Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Acceleration Platform
Audio signal representation &
preprocessing Neural network
15 1413 12
11 109 8
7 65 4
3 21
12
34
56
78
910
1112
1314
15
0
0.2
0.4
0.6
0.8
1
Activation precision [bits]Weight precision [bits]
The Xelera Suite Software accelerates Deep
Learning model executions in order to enable
inference with low and low-variance latency. Itachieves this by offloading the inference to
highly efficient hardware accelerators such as
FPGAs. The software is available where FPGAs
are available: in the public cloud, in on-premises
data centers or edge computing servers. Xelera
Suite ensures that the accelerator connects
seamlessly to the host CPU, network or sensor
I/O. The software can be used without
application code changes because it integrates
into common software frameworks. Xelera
provides the Deep Learning acceleration
software, and optional customization and
integration services.
Model parameter extraction
Model quantization
Tailored FPGA
architecture synthesis
Deep Learning Accelerator on FPGAs Tool Flow
Scalable FPGA Architecture
Control &
Tiling
Pooling &
Activation
Scalable
accelerator
Model Quantization
Reference Use Case – Real-Time Speaker Recognition – 6 ms Inference
AI
Engine
Tensor
precision [bits]
Weight
precision [bits]
Select optimal
quantization
Rel. error metric
FPGA-enabled standard
server or cloud instance
Deep Learning model
Convert
Keras/Tensorflow
model into
reduced-bitwidth
fixed-point
representation
Mic Audio signal
Detected
speaker
6 ms
Platform Parallel AI engines
per FPGA
Xilinx® AlveoTM
U20010
Xilinx® AlveoTM
U25016
AWS F1.2
instance10
AWS F1.16
instance80