ZLUDA

ZLUDA lets you run unmodified CUDA applications with near-native performance on AMD GPUs.

ZLUDA is currently alpha quality, but it has been confirmed to work with a variety of native CUDA applications: Geekbench, 3DF Zephyr, Blender, Reality Capture, LAMMPS, NAMD, waifu2x, OpenFOAM, Arnold (proof of concept) and more.

interesting project that sounds like it had some promise. sounds like AMD pulled funding and the developer just dumped it on github. defacto abandoned at this time until someone else decides to pickup where it left off.

but this kind of proves how superior CUDA is at the software level than OpenCL since they had performance boosts running the adapted CUDA native code vs the existing OpenCL implementations.

also sounds like it requires the code to be compiled with PTX code, which a lot of existing CUDA code from BOINC projects is not compiled this way, including GPUGRID as far as i know. so it will require some amount of effort from application developers still (unless they happen to be shipping PTX already), which has always been the biggest challenge.

---

interesting project that sounds like it had some promise. sounds like AMD pulled funding and the developer just dumped it on github. defacto abandoned at this time until someone else decides to pickup where it left off.

Indeed we hope that AMD will help this project (before or later).

but this kind of proves how superior CUDA is at the software level than OpenCL since they had performance boosts running the adapted CUDA native code vs the existing OpenCL implementations.

No, Cuda is not "superior".
If implemented correctly OpenCl/Sycl/Rocm/OneApi/ect have the same (if not better) performances of Cuda.
We are speaking about a "translator".

Cuda is better manteined by Nvidia, for sure.

I dunno. it seems clear. they saw better performance with this adapted CUDA implementation than with the native HIP one. I've not seen a single instance where OpenCL performed better, objectively, than CUDA. previous comparisons were always only possible on Nvidia, since AMD couldnt run CUDA. and CUDA was always faster on Nvidia.

now that AMD can run CUDA, even in this alpha state, CUDA ends up faster again.

---

I've not seen a single instance where OpenCL performed better, objectively, than CUDA.

https://github.com/ccsb-scripps/AutoDock-GPU/issues/239

and CUDA was always faster on Nvidia.

That's incredible!! A proprietary closed framework on the hw for which is was born is faster than a generic implementation of an open-source framework.
Who would have thought??

Ps. OpenCl could run on cpus. CUDA not.

Sounds like poor code optimization on the CUDA branch. And one of the comments on that issue suggests as much claiming the CUDA code didn’t allocate as much memory as the OpenCL variant. Not really a fair comparison.

But still you seem to be ignoring that alpha version of AMD CUDA implementation is already faster than native OpenCL and/or HIP implementations. That says a lot.

---

but this kind of proves how superior CUDA is at the software level than OpenCL since they had performance boosts running the adapted CUDA native code vs the existing OpenCL implementations.

Or how poorly AMD's driver stack is for handling OpenCL code. Not saying it's definitely the case, but feels like a reasonable possibility given how much variance of performance across different AMD OpenCL stack shows on Linux.

At the end the conclusion is same anyway. AMD hardware + software is just not as competitive for compute than Nvidia for the past few generations, even without involving CUDA. I can only hope the AI craze will force AMD to invest more in software stack and eventually bring some improvement to RDNA compute.

Meanwhile, Nvidia takes his countermeasures, blocking with CUDA 11.5 the porting of the code
https://twitter.com/never_released/status/1758946808183525702

Meanwhile, Nvidia takes his countermeasures, blocking with CUDA 11.5 the porting of the code
https://twitter.com/never_released/status/1758946808183525702

Your post seems to be missing some context. Blocking how? Your link doesn’t really elaborate. Do you mean only some notes in some TOS somewhere? And CUDA 11.5? That’s 2.5 years old.

---

Your post seems to be missing some context. Blocking how? Your link doesn’t really elaborate. Do you mean only some notes in some TOS somewhere? And CUDA 11.5? That’s 2.5 years old.

A simple research reports these words in many license agreements of Nvidia products (for example, here: https://github.com/NVIDIA/spark-rapids-container/blob/dev/NOTICE-binary)

Cuda 11.5....and all following releases.

So, bye bye Zluda and thanks for all the fish

Blocking how?

With the law/copyright/agreements/etc

By the way i don't know if this clause is legal in EU...

but again. that was more than 2 years ago. and ZLUDA is working fine with all recent CUDA releases.

this is just words in the TOS, and was probably there before this guys even ported ZLUDA to AMD, and obviously didnt stop it. it's not actively "blocking" anything.

---

zluda does work on my radeon 7900 xtx and the performance also looks good.
https://github.com/lshqqytiger/ZLUDA

https://www.gpugrid.net/result.php?resultid=37911665

Have you used any special params to get compute working?
I tried, and failed.

with different versions of zluda on that repo you noted.
https://www.gpugrid.net/results.php?userid=60896

nothing special, v3.8.7 zluda (ZLUDA-windows-rocm6-amd64.zip) + 24.12.1 drivers (latest)

Add the Zluda directory to your environment paths. This will enable BOINC to recognize it as a CUDA device and process CUDA tasks.
16.01.2025 14:34:17 | | CUDA: NVIDIA GPU 0: AMD Radeon RX 7900 XTX [ZLUDA] (driver version unknown, CUDA version 12.2, compute capability 8.8, 24560MB, 24410MB available, 30499 GFLOPS peak)
16.01.2025 14:34:17 | | OpenCL: AMD/ATI GPU 0: AMD Radeon RX 7900 XTX (driver version 3628.0 (PAL,LC), device version OpenCL 2.0 AMD-APP (3628.0), 24560MB, 24560MB available, 30499 GFLOPS peak)

I didn't test it on Linux, but since you're also on Windows, that shouldn't be the issue.

perhaps some Cuda libraries are missing? Installing https://developer.nvidia.com/cuda-downloads may help, but I don't think it should be necessary.

can you post the environmental name you are ref. to?
(or all your environmentals)

Thanks

a system variable "PATH"
https://www.computerhope.com/issues/ch000549.htm
just add the ZLUDA directory and restart the BOINC client, but if BOINC recognizes the GPU, you probably already have it right.

do you still have a custom coproc_info.xml?
http://www.gpugrid.net/forum_thread.php?id=5366&nowrap=true#61656
This might cause a problem. Zluda utilizes 8.8 as a placeholder CUDA architecture, but specifying it this way could confuse the application.

Adding zluda into path and launching boinc directly (not through zluda.exe file) seem to work.

Thanks (its running now)

ok, the wrapper doesn't share it with BOINC applications, it only makes it available to the BOINC process.

GPU: NVIDIA AMD Radeon RX 7900 XTX [ZLUDA (), [24560MB] 1 (0)

Without any doubt, this is the weirdest thing I've seen here in 10 years.