Home Tech Intel Core Display Can Turn on NVIDIA CUDA Performance Running Score 52%...

Intel Core Display Can Turn on NVIDIA CUDA Performance Running Score 52% Higher

ZLUDA Geekbench Test

SCH.cxIntel core display can turn on NVIDIA CUDA acceleration running score 52% higher. NVIDIA CUDA is famous, but from the beginning, this technology has been exclusive to the N card. Although there are some tools that allow CUDA to be used in the OpenCL environment, even advanced such as HIPCL is still a semi-automated tool that requires manual intervention by the developer.



The good news is that a new tool named ZLUDA was born. Simply put, systems with Intel core graphics (6th generation Core Skylake and above) can still use CUDA acceleration.

ZLUDA Geekbench Test
ZLUDA Geekbench Test

Although it is said that efficiency and localization are almost the same, it seems to be used only well on Geekbench. The official Linux-based test shows that the UHD 630 core display (Core i5-8700K) can run to 6333 after CUDA is turned on. Although the score is lower than the native OpenCL 6482, the CUDA acceleration still plays a big role in some test items. The buff bonus has a maximum range of 52%. Unfortunately, after some media tried to download ZLUDA from the GitHub page, and operated on the Tiger Lake notebook, Geekbench 5 reported an error…


ZLUDA is a drop-in replacement for CUDA on Intel GPU. ZLUDA allows to run unmodified CUDA applications using Intel GPUs with near-native performance (more below). It works with current integrated Intel UHD GPUs and will work with future Intel Xe GPUs


ZLUDA performance has been measured with GeekBench 5.2.3 on Intel UHD 630.
One measurement has been done using OpenCL and another measurement has been done using CUDA with Intel GPU masquerading as a (relatively slow) NVIDIA GPU with the help of ZLUDA. Both measurements use the same GPU.

The performance below is normalized to OpenCL performance. 110% means that ZLUDA-implemented CUDA is 10% faster on Intel UHD 630.

Performance graph

ZLUDA – detailed results on Geekbench.com

OpenCL – detailed results on Geekbench.com

Overall, ZLUDA is slower in GeekBench by roughly 2%.

Explanation of the results

  • Why is ZLUDA faster in some benchmarks?
    This has not been precisely pinpointed to one thing or another but it’s likely a combination of things:
    • ZLUDA uses Level 0, which in general is a more low level, high-performance API than OpenCL
    • Tying to the previous point, currently, ZLUDA does not support asynchronous execution. This gives us an unfair advantage in a benchmark like GeekBench. GeekBench exclusively uses CUDA synchronous APIs
    • There is a set of GPU instructions that are available on both NVIDIA hardware and Intel hardware but are not exposed through OpenCL. We are comparing NVIDIA GPU optimized code with the more general OpenCL code. It’s a lucky coincidence (and a credit to the underlying Intel Graphics Compiler) that this code also works well on an Intel GPU
  • Why is OpenCL faster in Canny and Horizon Detection?
    Authors of CUDA benchmarks used CUDA functions atomicInc and atomicDec which have direct hardware support on NVIDIA cards, but no hardware support on Intel cards. They have to be emulated in software, which limits performance
  • Why is ZLUDA slower in the remaining benchmarks?
    The reason is unknown. Most likely, in some tests, we compile from suboptimal NVIDIA GPU code and in other tests, ZLUDA itself is emitting suboptimal Intel GPU code. For example, SFFT used to be even slower before PR #22


  • Is ZLUDA a drop-in replacement for CUDA?
    Yes, but certain applications use CUDA in ways that make it incompatible with ZLUDA
  • What is the status of the project?
    This project is a Proof of Concept. About the only thing that works currently is Geekbench. It’s amazingly buggy and incomplete. You should not rely on it for anything serious
  • Is it an Intel project? Is it an NVIDIA project?
    No, it’s a private project
  • What is the performance?
    Performance can be close to the performance of similarly written OpenCL code (see GeekBench results in the previous section). NVIDIA GPUs and Intel GPUs have different architecture and feature sets. Consequently, certain NVIDIA features have to be emulated in ZLUDA with a performance penalty. Additionally, the performance of ZLUDA will be always lower than the performance of code specifically optimized for Intel GPUs
  • How it’s different from AMD HIP or Intel DPC++ Compatibility toolkit?
    Both are porting toolkits that require the programmer’s effort to port applications to the API in question. With ZLUDA existing applications “just work” on an Intel GPU (if you are lucky and ZLUDA supports the particular subset of CUDA)
  • Which Intel GPU are supported?
    Intel Gen9 and newer (Skylake and newer) which are supported by Intel Level 0
  • Does ZLUDA support AMD GPUs?
    Certainly not currently, but it might be technically possible


Warning: this is a very incomplete proof of concept. It’s probably not going to work with your application. ZLUDA currently works only with applications that use CUDA Driver API or statically-linked CUDA Runtime API – dynamically-linked CUDA Runtime API is not supported at all


You should have the most recent Intel GPU drivers installed.
Run your application like this:



You should install the most recent run-time driver packages as outlined here: https://dgpu-docs.intel.com/installation-guides/index.html.
Run your application like this:



You should have a relatively recent version of Rust installed, then you just do:

cargo build --release

in the main directory of the project.


You should install the most recent run-time and developer driver packages as outlined here: https://dgpu-docs.intel.com/installation-guides/index.html. Additionally, you should have ocl-icd-opencl-dev (or equivalent) installed.
If you are building on Linux you must also symlink (or rename) the ZLUDA output binaries after ZLUDA build finishes:

ln -s libnvcuda.so target/release/libcuda.so
ln -s libcuda.so target/release/libcuda.so.1


If you want to develop ZLUDA itself, read CONTRIBUTING.md, it contains instructions how to set up dependencies and run tests


This software is dual-licensed under either the Apache 2.0 license or the MIT license. See LICENSE-APACHE or LICENSE-MIT for details

Previous articleThe Fifth Weekly Release Candidate Version of Linux 5.11 Begins Testing
Next articleIntel’s February Wireless Driver Update Fixes Big Issues on Windows 10


Please enter your comment!
Please enter your name here