Skip to content
This repository has been archived by the owner on Jan 7, 2021. It is now read-only.

Latest commit

 

History

History
119 lines (95 loc) · 13.9 KB

README.md

File metadata and controls

119 lines (95 loc) · 13.9 KB
Raw

Overlay Selector Guide

(Only overlay_4 is supported going forward)

This directory is used to hold various hardware overlays for accelerating neural networks on different platforms.
Overlays are compatible with specific hardware platforms. Alveo platform overlays must be downloaded from Xilinx and unzipped within this directory:
ml-suite/overlaybins/

AWS and Nimbix overlays are included directly in the github repository.

Background

Field Programmable Gate Arrays - FPGAs - are semiconductor devices, which by design implement an array of logic blocks with a programmable interconnect. Unlike "hardened" devices (i.e. CPU/GPU) FPGAs can be programmed to implement a hardware design that does specifically what the user wants. After the design of the hardware system, the FPGA must be programmed using a binary file. This process is typically referred to as configuration. Furthermore, in a usecase where there is fixed functionality, and dynamic functionality an FPGA can be partially reconfigured. In a datacenter environment, the FPGA is always connected to the CPU via PCIe, and it is always connected to external off-chip memory. Given the above assumption the FPGA binary can be partitioned into a static shell (Xilinx uses the term DSA), and a dynamic overlay (Xilinx uses the term xclbin). The static shell must be loaded prior to the loading of any overlay. Xilinx's cloud partners have already loaded the shell. If this explanation seems a bit confusing, don't worry too much about it. All you need to know is that you need to pick an overlay that will best meet your application needs. It will be loaded by the Python API.

ALVEO-U250 Hardware Overlays (XCLBINs)

The U250 board is similar to U200, but with a larger FPGA which can handle more accelerator cores.

Overlay # of PE Bitwidth DSP Array Dim. Max Image H/W Max Image Depth On-Chip Activation Memory Max Filter Size (HxWxD) DSA Version XDNN Version
overlay_0.xclbin* 6 8 32x28 1023 4095 4MB 4608 xilinx_u250_xdma_201820_1 2.3
overlay_1.xclbin* 6 16 32x28 1023 4095 4MB 4608 xilinx_u250_xdma_201820_1 2.3
overlay_2.xclbin 4 8 32x56 1023 4095 6MB 9792 xilinx_u250_xdma_201820_1 2.3
overlay_3.xclbin 4 16 32x56 1023 4095 6MB 9792 xilinx_u250_xdma_201820_1 2.3
overlay_4.xclbin 4 8 16x96 1023 4095 9MB 9792 xilinx_u250_xdma_201820_1 3.0

*Note: Currently, overlay_0 and overlay_1 are not available on the alveo-u250 platform. Most likely, these will not be provided, as we migrate to XDNNv3 usage.

Overlay Notes
overlay_0.xclbin This overlay is best for high throughput applications as it can process 8 images in parallel, it does not support networks with very large filters such as YOLOv2
overlay_1.xclbin This overlay is great for high throughput applications as it can process 4 images in parallel, and it will maintain accuracy nearly equivalent to float32 models
overlay_2.xclbin This overlay is best for high throughput applications as it can process 4 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2
overlay_3.xclbin This overlay is great for high throughput applications as it can process 2 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models
overlay_4.xclbin This overlay is best for all applications. It can process 1 image at a time at ~3x speed compared to 32x56 alternative, the optimized DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models

ALVEO-U200 Hardware Overlays (XCLBINs)

Overlay # of PE Bitwidth DSP Array Dim. Max Image H/W Max Image Depth On-Chip Activation Memory Max Filter Size (HxWxD) DSA Version XDNN Version
overlay_0.xclbin 4 8 32x28 1023 4095 4MB 4608 xilinx_u200_xdma_201820_1 2.3
overlay_1.xclbin 4 16 32x28 1023 4095 4MB 4608 xilinx_u200_xdma_201820_1 2.3
overlay_2.xclbin 2 8 32x56 1023 4095 6MB 9792 xilinx_u200_xdma_201820_1 2.3
overlay_3.xclbin 2 16 32x56 1023 4095 6MB 9792 xilinx_u200_xdma_201820_1 2.3
overlay_4.xclbin 2 8 16x96 1023 4095 9MB 9792 xilinx_u200_xdma_201820_1 3.0
Overlay Notes
overlay_0.xclbin This overlay is best for high throughput applications as it can process 8 images in parallel, it does not support networks with very large filters such as YOLOv2
overlay_1.xclbin This overlay is great for high throughput applications as it can process 4 images in parallel, and it will maintain accuracy nearly equivalent to float32 models
overlay_2.xclbin This overlay is best for high throughput applications as it can process 4 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2
overlay_3.xclbin This overlay is great for high throughput applications as it can process 2 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models
overlay_4.xclbin This overlay is best for all applications. It can process 1 image at a time at ~3x speed compared to 32x56 alternative, the optimized DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models

NIMBIX Hardware Overlays (XCLBINs)

Overlay # of PE Bitwidth DSP Array Dim. Max Image H/W Max Image Depth On-Chip Activation Memory Max Filter Size (HxWxD) DSA Version XDNN Version
overlay_0.xclbin 4 8 32x28 1023 4095 4MB 4608 xilinx:vcu1525:dynamic:5.1 2.3
overlay_1.xclbin 4 16 32x28 1023 4095 4MB 4608 xilinx:vcu1525:dynamic:5.1 2.3
overlay_2.xclbin 2 8 32x56 1023 4095 6MB 9792 xilinx:vcu1525:dynamic:5.1 2.3
overlay_3.xclbin 2 16 32x56 1023 4095 6MB 9792 xilinx:vcu1525:dynamic:5.1 2.3
overlay_4.xclbin 2 8 16x96 1023 4095 9MB 9792 xilinx:vcu1525:dynamic:5.1 3.0
Overlay Notes
overlay_0.xclbin This overlay is best for high throughput applications as it can process 8 images in parallel, it does not support networks with very large filters such as YOLOv2
overlay_1.xclbin This overlay is great for high throughput applications as it can process 4 images in parallel, and it will maintain accuracy nearly equivalent to float32 models
overlay_2.xclbin This overlay is best for high throughput applications as it can process 4 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2
overlay_3.xclbin This overlay is great for high throughput applications as it can process 2 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models
overlay_4.xclbin This overlay is best for all applications. It can process 1 image at a time at ~3x speed compared to 32x56 alternative, the optimized DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models

AWS Hardware Overlays (XCLBINs)

Overlay # of PE Bitwidth DSP Array Dim. Max Image H/W Max Image Depth On-Chip Activation Memory Max Filter Size (HxWxD) DSA Version XDNN Version
overlay_0.xclbin 4 8 32x28 1023 4095 4MB 4608 xilinx:aws-vu9p-f1-04261818:dynamic:5.0 2.3
overlay_1.xclbin 4 16 32x28 1023 4095 4MB 4608 xilinx:aws-vu9p-f1-04261818:dynamic:5.0 2.3
overlay_2.xclbin 1 8 32x56 1023 4095 5MB 9792 xilinx:aws-vu9p-f1-04261818:dynamic:5.0 2.3
overlay_3.xclbin 1 16 32x56 1023 4095 5MB 9792 xilinx:aws-vu9p-f1-04261818:dynamic:5.0 2.3
overlay_4.xclbin* 1 8 16x96 1023 4095 9MB 9792 xilinx:aws-vu9p-f1-04261818:dynamic:5.0 3.0

*Note: Currently, overlay_4 is not available on the aws platform. This will be updated shortly.

Overlay Notes
overlay_0.xclbin This overlay is best for high throughput applications as it can process 8 images in parallel, it does not support networks with very large filters such as YOLOv2
overlay_1.xclbin This overlay is great for high throughput applications as it can process 4 images in parallel, and it will maintain accuracy nearly equivalent to float32 models
overlay_2.xclbin This overlay is best for low latency applications as it can process 2 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2
overlay_3.xclbin This overlay is good for low latency applications as it can process 1 image at a time at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models
overlay_4.xclbin This overlay is best for all applications. It can process 1 image at a time at ~3x speed compared to 32x56 alternative, the optimized DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models

VCU1525 Hardware Overlays (XCLBINs)

Overlay # of PE Bitwidth DSP Array Dim. Max Image H/W Max Image Depth On-Chip Activation Memory Max Filter Size (HxWxD) DSA Version XDNN Version
overlay_0.xclbin 4 8 32x28 1023 4095 4MB 4608 xilinx:vcu1525:dynamic:5.1 2.3
overlay_1.xclbin 4 16 32x28 1023 4095 4MB 4608 xilinx:vcu1525:dynamic:5.1 2.3
overlay_2.xclbin 2 8 32x56 1023 4095 6MB 9792 xilinx:vcu1525:dynamic:5.1 2.3
overlay_3.xclbin 2 16 32x56 1023 4095 6MB 9792 xilinx:vcu1525:dynamic:5.1 2.3
overlay_4.xclbin 2 8 16x96 1023 4095 9MB 9792 xilinx:vcu1525:dynamic:5.1 3.0
Overlay Notes
overlay_0.xclbin This overlay is best for high throughput applications as it can process 8 images in parallel, it does not support networks with very large filters such as YOLOv2
overlay_1.xclbin This overlay is great for high throughput applications as it can process 4 images in parallel, and it will maintain accuracy nearly equivalent to float32 models
overlay_2.xclbin This overlay is best for high throughput applications as it can process 4 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2
overlay_3.xclbin This overlay is great for high throughput applications as it can process 2 images in parallel at 2x speed compared to 32x28 alternative, the larger DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models
overlay_4.xclbin This overlay is best for all applications. It can process 1 image at a time at ~3x speed compared to 32x56 alternative, the optimized DSP compute array provides minimum latency, and the larger supported filter size accomodates large networks like YOLOv2, and it will maintain accuracy nearly equivalent to float32 models