Contents
Intel's media accelerators featuring Intel® Quick Sync Video transcode optimize both throughput and visual quality for video cloud distribution. Three classes of accelerator engines target high density video decode, encode, and video processing. These unlock real-time customers needing:
- File based, Just-in-time, and Live consumption and creation
- Bandwidth-efficient media delivery for Adaptive bitrate streaming, user generated content upload, and high visual quality broadcast
Intel's discrete graphics accelerators are well integrated into open source media frameworks such as FFmpeg and Intel® oneAPI Video Processing Library. These popular frameworks allow both complex pipeline support as well as extreme customization of accelerator control. To make these tools even more accessible to Linux developers we're now providing build scripts in Docker on the latest Linux kernels.
Intel® Data Center GPU Flex Series (products formerly Arctic Sound) comes in two flavors:
These accelerators have 4 classes of video accelerator engines:
- 2 engines (VDBOX) accelerate video decode and encode
- 2 video enhancement engines (VEBOX) and two Scale and Format Converters (SFC) accelerate video scaling, color space conversion, denoise, deinterlace and more
- 1 render engine provides distributed execution units combined with media samplers
- 1 (or 4 for ATS-M150) compute engines allows parallel access to distributed execution units to accelerate AI and intensive compute workloads
The combination of 4 accelerator types ensures typical transcode operations can pipeline execution to minimize latency. Multiple accelerator units allow concurrent execution of multiple frames to maximize throughput.
Caption: Picture provides brief overview of Intel® Data Center GPU Flex 140 (ATS-M75) and Intel® Data Center GPU Flex 170 (ATS-M150) SoC. Other Adapters might have different architecture and features.
| Intel® Data Center GPU Flex 140 | Intel® Data Center GPU Flex 170 | |
|---|---|---|
|
|
|
| Open Source Reference | ATS-M75 | ATS-M150 |
| SoC number | 2 | 1 |
| Target Workloads | Media processing and delivery, Windows and Android cloud gaming, virtualized desktop infrastructure, AI visual inference | |
| Card Form Factor | Half height, half length, single wide, passive cooling | Full height, three-quarter lenfth, single wide, passive cooling |
| Card TDP | 75 watts | 150 watts |
| GPUs per Card | 2 | 1 |
| GPU Microarchitecture | Xe-HPG | |
| Xe Cores | 16 (8 per GPU) | 32 |
| Fixed Function Media | 4 (2 per GPU) | 2 |
| Ray Tracing | Yes | |
| Peak Compute (Systolic) | 8 TFLOPS (FP32)/105 TOPS (INT8) | 16 TFLOPS (FP32)/250 TOPS (INT8) |
| Memory Type | GDDR6 | |
| Memory Capacity | 12 GB (6 per GPU) | 16 GB |
| Virtualization (Instances) | SR-IOV (62) | SR-IOV (31) |
| Host Bus | PCIe Gen 4 | |
| Host CPU Support | 3rd Generation Intel® Xeon® Scalable Processors | |
For more details refer to the following materials:
- Intel® Data Center GPU Flex Series - Overview
- Intel® Data Center GPU Flex Series - Product Brief
- Intel® Data Center GPU Flex Series - Media Processing & Delivery Solution Brief
Intel® Data Center GPU Flex Series comes with the following exciting new features for media:
- AV1 Hardware Encoding support
- Advanced software bitrate controller to boost hardware encoding quality
Below we give brief overview of Intel® Data Center GPU Flex Series Graphics Adapter’s Media capabilities. Other Adapters might have different features. You can find more information about supported media features in media-driver documentation.
| Fixed Function | ATS-M75 | ATS-M150 |
|---|---|---|
| VEBOX | 4 (2 per SoC) | 2 (2 per SoC) |
| VDBOX | 4 (2 per SoC) | 2 (2 per SoC) |
| SFC | 4 (2 per SoC) | 2 (2 per SoC) |
| VME | 0 | 0 |
| Codec | Sampling Depth | BPP | Format | Max Resolution |
|---|---|---|---|---|
| AVC | 4:2:0 | 8 | Progressive | 4K |
| HEVC | 4:2:0 | 8, 10 | Progressive | 8K |
| HEVC | 4:4:4 | 8, 10 | Progressive | 5K |
| VP9 | 4:2:0 | 8, 10 | Progressive | 8K |
| VP9 | 4:4:4 | 8, 10 | Progressive | 5K |
| AV1 | 4:2:0 | 8, 10 | Progressive | 8K |
| JPEG | 4:2:0 | 8 | Progressive | |
| JPEG | 4:2:2 | 8 | Progressive | |
| JPEG | 4:4:4 | 8 | Progressive |
| Codec | Sampling Depth | BPP | Format | Max Resolution |
|---|---|---|---|---|
| MPEG2 | 4:2:0 | 8 | Progressive, Interlaced | FHD |
| AVC | 4:2:0 | 8 | Progressive, Interlaced | 4K |
| HEVC | 4:2:0 | 8, 10, 12 | Progressive | 8K |
| HEVC | 4:2:2 | 8, 10, 12 | Progressive | 8K |
| HEVC | 4:4:4 | 8, 10, 12 | Progressive | 5K |
| VP9 | 4:2:0 | 8, 10, 12 | Progressive | 8K |
| VP9 | 4:4:4 | 8, 10, 12 | Progressive | 5K |
| AV1 | 4:2:0 | 8, 10 | Progressive | 8K |
| JPEG | 4:2:0 | 8 | Progressive | |
| JPEG | 4:2:2 | 8 | Progressive | |
| JPEG | 4:4:4 | 8 | Progressive |
Each Intel® Data Center GPU Flex Series graphics die achieves the highest levels of performance for the modern generation video standards like HEVC and AV1, while still supporting ultra-high density and high quality transcode.
When using high level API’s like FFMPEG, we provide three convenient operating presets that offer different tradeoffs between speed and quality (many additional controls are available for developers use). For more details check Video Performance Command Linux and Measuring Methodology.
See key platform capabilities highlight below:
|
|
|
We evaluate performance on the following streams:
- https://repositories.intel.com/media/bbb_sunflower_1080p_60fps_4Mbps_38kframes.h264
- https://repositories.intel.com/media/bbb_sunflower_1080p_60fps_4Mbps_38kframes.h265
- https://repositories.intel.com/media/bbb_sunflower_1080p_60fps_4Mbps_38kframes_av1.ivf
- https://repositories.intel.com/media/bbb_sunflower_2160p_60fps_8Mbps_38kframes.h264
- https://repositories.intel.com/media/bbb_sunflower_2160p_60fps_8Mbps_38kframes.h265
- https://repositories.intel.com/media/bbb_sunflower_2160p_60fps_8Mbps_38kframes_av1.ivf
See the following files for attributions:
To verify downloads, use:
Intel® Data Center GPU Flex Series offers significant HEVC and AVC encode improvements over the previous generation of hardware encoders. When compared to typical presets on popular software video encoders (x264* and x265*) Intel® Data Center GPU Flex Series provides acceleration at similar quality level. AV1 encoding offers bandwidth and bitrate savings if 30% over AVC in Low Delay.
AV1 Bandwidth or bitrate savings of 30% over AVC in low delay encoding
The graphs (below) illustrate the video bitrate savings of Intel® Data Center GPU Flex Series graphics compared with the most common presets.
For more details about testing methodology in case of Random Access encoding, check Video Quality Command Lines and Measuring Methodology.
Caption: These charts illustrate quality assessment of our Intel® Data Center GPU Flex Series encoders (in random access use case) expressed as a percent of bitrate saved for 8-bit 420 720p and 1080p compressed video streams. Bitrate savings are computed as BDRATE (using piecewise linear approach). Each point on the chart is the average BDRATE computed across 27 standard short sequences generated in both CBR and VBR. The objective visual quality metric used in the BDRATE calculation is Luma PSNR, averaged across frames. BDRATE is calculated using baselines of x264 medium for AVC and AV1 or x265 medium for HEVC.
Caption: Using the aforementioned methodology for bitrate savings assessment, the above charts show bitrate savings for individual test sequences with respect to the reference. Bitrate savings S-curves for Quality and Balanced random access modes are shown for the Intel® Data Center GPU Flex Series encoders across different content.
For more details about testing methodology in case of Low Delay encoding, check Low Delay Video Quality Command Lines and Measuring Methodology.
Caption: These charts illustrate quality assessment of our Intel® Data Center GPU Flex Series encoders (in low delay use case) expressed as a percent of bitrate saved for 8-bit 420 720p and 1080p compressed video streams. Bitrate savings are computed as BDRATE (using piecewise linear approach). Each point on the chart is the average BDRATE computed across 27 standard short sequences generated in both CBR and VBR. The objective visual quality metric used in the BDRATE calculation is Luma PSNR, averaged across frames. BDRATE is calculated using baselines of x264 medium for AVC and AV1 or x265 medium for HEVC.
Caption: This chart uses the same approach but illustrates the consistent average bitrate savings of the Intel AV1 encoder over (Intel) AVC and software x264 encoder with medium preset and tuned for low delay use case.
Caption: Using the aforementioned methodology for bitrate savings assessment, the above charts show bitrate savings for individual test sequences with respect to the reference. Bitrate savings S-curves for Quality and Balanced low delay modes are shown for the the Intel® Data Center GPU Flex Series encoders across different content.
Below we will provide command line recommendations for ffmpeg transcoding in Random Access and Low Delay modes.
For more details on ffmpeg-qsv supported features, see ffmpeg-qsv capabilities.
For more information on how to engage with Intel GPU encoding, decoding and transcoding as well as deal with multiple GPUs, please refer to ffmpeg-qsv multi-GPU selection document.
The recommended good practices are used throughout this project: in the demo examples as well as in the quality and performance measuring tools. The following links provide additional information:
- Random Access Video Quality Command Lines and Measuring Methodology
- Low Delay Video Quality Command Lines and Measuring Methodology
- Video Performance Command Linux and Measuring Methodology
Intel’s advanced software bitrate controller (dubbed “EncTools”) has been designed to boost GPU video quality for AVC, HEVC and (coming soon) AV1 using various compression efficiency technologies and content adaptive quality optimization tools while at the same time having minimal impact on the coding performance (speed). EncTools technology includes tools such as adaptive pyramid quantization, persistence adaptive quantization, low power look ahead, advanced scene change detection and more.
The recommended random access transcoding ffmpeg-qsv (Intel GPU integration with ffmpeg) command lines optimized for high quality and performance are given below:
AVC/H.264:
ffmpeg -hwaccel qsv -qsv_device ${DEVICE:-/dev/dri/renderD128} -c:v $inputcodec -extra_hw_frames 8 -an -i $input \
-frames:v $numframes -c:v h264_qsv -preset $preset -profile:v high -async_depth 1 \
-b:v $bitrate -maxrate $((2 * $bitrate)) -bitrate_limit 0 -bufsize $((4 * $bitrate)) \
-rc_init_occupancy $((2 * $bitrate)) -low_power ${LOW_POWER:-true} \
-look_ahead_depth 8 -extbrc 1 -b_strategy 1 \
-adaptive_i 1 -adaptive_b 1 -bf 7 -refs 5 -g 256 -strict -1 \
-vsync passthrough -y $output
HEVC/H.265:
ffmpeg -hwaccel qsv -qsv_device ${DEVICE:-/dev/dri/renderD128} -c:v $inputcodec -extra_hw_frames 8 -an -i $input \
-frames:v $numframes -c:v hevc_qsv -preset $preset -profile:v main -async_depth 1 \
-b:v $bitrate -maxrate $((2 * $bitrate)) -bufsize $((4 * $bitrate)) \
-rc_init_occupancy $((2 * $bitrate)) -low_power ${LOW_POWER:-true} \
-look_ahead_depth 8 -extbrc 1 -b_strategy 1 \
-bf 7 -refs 4 -g 256 -idr_interval begin_only -strict -1 \
-vsync passthrough -y $output
AV1 (HW-based BRC, EncTools coming soon):
ffmpeg -hwaccel qsv -qsv_device ${DEVICE:-/dev/dri/renderD128} -c:v $inputcodec -an -i $input \
-frames:v $numframes -c:v av1_qsv -preset $preset -profile:v main -async_depth 1 \
-b:v $bitrate -maxrate $((2 * $bitrate)) -bufsize $((4 * $bitrate)) \
-rc_init_occupancy $(($bufsize / 2)) -b_strategy 1 -bf 7 -g 256 \
-vsync passthrough -y $output
Extra quality boost can be achieved with use of low power look ahead (by setting “-look_ahead_depth 40” option) at the expense of a slight performance impact (10-20%). The use of "-extra_hw_frames" option is currently required for transcoding with look ahead due to the increased GPU memory requirements. Please set the value for "-extra_hw_frames" to be the same as the number of lookahead frames.
For best single stream performance on low density use case with high resolutions such as 4K, “-async_depth 2” option is recommended (yielding only negligible quality loss compared to “-async_depth 1”).
Recommendations for more specific use cases as well as additional information on developer configurable bitrate controllers and available advanced coding options is provided in the supplementary Video Quality document.
The recommended low delay transcoding ffmpeg-qsv (Intel GPU integration with ffmpeg) command lines optimized for high quality and performance are given below:
AVC/H.264:
ffmpeg -hwaccel qsv -qsv_device ${DEVICE:-/dev/dri/renderD128} -c:v $inputcodec -an -i $input \
-frames:v $numframes -c:v h264_qsv -preset $preset -profile:v high -async_depth 1 \
-b:v $bitrate -maxrate $bitrate -minrate $bitrate -bufsize $((bitrate / 4)) \
-rc_init_occupancy $((bitrate / 8)) -bitrate_limit 0 \
-bf 0 -refs 5 -g 9999 -strict 1 -vsync passthrough -y $output
HEVC/H.265:
ffmpeg -hwaccel qsv -qsv_device ${DEVICE:-/dev/dri/renderD128} -c:v $inputcodec -an -i $input \
-frames:v $numframes -c:v hevc_qsv -preset $preset -profile:v main -async_depth 1 \
-b:v $bitrate -maxrate $bitrate -minrate $bitrate -bufsize $((bitrate / 4)) \
-rc_init_occupancy $((bitrate / 8)) \
-bf 0 -refs 4 -g 9999 -strict 1 -vsync passthrough -y $output
AV1:
ffmpeg -hwaccel qsv -qsv_device ${DEVICE:-/dev/dri/renderD128} -c:v $inputcodec -an -i $input \
-frames:v $numframes -c:v av1_qsv -preset $preset -profile:v main -async_depth 1 \
-b:v $bitrate -maxrate $bitrate -minrate $bitrate -bufsize $((bitrate / 2)) -rc_init_occupancy $((bitrate / 4)) \
-bf 0 -g 9999 -vsync passthrough -y $output
Recommendations for more specific use cases as well as additional information on developer configurable bitrate controllers and available advanced coding options is provided in the supplementary Low Delay Video Quality document.
Intel® Data Center GPU Flex Series uses the same API’s and software components as Intel’s integrated graphics adapters. Developers and users can easily access Linux drivers, oneVPL, Intel Media SDK, and FFMPEG. Our commitment to open source allows developers to easily customize these components for any video application.
Use FFmpeg command line tool to perform basic transcode operations. See command line examples to achieve optimal quality level for content delivery usage scenarios. Check out generic examples for Intel oneVPL and Media SDK Plugins for FFmpeg.
Start developing or enhance your own application reading oneVPL specification.
- Intel® Media Delivery Software Stack
- Intel® Media Driver
- Intel® oneAPI Video Processing Library
- Intel® Media SDK
- Intel® Media SDK Plugins for FFmpeg (also known as Intel® Quick Sync Video Plugins for FFmpeg)
Test by Intel 9/16/2020, Intel® Server System M50CYP2UR208 (products formerly Coyote Pass), Intel® Xeon® Gold 6336Y Processor @ 2.4GHz, 1 node, 2 sockets, 24 cores/socket, 2 threads/core, 1024GB 32-ch 3200 MT/sec, Intel Turbo Boost enabled, Intel Hyper Threading enabled, BIOS: 0020P41_CoyotePass_LBG_ICX_UpdateCapsule_prd.bin. Intel® Data Center GPU Flex Series 140, ATS_M75_128_B0_ES_023_22WW29_03_GS1792_PC9707A_OP1051_ECC_ON_GFX2267.bin, 2 devices per-card, both used to compute performance (stream density). N concurrent sessions’ average fps used for performance stream density report. Data collected on PVT card < 65C Temperature. Ubuntu 20.04 LTS, linux-image-unsigned-5.14.0-1045-oem kernel. Commercial products may operate at higher or lower frequency.
DG1 reference quality data is measured by Intel 11/20/2020, 11th Gen Intel® Core™ i7-1185G7 (Product formerly Tiger Lake) @ 3.00GHz, 1 socket, 2 threads/core, 8 total CPUs, Intel Turbo Boost enabled, Total Memory 7714372kB, BIOS: TGLSFWI1.R00.3373.A00.2009091720 (ucode: 0x60), Ubuntu 20.04 LTS, gcc (Ubuntu 9.3.0-10ubuntu2) 9.3.0. For more details see Accelerating Media Delivery with Intel® Iris® Xe MAX graphics.
CPU reference quality data is measured by Intel 02/09/2022 on an Intel SawTooth Pass Server, 1-node, 2-socket, 28 cores/socket, 2 threads/core, Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz with enabled Intel Hyper Threading, enabled Intel Turbo Boost, CPU Microcode 0x4D, Windows 10 Enterprise LTSC 64-bit, 240GB 15-ch, DDR4-2666 DRAM.
Multi-stream performance data is collected using scripts noted above running file-to-file transcode. The scripts execute multiple concurrent 720p, 1080p, or 4K content streams, measuring the average frame rate of the transcoding process, at increasing numbers of streams to seek a target (typically 30 fps or 60 fps). The maximum stream density that meets or exceeds 98% of the target fps is reported.
The following is a table of the project versions used.
Project versions
| Component | Version |
|---|---|
| I915 DKMS | UBUNTU2204_22WW34_419_5949_220707.2 |
| CSE DKMS | 22WW33_419.38_UBUNTU514 |
| PMT DKMS | 22WW33_419.38_UBUNTU514 |
| Intel® Media driver for VAAPI | intel-media-22.5.2 |
| Intel® oneAPI Video Processing Library | v2022.2.0 |
| Intel® oneVPL GPU Runtime | intel-onevpl-22.5.2 |
| Intel® Media SDK | intel-mediasdk-22.5.2 |
| libva | 2.15.0 |
| Intel® Graphics Memory Management Library | intel-gmmlib-22.1.7 |
| ffmpeg | f6a36c7 |
| ffmpeg-cartwheel | 53a3f44 |
Performance varies by use, configuration and other factors. Learn more at www.intel.com/PerformanceIndex
Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available updates. See backup for configuration details. No product or component can be absolutely secure.
See backup for configuration details. For more complete information about performance and benchmark results, visit www.intel.com/benchmarks.
Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors.
Intel technologies may require enabled hardware, software or service activation.
Your costs and results may vary.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.