RADEON Instinct MI8 Data Sheet

User Manual: RADEON-Instinct-MI8 DataSheet

Open the PDF directly: View PDF PDF.
Page Count: 3

DEEP LEARNING INFERENCE ACCELERATOR WITH 8.2 TFLOPS SINGLE
PRECISION COMPUTE PERFORMANCE IN EFFICIENT SCALE-OUT DESIGN1
Datacenter designers deploying Machine Learning and AI
systems today require systems capable of running more
complex workloads that are massively parallel in nature, while
continuing to improve system eciencies. Improvements in
the capabilities of accelerators over the last decade, along with
the advancement in software, are providing designers with
the option to build more ecient heterogeneous computing
systems to help them meet today’s challenges.
The Radeon Instinct™ MI8 server accelerator is a highly
ecient, cost-eective inference and HPC solution delivering
8.2 TFLOPS of peak performance with 4GB of ultrafast HBM1
memory, making it the perfect solution for running deep learning
inference applications, where lots of new smaller data set inputs
are being run at half or single precision against trained neural
networks to discover new knowledge.1 The MI8 accelerator
is also the perfect open solution for general purpose HPC
systems deployed in Financial, Energy, Life Science, Automotive,
Academic (Research & Teaching), Government Labs and other
Open Ecosystem For Machine Intelligence
Highlights
• 8.2 TFLOPS FP16 or FP32 Performance1
• Up To 47 GFLOPS Per Watt FP16 or FP32 Performance2
• 4GB HBM1 on 512-bit Memory Interface
• Passively Cooled Server Accelerator
• Large BAR Support for Multi GPU Peer to Peer
• ROCm Open Platform for HPC-Class Rack Scale
• Optimized MIOpen Libraries for Deep Learning
• MxGPU SR-IOV Hardware Virtualization
Key Features
GPU Architecture: AMD “Fiji”
Stream Processors: 4,096
GPU Memory: 4GB HBM1
Memory Bandwidth: Up to 512 GB/s
Performance:
Half-Precision (FP16) 8.2 TFLOPS
Single-Precision (FP32) 8.2 TFLOPS
Double-Precision (FP64) 512 GFLOPS
Bus Interface: PCIe® Gen 3 x16
MxGPU Capability: Yes
Board Form Factor: Full-Height, Dual-Slot
Length: 6”
Thermal Solution: Passively Cooled
Standard Max Power: 175W TDP
AMD’s Radeon Instinct™ MI8 accelerator
delivers superior half precision performance
with 4GB of high-bandwidth HBM1 memory
in an ecient, passively cooled, scale out
accelerator form factor.1
Warranty: Three Year Limited3
OS Support: Linux® 64-bit
ROCm Software Platform: Yes
Programing Environment:
ISO C++, OpenCL™, CUDA (via AMD’s HIP conversion tool)
and Python4 (via Anaconda’s NUMBA)
For more information, visit: Radeon.com/Instinct
Deep Learning Inference
P4
P40
MI8 8.2
0.19
0.09 Up to 90x
Up to 42x
Peak FP16 (TFLOPS)
OUTSTANDING SINGLE AND
HALF PRECISION PERFORMANCE
The Radeon Instinct™ MI8 server accelerator
is based on the “Fiji” architecture which is
built with AMD’s 3rd Generation Graphics
Core Next (GCN) packing 64 compute
units (CU) with 64 stream processor per
CU delivering 8.2 TFLOPS FP16 or FP32
compute performance in a single GPU card.1
The Radeon Instinct MI8 accelerator based on AMD’s 3rd generation “Fiji” architecture with
improved data-parallel processing and ultra-fast HBM1 memory delivers 8.2 TFLOPS of peak
performance with up to 512 GB/s of memory bandwidth in a single, passively cooled GPU card.1
The MI8 accelerator, combined with AMD’s ROCm open software platform, is the perfect solution
for cost sensitive system deployments for Machine Intelligence, Deep learning and HPC workloads,
where performance and eciency are key system requirements.
Design with support of AMD’s MxGPU SR-
IOV hardware virtualization technology for
optimized datacenter cycle utilization, the
Radeon Instinct MI8 provides a virtualization
solution with dedicated user GPU resources,
data security and version control, a cost
eective licensing model with no additional
hardware licensing fees, and a simplied
native driver model ensuring operating
system and application compatibility.
The Radeon Instinct MI8 design is a
passively-cooled accelerator design
for large-scale server deployments.
3RD GENERATION “FIJI”
ARCHITECTURE
MxGPU SR-IOV HARDWARE
VIRTUALIZATION
PASSIVELY
COOLED
4GB of ultrafast HBM1 GPU memory
delivering up to 512 GB/s of memory
bandwidth. HBM1 is a modern type of
memory design with low power consumption
and ultra-wide communication lanes.
HBM1: ULTRAFAST
MEMORY BANDWIDTH
AMD’s ROCm platform provides a scalable,
fully open source software platform
optimized for large-scale heterogeneous
system deployments with an open source
headless Linux driver, HCC compiler, rich
runtime based on HSA, tools and libraries.
ROCm OPEN SOFTWARE
PLATFORM
Deep Learning Inference
Radeon Instinct MI8
Nvidia Tesla P40
Nvidia Tesla P4
47
Peak FP16 GFLOPS/Watt
0.76
1.2
Up to 60x
Up to 38x
For more information, visit:
Radeon.com/Instinct
ROCm.github.io
©2017 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD arrow logo, Radeon, and combinations thereof, are trademarks of Advanced Micro Devices, Inc.
All other product names are for reference only and may be trademarks of their respective owners. “Fiji” is an internal architecture code name and not a product name.
SUPERIOR FP16 PERFORMANCE FOR INFERENCE2
1. Measurements conducted by AMD Performance Labs as of June 2, 2017 on the Radeon Instinct™ MI8 “Fiji” architecture based accelerator. Results
are estimates only and may vary. Performance may vary based on use of latest drivers. PC/system manufacturers may vary congurations yielding
dierent results. The results calculated for MI8 resulted in 8.2 TFLOPS peak half precision (FP16) performance and 8.2 TFLOPS peak single precision
(FP32) oating-point performance.
AMD TFLOPS calculations conducted with the following equation: FLOPS calculations are performed by taking the engine clock from the highest
DPM state and multiplying it by xx CUs per GPU. Then, multiplying that number by xx stream processors, which exist in each CU. Then, that number is
multiplied by 2 FLOPS per clock for FP32. To calculate TFLOPS for FP16, 4 FLOPS per clock were used.
Measurements on the Nvidia Tesla P40 resulted in 0.19 TFLOPS peak half precision (FP16) peak oating-point performance with 250w TDP GPU
card from external source.
Sources:
https://devblogs.nvidia.com/parallelforall/mixed-precision-programming-cuda-8/
http://images.nvidia.com/content/pdf/tesla/184427-Tesla-P40-Datasheet-NV-Final-Letter-Web.pdf
Measurements on the Nvidia Tesla P4 resulted in 0.09 TFLOPS peak half precision (FP16) oating-point performance with 75w TDP GPU card from
external source.
Sources:
https://devblogs.nvidia.com/parallelforall/mixed-precision-programming-cuda-8/
http://images.nvidia.com/content/pdf/tesla/184457-Tesla-P4-Datasheet-NV-Final-Letter-Web.pdf
AMD has not independently tested or veried external and/or third party results/data and bears no responsibility for any errors or omissions therein.
RIF-1
2. Measurements conducted by AMD Performance Labs as of June 2, 2017 on the Radeon Instinct™ MI8 “Fiji” architecture based accelerator. Results
are estimates only and may vary. Performance may vary based on use of latest drivers. PC/system manufacturers may vary congurations yielding
dierent results. The results calculated for Radeon Instinct MI8 resulted in 47 GFLOPS/watt peak half precision (FP16) performance and 47 GFLOPS/
watt peak single precision (FP32) oating-point performance.
AMD GFLOPS per watt calculations conducted with the following equation: FLOPS calculations are performed by taking the engine clock from the
highest DPM state and multiplying it by xx CUs per GPU. Then, multiplying that number by xx stream processors, which exist in each CU. Then, that
number is multiplied by 2 FLOPS per clock for FP32. To calculate TFLOPS for FP16, 4 FLOPS per clock were used.
Once the TFLOPs are calculated, the number is divided by the 175w TDP power and multiplied by 1,000.
Measurements on the Nvidia Tesla P40 based on 0.19 TFLOPS peak FP16 with 250w TDP GPU card result in 0.76 GFLOPS/watt peak half precision
(FP16) performance.
Sources for Nvidia Tesla P40 FP16 TFLOPs number:
https://devblogs.nvidia.com/parallelforall/mixed-precision-programming-cuda-8/
http://images.nvidia.com/content/pdf/tesla/184427-Tesla-P40-Datasheet-NV-Final-Letter-Web.pdf
Measurements on the Nvidia Tesla P4 based on 0.09 TFLOPS peak FP16 with 75w TDP GPU card result in 1.2 GFLOPS/watt peak half precision
(FP16) performance.
Sources for Nvidia Tesla P40 FP16 TFLOPs number:
https://devblogs.nvidia.com/parallelforall/mixed-precision-programming-cuda-8/
http://images.nvidia.com/content/pdf/tesla/184457-Tesla-P4-Datasheet-NV-Final-Letter-Web.pdf
AMD has not independently tested or veried external and/or third party results/data and bears no responsibility for any errors or omissions therein.
RIF-2
3. The Radeon Instinct GPU accelerator products come with a three year limited warranty. Please visit www.AMD.com/warranty for details on the
specic graphics products purchased. Toll-free phone service available in the U.S. and Canada only, email access is global.
4. Support for Python is planned, but still under development.
FOOTNOTES

Navigation menu