Innovative Solutions for Cloud Gaming, Media, Transcoding, & AI Inferencing
- 2. Better Faster Greener™ © 2022 Supermicro
Supermicro Solutions Using
Intel® Data Center GPU Flex Series
Radhika Rao
Dir. Product and Business Management,
Data Center GPUs
Intel Corporation
Thomas Jorgensen
Sr. Dir. Technology Enablement
Supermicro
Michael Schulman
Sr. Content Manager
Supermicro
- 4. Better Faster Greener™ © 2022 Supermicro
4
Topics To Be Covered
Details About the Intel® Data Center GPU Flex Series
What New Markets Does the Intel® Data Center GPU Flex Series Address
- 5. Super Flexible Data Center GPU
Better Faster Greener™ © 2022 Supermicro
5
For Visual Cloud, Media and Cloud Gaming
Cloud Gaming
AI Edge Inferencing
Media Delivery
Virtual Desktop Infrastructure
- 6. Intel Data Center GPUs
Better Faster Greener™ © 2022 Supermicro
6
FLEX SERIES
Intel Data Center GPU Flex 140
Intel Data Center GPU Flex 170
- 7. Optimized forTCO and Density
PCIe Gen 4 Cards
Half Height, half length, simple Passive Cooling
2 GPUs / card
Launching with Intel® Scalable Gen3 Processors
See the Vision section of intel.com/performanceindex for workloads and configurations. Results may vary.
FLEX SERIES 140
4
Media
Engines
75W
Power Envelope
8
RayTracing
Units
Architecture
Half
Height
PCIe
8
Xe cores
- 8. See the Vision section of intel.com/performanceindex for workloads and configurations. Results may vary.
FLEX SERIES 170
2
Media
Engines
150W
Power Envelope
32
RayTracing
Units
Architecture
Full
Height
PCIe
32
Xe cores
Optimized for Peak Performance
PCIe Gen 4 Cards
Full Height, ¾ length, single-wide, Passive Cooling
1 GPU / card
Launching with Intel® Scalable Gen3 Processors
- 9. Cloud Gaming
Better Faster Greener™ © 2022 Supermicro
9
• High density Cloud Gaming with
support for Android & Windows
• Up to 54 game streams per card
and 540 game streams per server
• Biggest 3rd party Game Engines
Supported
• Run Across CPU & GPU
Seamlessly
- 10. Supermicro Solutions – Cloud Gaming
Better Faster Greener™ © 2022 Supermicro
10 Better Faster Greener™ © 2022 Supermicro
10
X12 CloudDC: SYS-620C-TN12R
Supports Up to 6 Intel Data Center GPUs
Cloud Gaming
- 11. Media Processing and Delivery
Better Faster Greener™ © 2022 Supermicro
11
• Industry first AV1encoding and decoding in
HW
• >30% distribution cost savings
• 30+ 1080p HEVC & AV1 Streams per card
• Industry’s biggest media frameworks
supported
• oneVPL library for encode/decode and
processing unified API across CPU & GPU
• Open-Source SW
- 12. Media Delivery
Supermicro Solutions – Media Processing & Delivery
Better Faster Greener™ © 2022 Supermicro
12 Better Faster Greener™ © 2022 Supermicro
12
X12 BigTwin: SYS-220BT-DNTR/HNTRX
Supports Up to 8 Intel Data Center GPUs
- 13. Supermicro Solutions – Transcoding
Better Faster Greener™ © 2022 Supermicro
13 Better Faster Greener™ © 2022 Supermicro
13
X12 4U 10 GPU System: SYS-420GP-TNR
Supports Up to 10 Intel Data Center GPUs
Media Transcoding
- 14. AI Edge Inferencing
Supermicro Solutions – AI Edge Inferencing
Better Faster Greener™ © 2022 Supermicro
14 Better Faster Greener™ © 2022 Supermicro
14
X12 CloudDC: SYS-620C-TN12R
Supports Up to 6 Intel Data Center GPUs
- 15. Intel® Data Center GPU Flex Series Micro Architecture
Better Faster Greener™ © 2022 Supermicro
15
Scalable High-Performance Graphics
Architecture for visually immersive workloads
Xe core
Dedicated
RayTracing
Units
16
Vector
Engines
16
XMX
Engines
- 16. Better Faster Greener™ © 2022 Supermicro
16
Intel® Data Center GPU Flex Series
140
Intel Data Center GPU Flex Series 170
Card Form Factor
Half height, half length, single wide,
passive cooling
Full height, three-quarter length,
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
Systolic Arrays (Relative) 1x 2.5x
Ray Tracing Yes
Peak Compute (Systolic) 8 TFLOPS (FP32) / 105 TOPS (INT8) 16 TFLOPS (FP32) / 250 TOPS (INT8)
Cache 8 MB 16 MB
Memory Type GDDR6
Memory Capacity 12 GB (6 per GPU) 16 GB
Memory Bandwidth 300 GB/s (150 per GPU) 600 GB/s
Memory Bus Width 192 bits (96 per GPU) 256 bits
Virtualization (Instances)6 SR-IOV (62) SR-IOV (31)
Host Bus PCIe Gen 4
Host CPU Support 3rd Generation Intel Xeon® Scalable Processors
FLEX SERIES
Intel Data Center GPU Flex 140
Intel Data Center GPU Flex 170
- 18. Next Gen Video Experiences: AV1 on Intel® Data Center GPU Flex Series
Ushering in a new era of video with groundbreaking AV1
codec technology
Intel is leading the industry in bringing a suite of AV1
solutions to the market
Ultimate quality and performance optimized SVT-
AV1 CPU Software Encoder for Intel® Xeon™ and
Core™ processors
Intel® Data Center GPU Flex Series bringing
professional quality, performance AV1 HW Encode to
serve Data Center
- 19. Targeted Usages
Object Classification
Object Detection
Image Segmentation
Frameworks
Media
AVC/HEVC/AV1
JPEG/MJPEG
Upto 4k
Visual Inference, Media Analytics on Intel® Data Center GPU Flex Series
Xe Matrix
Extensions
Dedicated built-inAI functionality
Accelerating most AI DataTypes
256
INT8
ops/clock
128
FP16/BF16
ops/clock
512
INT4/INT2
ops/clock
- 20. Virtual Desktop Infrastructure
Better Faster Greener™ © 2022 Supermicro
20
• High density distributed virtual desktops
• Up to 62 virtual functions SRIOV
• No additional Virtualization costs for GPU
SW licensing
- 21. Virtual Desktop Infrastructure
Supermicro Solutions – Virtual Desktop Infrastructure
Better Faster Greener™ © 2022 Supermicro
21 Better Faster Greener™ © 2022 Supermicro
21
X12 BigTwin: SYS-220BT-DNTR/HNTRX
Supports Up to 8 Intel Data Center GPUs
- 23. Q & A
Learn More at https://www.supermicro.com/en/accelerators/intel
- 25. Intel Data Center GPUs
Better Faster Greener™ © 2022 Supermicro
25
Video Transcoding Cloud Gaming VDI AI Inference
30+
1080p streams
40+
Game streams
62
Virtualized
functions
150
AI TOPS
4
Xe Media
Engines
32
Xe Core & Ray
Tracing Units
AV1
HW Encode &
Decode
XMX
AI Acceleration
Built in
Peak Performance
Intel Data Center
GPU 140 (75W)
Maximum Density
Intel Data Center
GPU 170 (150W)
Editor's Notes
- Let’s first talk about the ATS-M150 optimized for peak performance.
- Let’s first talk about the ATS-M150 optimized for peak performance.
- As I said great HW needs great SW.
Intel Extensible and Open Software Architecture is focusing on optimizing end-to-end workloads.
Starting with low level GPU drivers and firmware, Intel’s OneAPI foundational layer, through use case specific open-source enablement of video, inference and gaming frameworks and APIs.
Resulting in creation of reference solutions for our partners.
- Video streaming currently accounts for over 80% of global internet traffic which drives significant demand for video processing in the datacenter and increases cost for service providers.
Providers are looking to simultaneously improve bitrate savings to keep costs low, while maintaining the visual quality needed to satisfy their customers.
Arctic Sound-M is well positioned to help solve this challenge, bringing its industry leading AV1 HW encoder paired with Intel’s oneVPL library to the market later this year.
ATSM with AV1 will deliver close to 30% distribution cost savings, which in todays market could translate to about $20M per year savings for 100K users.
-
Currently inference solutions are on a CPU and GPU in a closed ecosystem.
We have been working with customers to be able to seamless move their workloads to all Intel open source ecosystem.
There are three major buckets of usescases within the media analytics segment that we’re targeting with ATS-M
First : IOT and Video Analytics where video is coming in directly from a camera into the network and being processed to gather information. These typically object detection, object classification, and segmentation algorithms
Next: These algorithms are used “Library Indexing and Compliance” category to really understand what is going in the video and extracting more information.
Lastly is “AI Guided Enhancement” which is used to help improve video quality or drive down the cost of transmitting the video.
Here is what we are doing on ATS-M:
One: Intel is optimizing the most popular AI frameworks with TensorFlow, pytorch, and openvino to work efficiently with ATS-M on visual inference models.
Two: You can use the same tools in the openvino tool suite to enable hybrid workflows between CPUs and GPUs
Three: In addition to AV1, ATS-M will continue to support multiple video CODECs