If I'm looking to play around with some python code that needs a GPU and will not run on a CPU only machine, what's a good laptop I can buy? I keep running into CUDA not installed errors and I can't install CUDA.
Why do you want a laptop? You'll be paying for worse performance and less capable cooling. If you already have a primary laptop you use, it would be cheaper and higher-performance to build a CUDA-oriented homeserver and SSH into that.
A link to a homeserver? You'd probably end up building it yourself, unless you want to pay a premium for a prebuilt.
A good LLM build maximizes the amount of VRAM you're getting for AI workloads. Shop around for a 3090 24gb if that's within your price range, and then build backwards from there. Cheaper options include the 3060 12gb and the various other 16-20gb CUDA-enabled cards. Then find a motherboard that maxes out your card's PCI bandwidth and select a budget CPU that fits the socket and won't bottleneck your GPU. Add your memory, storage and install an OS. Then you're done.
If you want to be on the cutting edge (eg. HEDT, ECC, 14th gen/Zen 4 CPUs, etc.) then there aren't many integrators that will be able to bundle all the newest tech. If you're looking to maximize the performance-per-dollar and efficency for your usecase, self-built is really the only option.
If you still insist on paying next-gen prices for last-gen hardware, Lenovo and Dell have you covered:
CUDA is specific to Nvidia GPUs. So you'll need a computer (laptop, in your case) with an Nvidia video card which supports CUDA.
Does the Python code you want to run use PyTorch? Is it related to LLMs or Stable Diffusion? Because those should also run on CPU alone. Not very fast, but good enough to play around with.
If you're thinking of renting a machine from AWS, I'm currently using a "g4dn.xlarge" instance which does support CUDA and can run Stable Diffusion comfortably. It's a bit pricey though and I had to wait and get approved before I could rent it.
One option is to go for Dell's Alienware line, depending on your needs you need to choose how much RAM you need on the GPU card. RTX 4000 generation seems to have at minimum 8 GB, which is sufficient for most problems for now.
10 comments
[ 0.30 ms ] story [ 36.4 ms ] threadA good LLM build maximizes the amount of VRAM you're getting for AI workloads. Shop around for a 3090 24gb if that's within your price range, and then build backwards from there. Cheaper options include the 3060 12gb and the various other 16-20gb CUDA-enabled cards. Then find a motherboard that maxes out your card's PCI bandwidth and select a budget CPU that fits the socket and won't bottleneck your GPU. Add your memory, storage and install an OS. Then you're done.
If you still insist on paying next-gen prices for last-gen hardware, Lenovo and Dell have you covered:
https://www.lenovo.com/us/en/p/workstations/thinkstation-p-s...
https://www.dell.com/en-us/shop/scc/sr/workstations/precisio...
Does the Python code you want to run use PyTorch? Is it related to LLMs or Stable Diffusion? Because those should also run on CPU alone. Not very fast, but good enough to play around with.
If you're thinking of renting a machine from AWS, I'm currently using a "g4dn.xlarge" instance which does support CUDA and can run Stable Diffusion comfortably. It's a bit pricey though and I had to wait and get approved before I could rent it.