3 ms·

> If only GPUs were more common They are very common. In 2021 on Windows, is safe to assume the GPU does at least D3D 11.0. The last GPU which does not is Int

> If only GPUs were more common They are very common. In 2021 on Windows, is safe to assume the GPU does at least D3D 11.0. The last GPU which does not is Intel Sandy Bridge, discontinued in 2013. D3D 11.0 supports lots of features about GPGPU. The main issue with D3D11 is FP64 support being optional, the hardware may or may not support. > and easier to work with I don’t find neither DirectCompute 11, nor CUDA, exceptionally hard to work with. D3D12 and Vulkan are indeed hard, IMO.

amelius 5y ago
> I don’t find neither DirectCompute 11, nor CUDA, exceptionally hard to work with. It still sucks that, if you're not careful or unlucky, GPUs may interfere with normal video operation during setup. Also on one machine I'm getting "GPU lost" error messages every once in a while during a long computation, and I have no other option than to reboot my machine. Further, the form factor sucks. Every card comes with 4 or more DVI connectors which I don't need. If parallel is the future, then give me something that fits on the mainboard directly, and which is more or less a commodity. Not the proprietary crap that vendors are currently shoving at us.
- Const-me 5y ago
  > on one machine I'm getting "GPU lost" error messages every once in a while during a long computation, and I have no other option than to reboot my machine. If that’s your computer, just change the registry setting disabling the TDR. By default, Windows uses 2 seconds to limit max.pipeline latency. When exceeded, the OS resets the GPU, restarts the driver, and logs that “device lost” in the event log. If that’s your customer’s computer you have to be more creative and fix your compute shaders. When you have a lot of things to compute and/or you detect underpowered hardware, split your Dispatch() calls into a series of smaller ones. As a nice side effect, this minimizes effect of your application on 3D rendering things running on the same GPU by other programs. Don’t submit them all at once. Submit two initially then one at a time using ID3D11Query to track completion of these compute shaders. ID3D11Fence can do that better (can sleep on WaitForSingleObject saving electricity) but fences are less compatible unfortunately. That’s an optional feature introduced in Win10 post-release in some update, and e.g. VMWare SVGA 3D doesn’t support fences. > the form factor sucks. Every card comes with 4 or more DVI connectors which I don't need. Some datacenter GPUs, and newer mining GPUs come without connectors. The mainstream commodity cards to have connectors because the primary market for them is supposed to be gamers. > something that fits on the mainboard directly Unlikely to happen because thermals. Modern high-end GPUs consume more electricity than comparable CPUs, e.g. RTX 3090 needs 350W, similarly priced EPYC 7453 needs 225W.
  - amelius 5y ago
    Thanks for the advice (I'm on Linux), but my point is that the GPU really is not the "first-class citizen" that the CPU is.
    - Const-me 5y ago
      > my point is that the GPU really is not the "first-class citizen" that the CPU is. On Windows GPU is the first-class citizen since Vista. In Vista, Microsoft started to use D3D10 for their desktop compositor. In Windows 7 they have upgraded to Direct3D 11. The transition wasn’t smooth. Only gamers had 3D GPUs before Vista, many people needed new computers. Technically, Microsoft had to change driver model to support a few required features. On the bright side, now that XP->Win7 transition is long in the past, and 3D GPUs are used for everything on Windows. All web browsers are using D3D to render stuff, albeit not directly, through the higher-level libraries like Direct2D and DirectWrite. Linux doesn’t even have these higher-level libraries. They are possible to implement on top of whichever GPU API is available https://github.com/Const-me/Vrmac#vector-graphics-engine https://github.com/Const-me/Vrmac#vector-graphics-engine but so far nobody did it well enough. It’s very similar situation with GPU compute on Linux. The kernel and driver support has arrived by now, but the higher-level user mode things are still missing. P.S. If you’re on Linux, pretty sure you can still refactor your compute shaders the same way and they will work fine afterwards. You obviously don’t have ID3D11Query/ID3D11Fence but most GPU APIs have replacements: VkFence(3) in Vulkan, EGL_KHR_fence_sync/GL_OES_EGL_sync/VG_KHR_EGL_sync extensions for GL and GLES, etc.
- theandrewbailey 5y ago
  > Further, the form factor sucks. Every card comes with 4 or more DVI connectors which I don't need. DVI connectors are huge! I've never seen a card with more than 2 of them. You working with something more exotic than a PCI-e card?

amelius 5y ago

> I don’t find neither DirectCompute 11, nor CUDA, exceptionally hard to work with. It still sucks that, if you're not careful or unlucky, GPUs may interfere with normal video operation during setup. Also on one machine I'm getting "GPU lost" error messages every once in a while during a long computation, and I have no other option than to reboot my machine. Further, the form factor sucks. Every card comes with 4 or more DVI connectors which I don't need. If parallel is the future, then give me something that fits on the mainboard directly, and which is more or less a commodity. Not the proprietary crap that vendors are currently shoving at us.

Const-me 5y ago
> on one machine I'm getting "GPU lost" error messages every once in a while during a long computation, and I have no other option than to reboot my machine. If that’s your computer, just change the registry setting disabling the TDR. By default, Windows uses 2 seconds to limit max.pipeline latency. When exceeded, the OS resets the GPU, restarts the driver, and logs that “device lost” in the event log. If that’s your customer’s computer you have to be more creative and fix your compute shaders. When you have a lot of things to compute and/or you detect underpowered hardware, split your Dispatch() calls into a series of smaller ones. As a nice side effect, this minimizes effect of your application on 3D rendering things running on the same GPU by other programs. Don’t submit them all at once. Submit two initially then one at a time using ID3D11Query to track completion of these compute shaders. ID3D11Fence can do that better (can sleep on WaitForSingleObject saving electricity) but fences are less compatible unfortunately. That’s an optional feature introduced in Win10 post-release in some update, and e.g. VMWare SVGA 3D doesn’t support fences. > the form factor sucks. Every card comes with 4 or more DVI connectors which I don't need. Some datacenter GPUs, and newer mining GPUs come without connectors. The mainstream commodity cards to have connectors because the primary market for them is supposed to be gamers. > something that fits on the mainboard directly Unlikely to happen because thermals. Modern high-end GPUs consume more electricity than comparable CPUs, e.g. RTX 3090 needs 350W, similarly priced EPYC 7453 needs 225W.
- amelius 5y ago
  Thanks for the advice (I'm on Linux), but my point is that the GPU really is not the "first-class citizen" that the CPU is.
  - Const-me 5y ago
    > my point is that the GPU really is not the "first-class citizen" that the CPU is. On Windows GPU is the first-class citizen since Vista. In Vista, Microsoft started to use D3D10 for their desktop compositor. In Windows 7 they have upgraded to Direct3D 11. The transition wasn’t smooth. Only gamers had 3D GPUs before Vista, many people needed new computers. Technically, Microsoft had to change driver model to support a few required features. On the bright side, now that XP->Win7 transition is long in the past, and 3D GPUs are used for everything on Windows. All web browsers are using D3D to render stuff, albeit not directly, through the higher-level libraries like Direct2D and DirectWrite. Linux doesn’t even have these higher-level libraries. They are possible to implement on top of whichever GPU API is available https://github.com/Const-me/Vrmac#vector-graphics-engine https://github.com/Const-me/Vrmac#vector-graphics-engine but so far nobody did it well enough. It’s very similar situation with GPU compute on Linux. The kernel and driver support has arrived by now, but the higher-level user mode things are still missing. P.S. If you’re on Linux, pretty sure you can still refactor your compute shaders the same way and they will work fine afterwards. You obviously don’t have ID3D11Query/ID3D11Fence but most GPU APIs have replacements: VkFence(3) in Vulkan, EGL_KHR_fence_sync/GL_OES_EGL_sync/VG_KHR_EGL_sync extensions for GL and GLES, etc.
theandrewbailey 5y ago
> Further, the form factor sucks. Every card comes with 4 or more DVI connectors which I don't need. DVI connectors are huge! I've never seen a card with more than 2 of them. You working with something more exotic than a PCI-e card?