Amd 100-505950 Product Data Sheet

AMD FirePro™ S9300 x2 Server GPU

The World’s First GPU Accelerator with 1TB/s Memory Bandwidth

Speed up your most complex HPC workloads
in data analytics, or signal processing using
C++ or OpenCL™, and witness the power of
the world’s first GPU Accelerator with HBM
memory, the AMD FirePro™ S9300 x2.
Experience the fastest 32-bit compute GPU
accelerator, with up to 13.9 TFLOPS of peak
single-precision compute accelerating your
HPC workflows.

Heterogeneous-compute Interface for Portability
(HIP) Tool – Already have code written in CUDA? Don’t
want to be tied to a single vendor? Easily convert your
code to C++ with this free and open source tool, while
maintaining compatibility with CUDA compilers. The
HIP tool allows developers to port the majority of their
CUDA code over to C++ in a snap. Get started on the
AMD FirePro S9300 x2 GPU, an open-source friendly
accelerator from AMD, today.

Key Features

• Max Power: 300W

• Bus Interface: PCIe® Gen 3 x16

• Form Factor: Dual Slot, Full Length, Full Height

• Cooling: Passive

• Memory Size: 8GB HBM

• OS Support: Linux® 64-bit

1,3

Unparalleled Single Precision Performance

With 13.9 TFLOPS of peak single precision performance, the AMD FirePro S9300 x2 is the
fastest GPU accelerator available, delivering up to 2x peak single precision performance over
NVIDIA’s Tesla M40 and 1.6x peak single precision performance over NVIDIA’s Tesla K80.
The latest 3rd generation GCN architecture brings a number of enhancements to the GPU
including a new 16-bit floating point and integer instruction set for more ecient use of
memory bandwidth and reduced memory footprint.

World’s First GPU Accelerator with HBM Memory

The AMD FirePro™ S9300 x2 Server GPU is the industry’s first server GPU with 8GB of ultra­fast HBM memory and features 1 TB/s of memory bandwidth (512GB/s per GPU). HBM is a
new type of memory design with low power consumption and ultra-wide communication
lanes. It uses vertically stacked memory chips interconnected by microscopic wires called
“through-silicon vias,” or TSVs, placed directly onto the interposer, shortening the distance
information has to travel between memory and processor.

Open Developer Ecosystem

AMD FirePro™ S9300 x2 Server GPU is fully compatible with the new AMD GPUOpen software
stack, which includes a new open source Linux driver, built from the ground up and optimized
for compute. This driver includes native support for GPU-to-GPU communication both within
and between nodes.

2,3

Developers now have the option of using OpenCL™ or C++ to accelerate code on the GPU. The
free, open source Heterogeneous Compute Compiler (HCC) leverages popular CLANG/LLVM
technology and supports compilation of C++ code to x86 or GPU targets. Since a majority of
developers already use C++ programming, there’s no need to learn a new language to start
accelerating on the GPU. The open source HCC compiler will allow developers to have single
source code development for the new AMD FirePro™ S9300 x2 GPU.

AMD FirePro™ S9300 x2 Server GPU

FEATURES

13.9 TFLOPS of Peak Single Precision

Helps speed up time required to complete single precision floating
point operations used within Simulations, Video Enhancement,
Signal Processing, Video Transcoding and Digital Rendering
applications where high performance takes precedence over
accuracy. With the AMD FirePro™ S9300 x2 delivering 13.9 TFLOPS
of peak single precision compute performance, one can configure
a 2P server with 8 GPUs to achieve over 111 TFLOPS of peak single
precision compute performance. In a standard 42U rack with 10x
4U servers, that’s potentially over 1 PFLOP of single precision
compute performance!

870 GFLOPS of Peak Double Precision

Helps speed up time required to complete double precision
floating point operations used within Computational Fluid
Dynamics, Structural Mechanics, Reservoir Simulation and
Aerodynamics applications, where numerical precision is
mission critical.

Half Precision (FP16) Support

Developers who do not need the accuracy of 32-bit mathematical
operations can now use 16-bit operations to help achieve high
performance through a more ecient use of memory bandwidth
and reduced memory footprint.

GPUOpen Professional Compute

Comprised of an open-source Linux driver optimized for compute,
support for GPU acceleration using a new compiler to process
code written in the C++ programming language, and other
developer tools such as the Heterogeneous-compute Interface for
Portability (HIP) Tool to port code written for CUDA to C++.

OpenCL™ 1.2 Support

Helps professionals tap into the parallel computing power
of modern GPUs and multicore CPUs to accelerate compute­intensive tasks in leading CAD/CAM/CAE and Media &
Entertainment applications that support OpenCL. The AMD
FirePro S9300 x2 Server GPU supports OpenCL™ 1.2, allowing
developers to take advantage of new features that give GPUs
more freedom to do the work they are designed to do.

AMD PowerTune

AMD PowerTune Technology is an intelligent power management
system that monitors both GPU activity and power draw. AMD
PowerTune optimizes the GPU to deliver low power draw when
GPU workloads do not demand full activity and delivers the
optimal clock speed to ensure the highest possible performance
within the GPU’s power budget for high intensity workloads.

4

8GB HBM Memory

HBM is a new type of memory design with low power
consumption and ultra-wide communication lanes. It uses
vertically stacked memory chips interconnected by microscopic
wires called “through-silicon vias,” or TSVs, placed directly onto
the interposer, shortening the distance information has to travel
between memory and processor.

For more information, please visit AMD.com/firepro/hpc

1. The AMD FirePro™ S9300 x2 GPU is the first professional dual GPU accelerator card with HBM memory available as of March 31st, 2016. FP-186

2. AMD internal testing as of March 2016. System configurations may vary, yielding different results. The AMD FirePro™ S9300 x2 delivers up to 13.9 TFLOPS peak single-precision compute performance, while the NVIDIA Tesla M40 delivers up to 7 TFLOPS peak single-precision performance. For AMD, calculation
based on a formula using a combination of shader units and clock speeds. For Nvidia, source of data as of March 18, 2016: http://www.nvidia.com/object/tesla-m40.html FP-187.

3. AMD internal testing as of March 2016. System configurations may vary, yielding different results. The AMD FirePro™ S9300 x2 delivers 13.9 TFLOPS of peak single-precision compute performance. As of March 18th, 2016, NVIDIA’s fastest Tesla board is the Tesla K80, which delivers up to 8.73 TFLOPS of single­precision compute performance. For AMD, calculation based on a formula using a combination of shader units and clock speeds. For Nvidia, source of data as of March 18, 2016 - http://www.nvidia.com/object/tesla-k80.html FP-189

4. AMD PowerTune technology is offered by certain AMD FirePro™ products, which is designed to intelligently manage GPU power consumption in response to certain GPU load conditions. Not all products feature all technologies – check with your component or system manufacturer for specific model capabilities.

©2016 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD arrow, FirePro, and combinations thereof are trademarks of Advanced Micro Devices, Inc. in the United States and/or other jurisdictions. OpenCL is a trademark of Apple, Inc. and used by permission of Khronos. DirectX is a registered trademark of
Microsoft Corporation. PCIe and PCI Express are registered trademarks of the PCI-SIG Corporation. Other names are for informational purposes only and may be trademarks of their respective owners. PID 169019-A