GPU Computing in MATLAB

Speed up your MATLAB® applications using NVIDIA® GPUs without needing any CUDA® programming experience.
Parallel Computing Toolbox™ supports more than 700 functions that let you use GPU computing. Any GPU-supported function automatically runs using your GPU if you provide inputs as GPU arrays, making it easy to convert and evaluate GPU compute performance for your application.
In this video, watch a brief overview, including code examples and benchmarks. In addition, discover options for getting access to a GPU if you do not have one in your desktop computing environment. Also, learn about deploying GPU-enabled applications directly as CUDA code generated by GPU Coder™.
Learn more about Parallel Computing Toolbox: bit.ly/3JEiy5U
GPU Computing in MATLAB: bit.ly/3RAVCIN
Additional Resources:
Compare GPUs using standard numerical benchmarks in MATLAB: bit.ly/3telyk4
--------------------------------------------------------------------------------------------------------
Get a free product trial: goo.gl/ZHFb5u
Learn more about MATLAB: goo.gl/8QV7ZZ
Learn more about Simulink: goo.gl/nqnbLe
See what's new in MATLAB and Simulink: goo.gl/pgGtod
© 2022 The MathWorks, Inc. MATLAB and Simulink are registered trademarks of The MathWorks, Inc.
See www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand names may be trademarks or registered trademarks of their respective holders.

[@UsmanSaleemSulehri]

Wow this is really innovative? I wonder when will others start to know about how GPU is actually good for computations.

[@monkeydrufykun]

NVIDIA omniverse and Isaac Gym works exacly in the same way. While before you would have to train your RL agents on multiple cores of your CPU you can get an incredible speed up instantiating many environments in the GPU. It's a game changer in machine learning and robotics

[@stephaneduhamel7706]

Can we make CUDA a thing of the past and start using non brand-specific solutions like compute shaders or openCL?

[@wowimoldaf]

no since CUDA is widely spread, already. you cant replace C/C++ with Go/Rust for current computer industry. same goes for CUDA.

[@stephaneduhamel7706]

@@wowimoldaf Your analogy doesn't work because almost half of the used GPUs in the world can run shaders and openCL, but are unable to run CUDA code, while anything that runs Rust is still able to run C code. The worst part is that this limitation is entirely artificial, and the goal is to force developpers to buy Nvidia GPUs over the competition. I don't want CUDA to completely disappear, but i think people should start moving away from it as much as possible, just like it happend with languages like Pascal, COBOL, Lisp, BASIC and many more, which are still used for compatibility with legacy codebases, but are rarely used on new projects.

[@wowimoldaf]

@@stephaneduhamel7706 and again who uses "almost half of used GPUs w/o CUDA" for AI? because, in my understand you seems to be implying extremely old GPUs like Radeon 6k/7k area. i never saw single industry that using that area GPUs.

[@stephaneduhamel7706]

@@wowimoldaf CUDA isn't just about training AI, it's about speeding up parallelizable code. Also most GPUs sold in the last 5 years are made by Intel(70%) and are therefore unable to run cuda. AMD(no cuda) and Nvidia share the remaining 30% roughly evenly. Even accounting only for discrete GPUs, still more than 20% are made by AMD. And of course most people in the IA industry are using Nvidia cards, because most of the commonly used technologies rely on CUDA for hardware acceleration.

[@Keldor314]

Unfortunatly, OpenCL is unofficially deprecated by the entire industry, and I believe Intel is the single vendor that even has an up to date implementation. Cuda really is the only option in town these days, although Vulkan has some potential, although you'd have to write a SPIR-V build target instead of PTX, which really isn't a trivial undertaking. Also, Cuda has some very interesting functionality (kernel side memory management, for instance) which puts it head and shoulders above every other API. It really is no exaggeration to say that Vulkan, DirectX, Metal, and OpenCL are where Cuda was a decade ago. (Actually, Metal might be a bit less far behind. Pity it's Apple exclusive.)

[@salmiakki5638]

Does MATLAB support cuBLAS and cuDNN CUDA libraries?

[@williamgomez6087]

This is really usefull because you can actually prove your cuda-gpu drivers have got installed and are running¡ I mean, i jsut muved up to Linux, and this is a way to find out our GPUs are up... ThankYou

[@ticheroi]

))) you trolling or what?)

[@ggeilokowski]

@@ticheroi no, he probably isn’t

[@PhongNguyen-zz1ei]

It's kind a late for Matlab to have these features compared to popular deep learning frameworks such as Pytorch or Tensorflow. I would never use this when other frameworks are free and train models much faster and easier.

[@jonathanpuigvert7468]

Matlab is still much more like writing math, which makes it way more intuitive to use.

[@erickmaraz9753]

@@jonathanpuigvert7468 If you refer to linear algebra, i would not say so. Most deep learning framework have very robust libraries. However, for all other applications it can be very useful!

[@douwehuysmans5959]

Python + Numpy and R > Matlab

[@Chris-b-2]

Matlab is not just for deep learning; there are many other problems which can take advantage of GPGPU.

[@douwehuysmans5959]

@@Chris-b-2 all of which can be done with Python / R libraries

[@mrherpes2971]

Is this microsoft word 2000?

[@cloverlegend54]

how about providing the 2D wave equation simulation demo for two implementations

[@BraidenRobson]

Of course it only works with Nvdia GPUs

[@flamingoKnight]

cuda is an nvidia technology, what did you expect?

[@ggeilokowski]

Maybe it wouldn’t require an Nvidia GPU if AMD would have a technology thats even closely as good

[@PanosPitsi]

@Olaf Willocx nvidia has money to throw at them I am an amd user and i feel like i cant use my gpu to accelerate my deep learning work for no good reason even though it does have the power to

[@altairfoo1920]

@@PanosPitsi Every GPU has the power (OpenCL) to do so, but the lack of a sophisticated software stack makes it not an ideal platform.

[@PanosPitsi]

@@altairfoo1920 the way I see it youve let Nvidia monopolize the pro market for gpus that's why their ti gpus cost a fortune

[@fkurbaniii]

Around 2 minutes the text says paralyzed!

[@adt7058]

i wish octave can do the same TT

[@Kramer-tt32]

Matlab is great at taking open source free shared libraries, and then hiding it behind fancy easy to use toolboxes. And then charging astronomical prices for them... A fantastic business model, kudos to the mathworks team 👏

[@elise8619]

I think it's worth it, given all the time I've had to spend otherwise learning about CUDA programming. No thanks! Let me focus on my applications.

[@Kramer-tt32]

@elise8619 It's cheaper for a company to teach their employees how to write Cuda in the long run.

[@elise8619]

@@Kramer-tt32 Maybe in some cases? Depends on how into hardware/programming those employees are, how long that training takes, how many they are, and how much research/innovation time is lost.

[@Kramer-tt32]

@elise8619 always. You always have to renew the matlab subscription, but the training class is one time. It may 10 years till It becomes cheaper, but it will. Mathematically it has to

[@elise3455]

​@@Kramer-tt32Incorrect, as some licenses are life-time with the current version. Why do you feel the need to convince others to stop using software they find useful? I'm not a fan of python, but that doesn't mean I'll go onto python tutorials and tell people to stop using it.

[@jessegabriel1438]

Fantastic

[@imamjurjawi484]

good

[@magefront1485]

If only I can afford Matlab, Cuda codes are hideous!

[@kiranshila]

Imagine using nonfree software. I sure can't.

[@PanosPitsi]

@Olaf Willocx how is it a ripoff if the code is close sourced? 😂

[@PanosPitsi]

@Olaf Willocx when you say copy you mean they did it first? Then yeah by the same sense league of legegends copied chess because chess was the first strategy game. In reality, you can't claim code was copied when it was closed source, simply because there is no way to find the code in close source projects and there fore impossible to copy them.

[@annakquinn7084]

Let me guess...,no GTX support??? That is why octave and python are killing MATLAB.

[@annakquinn7084]

MATLAB unique feature is the UI. The rest is useless, Python is king.