GPU Info and selection#
FAQ#
Do I need a GPU?
Technically no, you can perform limited software rendering on linux using lavapipe (see drivers link below). However
fastplotlib is intentionally built for realtime rendering using the latest GPU technologies, so we strongly
recommend that you use a GPU. Note that modern integrated graphics is often sufficient for many use-cases.
My kernel keeps crashing.
This can happen under the following circumstances:
You have ran out of GPU VRAM.
Driver issues (see next section).
If you aren’t able to solve it please post an issue on GitHub. :)
Nothing renders or rendering is weird, or I see graphical artifacts.
Probably driver issues (see next section).
Drivers#
If you notice weird graphic artifacts, things not rendering, or other glitches try updating to the latest stable drivers.
More information is also available on the WGPU docs: https://wgpu-py.readthedocs.io/en/stable/start.html#platform-requirements
Windows#
Vulkan drivers should be installed by default on Windows 11, but you will need to install your GPU manufacturer’s driver package (Nvidia or AMD). If you have an integrated GPU within your CPU, you might still need to install a driver package too, check your CPU manufacturer’s info.
Linux#
You will generally need a linux distro that is from ~2020 or newer (ex. Ubuntu 18.04 won’t work), this is due to the glibc requirements of the wgpu-native binary.
Install the drivers directly from your GPU manufacturer’s website, after that you may still need to install mesa vulkan drivers:
Debian based distros:
sudo apt install mesa-vulkan-drivers
For other distros install the corresponding vulkan driver package.
Cloud compute#
See the WGPU docs: https://wgpu-py.readthedocs.io/en/stable/start.html#cloud-compute
Mac OSX#
WGPU uses Metal instead of Vulkan on Mac. You will need at least Mac OSX 10.13. The OS should come with Metal pre-installed, so you should be good to go!
GPU Info#
View available adapters#
You can get a summary of all adapters that are available to WGPU like this:
import fastplotlib as fpl
adapters = fpl.enumerate_adapters()
for a in adapters:
print(a.summary)
For example, on a Thinkpad AMD laptop with a dedicated nvidia GPU this returns:
AMD Radeon Graphics (RADV REMBRANDT) (IntegratedGPU) on Vulkan
NVIDIA T1200 Laptop GPU (DiscreteGPU) on Vulkan
llvmpipe (LLVM 15.0.6, 256 bits) (CPU) on Vulkan
AMD Radeon Graphics (rembrandt, LLVM 15.0.6, DRM 3.52, 6.4.0-0.deb12.2-amd64) (Unknown) on OpenGL
In jupyter all the available adapters are also listed when fastplotlib is imported.
You can get more detailed info on each adapter like this:
import pprint
for a in fpl.enumerate_adapters():
pprint.pprint(a.info)
- General description of the fields:
vendor: GPU manufacturer
device: specific GPU model
description: GPU driver version
adapter_type: indicates whether this is a discrete GPU, integrated GPU, or software rendering adapter (CPU)
backend_type: one of “Vulkan”, “Metal”, or “D3D12”
For more information on the fields see: https://gpuweb.github.io/gpuweb/#gpuadapterinfo
Adapter currently in use#
If you want to know the adapter that a figure is using you can check the adapter on the renderer:
# for example if we make a plot
fig = fpl.Figure()
fig[0, 0].add_image(np.random.rand(100, 100))
fig.show()
# GPU that is currently in use by the renderer
print(fig.renderer.device.adapter.summary)
Diagnostic info#
After creating a figure you can view WGPU diagnostic info like this:
fpl.print_wgpu_report()
Example output:
██ system:
platform: Linux-5.10.0-21-amd64-x86_64-with-glibc2.31
python_implementation: CPython
python: 3.11.3
██ versions:
wgpu: 0.15.1
cffi: 1.15.1
jupyter_rfb: 0.4.2
numpy: 1.26.4
pygfx: 0.2.0
pylinalg: 0.4.1
fastplotlib: 0.1.0.a16
██ wgpu_native_info:
expected_version: 0.19.3.1
lib_version: 0.19.3.1
lib_path: ./resources/libwgpu_native-release.so
██ object_counts:
count resource_mem
Adapter: 1
BindGroup: 3
BindGroupLayout: 3
Buffer: 6 696
CanvasContext: 1
CommandBuffer: 0
CommandEncoder: 0
ComputePassEncoder: 0
ComputePipeline: 0
Device: 1
PipelineLayout: 0
QuerySet: 0
Queue: 1
RenderBundle: 0
RenderBundleEncoder: 0
RenderPassEncoder: 0
RenderPipeline: 3
Sampler: 2
ShaderModule: 3
Texture: 6 9.60M
TextureView: 6
total: 36 9.60M
██ wgpu_native_counts:
count mem backend a k r e el_size
Adapter: 1 1.98K vulkan: 1 1 3 0 1.98K
BindGroup: 3 1.10K vulkan: 3 3 0 0 368
BindGroupLayout: 3 960 vulkan: 5 3 2 0 320
Buffer: 6 1.77K vulkan: 7 6 1 0 296
CanvasContext: 0 0 0 0 0 0 160
CommandBuffer: 1 1.25K vulkan: 0 0 0 1 1.25K
ComputePipeline: 0 0 vulkan: 0 0 0 0 288
Device: 1 11.8K vulkan: 1 1 0 0 11.8K
PipelineLayout: 0 0 vulkan: 3 0 3 0 200
QuerySet: 0 0 vulkan: 0 0 0 0 80
Queue: 1 184 vulkan: 1 1 0 0 184
RenderBundle: 0 0 vulkan: 0 0 0 0 848
RenderPipeline: 3 1.68K vulkan: 3 3 0 0 560
Sampler: 2 160 vulkan: 2 2 0 0 80
ShaderModule: 3 2.40K vulkan: 3 3 0 0 800
Texture: 6 4.94K vulkan: 7 6 1 0 824
TextureView: 6 1.48K vulkan: 6 6 1 0 248
total: 36 29.7K
* The a, k, r, e are allocated, kept, released, and error, respectively.
* Reported memory does not include buffer/texture data.
██ pygfx_adapter_info:
vendor: radv
architecture:
device: AMD RADV POLARIS10 (ACO)
description: Mesa 20.3.5 (ACO)
vendor_id: 4.09K
device_id: 26.5K
adapter_type: DiscreteGPU
backend_type: Vulkan
██ pygfx_features:
adapter device
bgra8unorm-storage: - -
depth32float-stencil8: ✓ -
depth-clip-control: ✓ -
float32-filterable: ✓ ✓
indirect-first-instance: ✓ -
rg11b10ufloat-renderable: ✓ -
shader-f16: - -
texture-compression-astc: - -
texture-compression-bc: ✓ -
texture-compression-etc2: - -
timestamp-query: ✓ -
MultiDrawIndirect: ✓ -
MultiDrawIndirectCount: ✓ -
PushConstants: ✓ -
TextureAdapterSpecificFormatFeatures: ✓ -
VertexWritableStorage: ✓ -
██ pygfx_limits:
adapter device
max-bind-groups: 8 8
max-bind-groups-plus-vertex-buffers: 0 0
max-bindings-per-bind-group: 1.00K 1.00K
max-buffer-size: 2.14G 2.14G
max-color-attachment-bytes-per-sample: 0 0
max-color-attachments: 0 0
max-compute-invocations-per-workgroup: 1.02K 1.02K
max-compute-workgroup-size-x: 1.02K 1.02K
max-compute-workgroup-size-y: 1.02K 1.02K
max-compute-workgroup-size-z: 1.02K 1.02K
max-compute-workgroup-storage-size: 32.7K 32.7K
max-compute-workgroups-per-dimension: 65.5K 65.5K
max-dynamic-storage-buffers-per-pipeline-layout: 8 8
max-dynamic-uniform-buffers-per-pipeline-layout: 16 16
max-inter-stage-shader-components: 128 128
max-inter-stage-shader-variables: 0 0
max-sampled-textures-per-shader-stage: 8.38M 8.38M
max-samplers-per-shader-stage: 8.38M 8.38M
max-storage-buffer-binding-size: 2.14G 2.14G
max-storage-buffers-per-shader-stage: 8.38M 8.38M
max-storage-textures-per-shader-stage: 8.38M 8.38M
max-texture-array-layers: 2.04K 2.04K
max-texture-dimension-1d: 16.3K 16.3K
max-texture-dimension-2d: 16.3K 16.3K
max-texture-dimension-3d: 2.04K 2.04K
max-uniform-buffer-binding-size: 2.14G 2.14G
max-uniform-buffers-per-shader-stage: 8.38M 8.38M
max-vertex-attributes: 32 32
max-vertex-buffer-array-stride: 2.04K 2.04K
max-vertex-buffers: 16 16
min-storage-buffer-offset-alignment: 32 32
min-uniform-buffer-offset-alignment: 32 32
██ pygfx_caches:
count hits misses
full_quad_objects: 1 0 2
mipmap_pipelines: 0 0 0
layouts: 1 0 3
bindings: 1 0 1
shader_modules: 2 0 2
pipelines: 2 0 2
shadow_pipelines: 0 0 0
██ pygfx_resources:
Texture: 8
Buffer: 23
Select GPU (adapter)#
You can select an adapter by passing one of the wgpu.GPUAdapter instances returned by fpl.enumerate_adapters()
to fpl.select_adapter():
# get info or summary of all adapters to pick an adapter
import pprint
for a in fpl.enumerate_adapters():
pprint.pprint(a.info)
# example, pick adapter at index 2
chosen_gpu = fpl.enumerate_adapters()[2]
fpl.select_adapter(chosen_gpu)
You must select an adapter before creating a Figure , otherwise the default adapter will be selected. Once a
Figure is created the adapter cannot be changed.
Note that using this function reduces the portability of your code, because it’s highly specific for your current machine/environment.
The order of the adapters returned by fpl.enumerate_adapters() is
such that Vulkan adapters go first, then Metal, then D3D12, then OpenGL.
Within each category, the order as provided by the particular backend is
maintained. Note that the same device may be present via multiple backends
(e.g. vulkan/opengl).
We cannot make guarantees about whether the order of the adapters matches
the order as reported by e.g. nvidia-smi. We have found that on a Linux
multi-gpu cluster, the order does match, but we cannot promise that this is
always the case. If you want to make sure, do some testing by allocating big
buffers and checking memory usage using nvidia-smi
Example to allocate and check GPU mem usage:
import subprocess
import wgpu
import torch
def allocate_gpu_mem_with_wgpu(idx):
a = wgpu.gpu.enumerate_adapters()[idx]
d = a.request_device()
b = d.create_buffer(size=10*2**20, usage=wgpu.BufferUsage.COPY_DST)
return b
def allocate_gpu_mem_with_torch(idx):
d = torch.device(f"cuda:{idx}")
return torch.ones([2000, 10], dtype=torch.float32, device=d)
def show_mem_usage():
print(subprocess.run(["nvidia-smi"]))
See pygfx/wgpu-py#482 for more details.