search

Xilinx / ml-suite

Getting Started with Xilinx ML Suite
https://aws.amazon.com/marketplace/pp/B077FM2JNS
Other
335 stars 152 forks source link

Raspi + Spartan 6 #76

Closed dadrake3 closed 5 years ago

dadrake3 commented 5 years ago

I am looking to implement an edge ML inferencing accelerator for the raspi. I am wondering if it is possible to use ml-suite to compile a TensorFlow graph to run on the spartan 6 FPGA. I am not sure where to start and I am wondering if this would be possible within ml-suite

Haffon commented 5 years ago

Hi, @dadrake3 we downloaded and installed xrt and xdma from https://www.xilinx.com/products/acceleration-solutions/xilinx-machine-learning-suite.html#gettingStartedU200 for cloud on ubuntu 16.04, you can check https://www.xilinx.com/products/design-tools/ai-inference/ai-developer-hub.html#edge for Xilinx edge AI solution.

wilderfield commented 5 years ago

Hey, while leflow looks super cool, and interesting...

ml-suite is software that wraps a DNN engine that is relatively fixed. Synthesis is already done. The tools are provided to map data flow graphs to the engine. The engine is designed for 16nm FPGAs.

In the future we may add kernels that are optimized for specific networks.

On Thu, Feb 21, 2019 at 10:19 PM Haffon notifications@github.com wrote:

Hi, @dadrake3 https://github.com/dadrake3 we downloaded and installed xrt and xdma from https://www.xilinx.com/products/acceleration-solutions/xilinx-machine-learning-suite.html#gettingStartedU200 for cloud on ubuntu 16.04, you can check https://www.xilinx.com/products/design-tools/ai-inference/ai-developer-hub.html#edge for Xilinx edge AI solution.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Xilinx/ml-suite/issues/76#issuecomment-466287197, or mute the thread https://github.com/notifications/unsubscribe-auth/AO_GpzRc1tbtVktVQR0GrRff2t_MKPiqks5vP4wJgaJpZM4bIfW3 .

dadrake3 commented 5 years ago

Im still a little confused, does the ml-suite not work with the spartan 6 fpga made by xilinx?

wilderfield commented 5 years ago

Correct.

The hardware design coupled with ml-suite makes heavy use of a new silicon feature called “URAM”. URAM cells are large on chip SRAM buffers that hold network activations.

The kernels also use DDR4 controllers which don’t exist in the era of Spartan 6.

On Fri, Feb 22, 2019 at 11:23 AM dadrake3 notifications@github.com wrote:

Im still a little confused, does the ml-suite not work with the spartan 6 fpga made by xilinx?

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/Xilinx/ml-suite/issues/76#issuecomment-466515747, or mute the thread https://github.com/notifications/unsubscribe-auth/AO_GpwVxfomI7VmcFlh720RzvWc4uAqtks5vQEOagaJpZM4bIfW3 .

Haffon commented 5 years ago

@dadrake3 wilderfield is correct, xilinx only provide precompiled overlaybins for VCU1525, Alveo-U200(KU9P)/250(KU13P) and so on. Write you own DNN layer/kernel for spartan 6 may be interesting, How complex edge ML could be?