why is it eternally inferior to cuda?
>>106776174It's literally half the speed is hilarious Nvidia I assume patented the only decent way to handle memory interleaving
>>106776174because of Lisa Su's collusion with her cousin Jensen Fag. Zluda proves that it's all artificial
>>106776255It's also half the price. You're not going with AMD because it's the fastest, you're going with them because they're the most cost-effective and will save your business money.
>>106776523>Half the speed>Half the price Were are the savings? You're paying more than that negligible difference in energy costs and dev time.
>>106776877Dev time is the same and electricity is grandfathered in.
>>106776174cuda is a good example of patent trollingthere's only so much you can do without infringing on some imaginary bullshit
>>106776336Honestly im glad Trump is getting rid of intel nvidia and AMD the new company's that will be replacing them will be 100% better
>>106776523but a business will lose money when amd crap will stop to workmore than that a business will have to hire own software developersmeanwhile nvidia provides support for that double price and cuda just works
>>106777307I hate AMD for the intentional gimping of their drivers but their hardware lasts. Nvidia hardware are ticking timebombs. I've had at least 7 nvidia GPUs prematurely die and all my AMD cards still work.
how could it not be when it can't even work on windows
>>106776174Someone at Nvidia actually had the foresight to make sure that their GPUs could also be used for scientific computing workloads, which led to CUDA and a massive headstart.AMD has just been horribly slow to catch up.
>>106777729> AMD has just been horribly slow to catch up.they didnt try or start
>>106777450>I've had at least 7 Nvidia GPUs prematurely dieDo you recall the lifespans?
>>106776174>why is it eternally inferior to cuda?every card since the 8000 series (2007) supported cuda, full stop. no further questions.rocm? lol. not even amd knows what version works with which cards for how long, maybe.
More than half of the nvidia engineers work on software. What's the percentage for AMD? ATI/AMD always was anemic with software.https://www.linkedin.com/posts/naderlikeladder_nvidia-has-more-software-engineers-than-hardware-activity-7304987279251415040-jrUl
>>106776174Because CUDA started first. That's the only reason.
>>106777729No it's a very simple architecture problem there is no "catch up" "better choosers"It's literally the memory and execution interleaving protocol One of them is the best, that's cuda. Doesn't get any betterProblemPatented So you must take second best, and ironically, obviously, AMD actually took 3rd best to have room to growBut you can't fuzz or shade between 3rd and 2nd, it's a single update stepAnd you go from 4 cycles per wave to 2. Nvidia has 1Which means it's literally a 50% increase in speed. Why the fuck would they do that today, when they could do it tomorrowSame with apple and no headphones jacks. It's guaranteed money better to wait, due to inflation etc, but it ends up never happening, bc the rate at which you get more money is always increasing
>>106776174Don't care. Still buying AMD.
>>106780149Why are they still incapable of building a Linux kernel driver and software stack?More engineers doesn't always mean better.
>>106776883time value of money, opportunity cost and first mover advantage
>>106779263>not even amd knows what version works with which cards for how long, maybe.Refer to:https://llvm.org/docs/AMDGPUUsage.html#processorsAnything supported by LLVM likely works with ROCm too.
>>106785752Not every business is looking to move first though. There is no doubt that NVIDIA is the market leader but most businesses will be perfectly content to follow rather than lead.AI will change the course of business and NVIDIA will play a part in that but not every business will need the most expensive solution to do that.
>>106785743Dogfeeding their AI coding
>>106776174>steals your proprietary anti-freeware with SCALEnot my problem
>>106777729cuda isnt hardware, its just regular gpu cores you have no clue what youre talking aboutif anything your criticisms only apply to intel for not properly implementing their oneapi plans and being too late
>>106785771that's the thing: as a business relying on gpgpu, you cannot bet on guesswork. shit has to work no matter what. nvidia offers just that, even if it's expensive af.
I think the reason things are so bad is academic usage. Nvidia won long ago there, and stuff like AI is mostly following a pretty short route from research to commercial use. Models get dropped and implemented into tools within 24h a lot of the time. People build runtimes for this stuff to depend on CUDA, and with that, you tend to get automated setup for Nvidia hardware.>t. attempting to get a ryzen AI APU to do useful local AI stuff and struggling for a multitude of reasons
>>106776255If only that were trueIt doesn't even work with an out of box installation and requires FAFOing about to even work on windows
>>106785743While it's true that hardware companies tend to be shit at software, in the case of nvidia and Linux for home users, it's likely that nvidia doesn't give a shit about that market segment so they don't allocate resources to it beyond the bare minimum to have a driver that displays a desktop.
>>106777648yea no the windows SDK came out ages agoin fact it worked on windows for like a year? before AMD even put out the sdk since it was secretly bundled in the drivers mostly just so blender could use it opencl style at runtime via hipewand true cuda style single binary compilation was possible even at that stage if you knew enough about clang's offloading infrastructure
>>106786840So buy one of the supported Instinct GPUs then?
>>106778729They all fail within 6 months of being in my cum splash zone. AMD fans blow through my dried gunk. Nvidia fans always seize after 6 months of buildup. What a pitiful company.
>>106788934skill issue
