White Papers

5 Deep Learning Inferencing with Mipsology using Xilinx ALVEO™ on Dell EMC Infrastructure
Why FPGA?
FPGAs achieve high computation throughput via a robust set of resources that comprises
substantial reprogrammable lookup tables (LUT) to implement millions of equivalent Boolean-
logic functions, a large assembly of multipliers/adders (MAC), numerous embedded memories
to accommodate a broad variety of logic circuitry. They can also support a high number of off-
chip memories if necessary. A series of auxiliary logic, such I/O interfaces, etc., complete the
device. The overall fabric is ideal for parallel processing as required by (neural networks) NN.
See figure 3.
Figure 3. (Left) Arrayed building blocks are connected via interconnect wires; (Right) Fully
featured FPGAs include a variety of advanced building blocks.
While CPUs/GPUs operate at the byte level, FPGAs function at the bit level, giving the user the
ability to design logic to perfectly fit a task. Users can handle irregular parallelism and fine-
grained computations much better with FPGAs than with CPUs/GPUS. FPGAs are ideal for
processing both sparse data and compact data types.
The re-programmability of the FPGA permits an unusual degree of customization after the
hardware was manufactured. The ability to tune the underlying hardware architecture and select
any computation quantization desired allows for FPGA-based platforms to support state-of-the-
art deep learning innovations as they emerge.
As mentioned before, the cornucopia of resources and the benefits come at a price. Not only do
FPGAs require unique skills and expertise to program and deploy them, but their compilation
process is very slow, and they need a laborious tuning process to achieve high frequency
computation.
Xilinx ALVEO™ U200
The Xilinx® ALVEO™ U200 and U250 data center accelerator cards are peripheral component
interconnect express (PCIe®) Gen3 x16 compliant cards featuring the Xilinx Virtex®
UltraScale+™ technology. These cards accelerate compute-intensive applications such as
machine learning, data analytics and video processing. The ALVEO™ U200 and U250 data
center accelerator cards are available in passive and active cooling configurations. See figure 4
and 5.