeah, the open source. Yeah, good point there -except, for when you need productivity software and need to use the closed driver - and then you are limited to a few enterprise distros. Right?
        Aside from the fact I am now worried you actually do have some form of memory loss as you responded to the same comment twice.
        Care to back up that you are limited to few Enterprise Distros? You can install AMDGPU-PRO on Gentoo, same with Arch through the AUR, just because it isn't listed on the AMD download page doesn't make it impossible (something they might have done so they don't have to bother with every tech illiterate distro hopper to come knocking at their door).

        Also it is pretty clear you are an Nvidia fanboy, you seem to never, ever respond to cases whenever AMD does score some better performance and go back on your tirade of CUDA better. People that aren't fanboys could face reality and see that while not necessarily common, there are cases where AMD fulfills a performance niche and for those reasons can be an option (or maybe as you said yourself due to the Open Source aspect of it, some people care, some people don't almost as if they have individuals wants and needs that can be met) but that obviously when it is beat for your use case you go with Nvidia.

        Comment


        • #74
          Originally posted by Panix View Post
          You're the one who's trolling - 7900 XTX cards are around the same price or more expensive than the 4080 Super and 4080s already didn't sell very many, apparently - but, the same can be said about 7900 XTX cards - in fact, they increased price for most of their shelf life. There was a temporary discount but then they all increased in price.
          It's pretty easy to check this.

          2024.02.14.gpu.chipset.radeon-rx-7900-xtx.a0b9da1129e3dd7bff4f61a4462c8af3.png

          Looks like a slight decline from launch and then flat.

          Comment


          • #75
            Originally posted by Developer12 View Post
            No, that's not how HIP works.
            Sure it is.

            Let me explain it with an example, let's assume we have a ray tracer written in CUDA3:
            • with ZLUDA you take the binary and run it on AMD card
            • with HIP you use the translator and may have to alternate the code a bit
            If the ray tracer gets updated to CUDA4 the old versions still runs in both cases, but if you want the new version:
            • ZLUDA needs support for the new version
            • HIP translator needs to support the new version or you need to manually put in much work to port it yourself
            So yes it's the same situation just more work for the user with the HIP translator.

            Comment


            • #76
              Originally posted by Anux View Post
              Sure it is.

              Let me explain it with an example, let's assume we have a ray tracer written in CUDA3:
              • with ZLUDA you take the binary and run it on AMD card
              • with HIP you use the translator and may have to alternate the code a bit
              If the ray tracer gets updated to CUDA4 the old versions still runs in both cases, but if you want the new version:
              • ZLUDA needs support for the new version
              • HIP translator needs to support the new version or you need to manually put in much work to port it yourself
              So yes it's the same situation just more work for the user with the HIP translator.
              Requiring that the user learn enough GPGPU programming to "alternate the code a bit" is a pretty enormous hurdle even assuming the application is open source.

              Comment


              • #77
                Originally posted by yump View Post
                Requiring that the user learn enough GPGPU programming to "alternate the code a bit" is a pretty enormous hurdle even assuming the application is open source.
                Of course that's not so easy, which was my point. More work with HIP but no advantages.

                Comment


                • #78
                  Originally posted by tenchrio View Post
                  You're being absolutely disingenuous.
                  In both cases they were above their respected MSRP when it came to AIBs, whether or not certain cards entered each others price territory in the after market is irrelevant,
                  do I really need to start digging up articles comparing the price? The cheapest RX 7900XTX still fell way below the cheapest RTX 4080.

                  No you live in a fairy tale were all everyone cares about is Ray Tracing, DLSS and CUDA. What are you, Sophisticles alt?
                  There are non gaming cases where the RX 7900XTX beats the 4080 in performance, there are cases the extra VRAM is the difference in working software or not (wait a minute I already linked this in my previous comment, almost as if you are selectively ignoring points that undermine your argument).
                  For some people that extra gaming performance alone can be worth the difference since they don't bother with any of the other stuff, that is just how capitalism works.

                  ZLUDA (you know what the article of this comment section is about) has proven that the large performance difference in some software is software related.
                  Would it be ideal if AMD could make it so ROCM/HIP can achieve that level (or even beyond as ZLUDA couldn't utilize Optix) of performance natively?
                  Absolutely, doesn't take away however that in certain uses cases the AMD cards already perform very well for their pricepoint and that the hardware has some potential for future performance gains in other places.

                  And before you go build the brainless strawman (or scarecrow I guess) on what I just said Oz-boy, no that obviously does not mean you should buy AMD on the premise of better future performance, if your software relies on CUDA or for any other reason runs undeniably better on Nvidia hardware you obviously go Nvidia. It's just that it is clear that there are scenario's where AMD can outperform Nvidia and maybe you or someone else could find yourself to be in those scenarios and have no care of need of performance outside of those.

                  Aside from the fact I am now worried you actually do have some form of memory loss as you responded to the same comment twice.
                  Care to back up that you are limited to few Enterprise Distros? You can install AMDGPU-PRO on Gentoo, same with Arch through the AUR, just because it isn't listed on the AMD download page doesn't make it impossible (something they might have done so they don't have to bother with every tech illiterate distro hopper to come knocking at their door).

                  Also it is pretty clear you are an Nvidia fanboy, you seem to never, ever respond to cases whenever AMD does score some better performance and go back on your tirade of CUDA better. People that aren't fanboys could face reality and see that while not necessarily common, there are cases where AMD fulfills a performance niche and for those reasons can be an option (or maybe as you said yourself due to the Open Source aspect of it, some people care, some people don't almost as if they have individuals wants and needs that can be met) but that obviously when it is beat for your use case you go with Nvidia.
                  I never disputed that AMD 7900 xtx cards were NOT a lot cheaper than 4080s - 4080s were way overpriced and that's why Nvidia did this "PR job" - your accusations that I'm a Nvidia 'fanboy' is laughable. I've only ripped Nvidia many times on here - as I think both companies are awful, greedy companies who have awful business practices and overpriced hardware. I'm just reacting to AMD fanboys like you and others who are always defending AMD like they are some sort of employees or pals of Su or something. This is done simply because AMD has an 'open source' driver in Linux, I guess? *Shrugs*

                  Their cards don't have to be so expensive - there's lots of claims online that they'd sell more 7900 xt and 7900 xtx cards if the prices were lower all that time and they STILL didn't lower them EVEN THOUGH the 4080 Super was at a much LOWER MSRP than the original 4080. Nvidia is now contributing to this 'out of stock' nonsense that we've seen before and scalpers are buying up the cards at msrp and then when they go out of stock, they put them on the 'used market' at prices over msrp. That's what I see, anyway and that's not on topic but I hope pointing that out will help ease you out of of your 'nvidia fanboy' delusions you have.

                  "ere are cases the extra VRAM[/URL] is the difference in working software" - where? In a few games? I don't care - that's not the normal or regular situation. If it's games - how many games? Which games? Also, DLSS is perceived as better than FSR (3) and that's in practically all games, afaik. Want to try again? I really don't care about that part - as I think the difference is probably not that great. But, if you want to make that point....

                  ZLUDA - this is just recent and AMD practically had nothing to do with it - they hired a guy - and they're afraid to associate themselves (officially) with the 'feat.' LOL! You bringing this point up just defeats your 'pro-AMD' nonsense .....it's actually laughable - AMD can only 'hack CUDA' to catch up or close the gap and their own software tech is so bad, they need to hack CUDA - and for e.g., in Blender - it's not CUDA that gives them a great advantage, it's the ray tracing acceleration - OptiX - and AMD can't even implement their own yet (HIP-RT) - it's just a mess and many theories out there are suggesting that whenever they get their head out of their arse with it - it will not perform well enough to close the gap.
                  They're barely catching up with CUDA - but, it's not Optix they have caught up to - which everyone on here forgets (looking at the Blender graph/performance info benchmarks).

                  It's pathetic that AMD had to get someone hack CUDA to catch up and then they're too afraid to associate with it (legal reasons, probably - what else could it be) - or that it would look bad as others on here suggest (theory: ppl would ask, 'why buy an AMD card when Nvidia will deliver the official 'CUDA' anyway?). Good work, AMD! This is not to ignore that employee's accomplishment - that's actually pretty cool and whatever it takes for AMD to catch up and offer competition and an alternative, is commendable and needed - to counter with Nvidia as it's practically a monopoly already and badly needs competition, whether from AMD or Intel.

                  Comment


                  • #79
                    Originally posted by yump View Post

                    It's pretty easy to check this.

                    2024.02.14.gpu.chipset.radeon-rx-7900-xtx.a0b9da1129e3dd7bff4f61a4462c8af3.png

                    Looks like a slight decline from launch and then flat.
                    It might have been flatter in the USA - also, the reference MBA card was cheaper than AIB cards - however, that card had several issues - including the Vapor Chamber debacle and I suspect sales of the reference/MBA version fell - but, I bet you can't find a graph with that one - that's just taking the average. However, in many countries not the USA - I think the price actually went up after launch - at least, in my country it did. It only recently reduced a bit but it's still $1300 cad/$960 USD for the cheapest ones. Most are well above that. I recall Americans who talked about the prices - saying $899 USD was somewhat acceptable (occasionally, there'd be a one-day sale at price but it was just temporary).
                    The price here was mostly $1400 (for most of its retail life) until the most recent price drop (around $100).

                    You can obviously check the graph of prices for the more common 7900 XTX AIB cards - to see how expensive they were - e.g. the XFX 7900 XTX - usually, one of the cheapest ones:


                    You can cross check your own country...maybe it's cheaper there, I dunno.

                    Comment


                    • #80
                      Originally posted by harakiru View Post
                      The article seems to mention that it was tested with llama.cpp however i cant find any information about it on the github page.

                      Edit: I tested it on blender 4.0 using my RX 6800 and it seems to be running really well. Props to the developer since the rocm backend always crashed my pc and i had to use cpu rendering for the past year.
                      Do you have any sources to compare? The open data has some sample scores and there are a few online\YT videos in which 6800 cards were benchmarked and tested against other cards, both AMD and Nvidia - various scenes were tested including Classroom, Lone Monk, Barbershop etc.

                      Comment

                      Working...
                      X