Dotan Levi
Research Subjects
Research Interests
Dotan Levi has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- Dror Porat
NVIDIA, Yokneam, Israel
2 shared publications - Eyal Frishman
Nvidia, Raanana, Israel
2 shared publications - Jant Chen
Nvidia, Shanghai, China
2 shared publications - Amit Hermony
Nvidia Corp
1 shared publication - Dimitris Syrivelis
University of Thessaly
1 shared publication - Elad Mentovich
Nvidia Corp
1 shared publication
Latest Publications
- Open Access
- 24 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 44 (3) , 1-35
FlexDriver: A Network Driver for Your Accelerator
We propose a new system design for connecting hardware and FPGA accelerators to the network, allowing them to directly control commodity ASIC NICs without using the CPU. This solves the key challenge of leveraging existing NIC hardware offloads such as RDMA and virtualization for hardware disaggregation and accelerator networking. Our approach supports a diverse set of use cases, from direct network access for disaggregated accelerators to inline acceleration of the network stack and disaggregation of main system memory, all without implementing complex networking logic. To demonstrate this approach, we build FlexDriver (FLD) , a hardware module that implements a NIC data-plane driver. Our main technical contribution is compressing NIC control structures by 5 ×, allowing FLD to achieve high scalability with low die area and no memory bandwidth interference. We build two prototypes – FLD core on NVIDIA Innova-2 FPGA SmartNICs and FLD with a load/store interface on IBM OpenCAPI FPGA deployment with ConnectX-6 Dx. We demonstrate four different use cases: a disaggregated LTE cipher, an IP-reassembly inline accelerator, an IoT cryptographic-token authentication offload, and a fine-grained memory disaggregation datapath over RDMA. These leverage the ASIC NIC for RDMA processing, VXLAN tunneling, and traffic shaping, without CPU involvement.
- 24 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 44 (3) , 1-35
Open AccessFlexDriver: A Network Driver for Your Accelerator
We propose a new system design for connecting hardware and FPGA accelerators to the network, allowing them to directly control commodity ASIC NICs without using the CPU. This solves the key challenge of leveraging existing NIC hardware offloads such as RDMA and virtualization for hardware disaggregation and accelerator networking. Our approach supports a diverse set of use cases, from direct network access for disaggregated accelerators to inline acceleration of the network stack and disaggregation of main system memory, all without implementing complex networking logic. To demonstrate this approach, we build FlexDriver (FLD) , a hardware module that implements a NIC data-plane driver. Our main technical contribution is compressing NIC control structures by 5 ×, allowing FLD to achieve high scalability with low die area and no memory bandwidth interference. We build two prototypes – FLD core on NVIDIA Innova-2 FPGA SmartNICs and FLD with a load/store interface on IBM OpenCAPI FPGA deployment with ConnectX-6 Dx. We demonstrate four different use cases: a disaggregated LTE cipher, an IP-reassembly inline accelerator, an IoT cryptographic-token authentication offload, and a fine-grained memory disaggregation datapath over RDMA. These leverage the ASIC NIC for RDMA processing, VXLAN tunneling, and traffic shaping, without CPU involvement.
- Open Access
- 2 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 42-43
Learned Image and Video Compression: Foundations, Algorithms, and Open Challenges
- 2 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 42-43
Open AccessLearned Image and Video Compression: Foundations, Algorithms, and Open Challenges
- Open Access
- 18 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 79-80
Harnessing Hardware-Generated Statistics For CU-Tree Decision Making
- 18 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 79-80
Open AccessHarnessing Hardware-Generated Statistics For CU-Tree Decision Making
- Open Access
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 131-132
Accelerated and Adaptive Non-local Mean Temporal Filtering for Improved HEVC Coding Efficiency
As modern video codecs exploit at their basis temporal and spatial correlations in video signal for compression, and such correlations are corrupted by noise that is present in the video signal, temporal filtering has been shown to be effective for improving coding efficiency in various encoder implementations [1], [2], [3]. In particular, the non-local mean approach in [1] has been applied in AV1, an open video coding format competing with HEVC (aka H.265), which includes built-in mechanisms that allow effective temporal filtering using non-displayable reference frames (aka ALTREF frames). NVIDIA NVENC (short for NVIDIA Encoder) is a hardware video encoding engine in NVIDIA graphics cards, aimed at offloading the compute-intensive task of video encoding from the CPU to a dedicated part of the GPU. Via the NVIDIA Video Codec SDK, a comprehensive set of APIs for hardware-accelerated video encode and decode are made available for several popular codecs, on both Windows and Linux. In this work, we propose an accelerated and adaptive nonlocal mean temporal filter for improved HEVC coding efficiency, which has been implemented in NVIDIA NVENC. Compared to prior work, our approach is novel and distinctive in three main aspects: (1) it applies a non-local mean temporal filter as a pre-processing step in HEVC, which unlike AV1 has no built-in support for temporal filtering, (2) it uses a novel locally-adaptive noise estimation procedure for block-level adaptation of filtering strength, and (3) it is implemented as an efficient CUDA kernel to minimize any performance impact on the NVENC encoder and reduce latency. NVENC already offers state-of-the-art coding efficiency (i.e., quality vs. rate tradeoff) in many settings, and with an upcoming release of NVIDIA Video Codec SDK will also offer significantly improved HEVC coding efficiency using an accelerated and adaptive non-local mean temporal filter implementation. A comprehensive evaluation of the temporal filter on both internal (NVIDIA proprietary) and external benchmarks demonstrated significant coding efficiency improvements for HEVC encoding. For example, the average compression gains achieved (as measured by PSNR BD-Rate, i.e., average bitrate savings for the same PSNR reconstruction quality) on the objective-1-fast standard benchmark dataset from Xiph.org [4] are approximately 7% for 720p sequences, and 6% for 1080p sequences. In terms of computational performance, our efficient, optimized temporal filter CUDA kernel is roughly 4000x faster than a (single-threaded) CPU implementation, reducing the filter processing time from seconds on a CPU to sub-millisecond on an NVIDIA RTX 4090 GPU (for a 1080p frame).
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 131-132
Open AccessAccelerated and Adaptive Non-local Mean Temporal Filtering for Improved HEVC Coding Efficiency
As modern video codecs exploit at their basis temporal and spatial correlations in video signal for compression, and such correlations are corrupted by noise that is present in the video signal, temporal filtering has been shown to be effective for improving coding efficiency in various encoder implementations [1], [2], [3]. In particular, the non-local mean approach in [1] has been applied in AV1, an open video coding format competing with HEVC (aka H.265), which includes built-in mechanisms that allow effective temporal filtering using non-displayable reference frames (aka ALTREF frames). NVIDIA NVENC (short for NVIDIA Encoder) is a hardware video encoding engine in NVIDIA graphics cards, aimed at offloading the compute-intensive task of video encoding from the CPU to a dedicated part of the GPU. Via the NVIDIA Video Codec SDK, a comprehensive set of APIs for hardware-accelerated video encode and decode are made available for several popular codecs, on both Windows and Linux. In this work, we propose an accelerated and adaptive nonlocal mean temporal filter for improved HEVC coding efficiency, which has been implemented in NVIDIA NVENC. Compared to prior work, our approach is novel and distinctive in three main aspects: (1) it applies a non-local mean temporal filter as a pre-processing step in HEVC, which unlike AV1 has no built-in support for temporal filtering, (2) it uses a novel locally-adaptive noise estimation procedure for block-level adaptation of filtering strength, and (3) it is implemented as an efficient CUDA kernel to minimize any performance impact on the NVENC encoder and reduce latency. NVENC already offers state-of-the-art coding efficiency (i.e., quality vs. rate tradeoff) in many settings, and with an upcoming release of NVIDIA Video Codec SDK will also offer significantly improved HEVC coding efficiency using an accelerated and adaptive non-local mean temporal filter implementation. A comprehensive evaluation of the temporal filter on both internal (NVIDIA proprietary) and external benchmarks demonstrated significant coding efficiency improvements for HEVC encoding. For example, the average compression gains achieved (as measured by PSNR BD-Rate, i.e., average bitrate savings for the same PSNR reconstruction quality) on the objective-1-fast standard benchmark dataset from Xiph.org [4] are approximately 7% for 720p sequences, and 6% for 1080p sequences. In terms of computational performance, our efficient, optimized temporal filter CUDA kernel is roughly 4000x faster than a (single-threaded) CPU implementation, reducing the filter processing time from seconds on a CPU to sub-millisecond on an NVIDIA RTX 4090 GPU (for a 1080p frame).
- Open Access
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 107-108
H/W Accelerated Unsupervised Codec Modernization
Following a decade where AVC/H.264 was the dominant codec in many video applications, the last years have finally seen more modern video codecs gain traction. This raises the need for cost effective, fast, fully automatic transcoding into modern codec formats which will preserve the input perceptual quality, at the lowest possible bitrate. The availability of such a solution makes the switchover to newer encoders a smooth, easy, and low-cost process accelerating adoption of HEVC, AVI or VVC.
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 107-108
Open AccessH/W Accelerated Unsupervised Codec Modernization
Following a decade where AVC/H.264 was the dominant codec in many video applications, the last years have finally seen more modern video codecs gain traction. This raises the need for cost effective, fast, fully automatic transcoding into modern codec formats which will preserve the input perceptual quality, at the lowest possible bitrate. The availability of such a solution makes the switchover to newer encoders a smooth, easy, and low-cost process accelerating adoption of HEVC, AVI or VVC.