Can an FPGA be used as a cloud instance?
Yes. Here a Zynq 7-series FPGA was provisioned and configured through OpenStack, so an orchestrator could load a design and route work to it like any other instance.
VIDEO / FPGA CLOUD
An H.264 decoder as an FPGA cloud instance, provisioned through OpenStack.
Video decode is cheap on a phone and expensive at fleet scale. Treating an FPGA as a cloud instance — something an orchestrator can provision, load and tear down — turns a fixed accelerator into elastic capacity, but only if the reconfiguration and the datapath both work through the cloud's own machinery rather than around it.
CONTEXT
Where this started.
The programme built exactly that: a Zynq 7-series design where an FPGA instance is configured through OpenStack to receive an H.264 stream, decode it, and send it back — with the decoder itself implemented in Vivado HLS rather than hand-written RTL.
The interesting part is the seam. The decoder is a hardware block; OpenStack is a cloud orchestrator; ICAP is the FPGA's own reconfiguration port. Making those cooperate — so a remote client sends video over Ethernet and gets decoded video back — is the work.
CHALLENGES
4 problems, named.
Stated before they had answers.
01
An FPGA that behaves like a cloud instance
The accelerator had to be provisioned and configured through OpenStack, not flashed by hand.
02
A decoder from HLS, not hand RTL
The H.264 decoder was implemented in Vivado HLS, trading hand-tuned RTL for a higher-level flow.
03
Reconfiguration through the fabric's own port
Loading the design used ICAP, the FPGA's internal configuration access port, under cloud control.
04
A clean network datapath
A remote client had to send the encoded stream and receive decoded video over Gigabit Ethernet, reliably.
ARCHITECTURE
How the system was built.
An OpenStack controller provisions and configures the FPGA instance over a Gigabit link, using ICAP to load the design. On the Zynq 7-series device the H.264 decoder — built in Vivado HLS — sits in the programmable logic alongside the MAC and the processing system. A remote client sends the encoded stream and receives decoded video back over 1 Gbps Ethernet.
CONTRIBUTION
What Faststream did.
The specific scope, rather than a capability list.
- FPGA cloud instance — a Zynq 7-series design provisioned and configured through OpenStack.
- H.264 decoder in Vivado HLS — the decode block implemented via high-level synthesis rather than hand RTL.
- Reconfiguration via ICAP — loading the design through the FPGA's internal configuration access port under cloud control.
- Network datapath — receiving the encoded stream and returning decoded video over 1 Gbps Ethernet.
- Cloud integration — making the OpenStack orchestrator, the reconfiguration port and the decoder cooperate as one flow.
WHAT WAS HARD
The parts that consumed the schedule.
Written out because a reader facing the same programme gets more from this than from a summary of what went well.
01
Making a decoder behave like an instance
The hard part was the seam between OpenStack, ICAP reconfiguration and the decoder — three systems that had to cooperate to look like one elastic service.
02
HLS for a stateful codec
Implementing H.264 decode in Vivado HLS meant expressing an inherently stateful, control-heavy algorithm in a flow that prefers streaming dataflow.
03
Reconfiguration under cloud control
Driving ICAP from the orchestrator, safely, so an instance can be loaded and reloaded without wedging the fabric.
04
A reliable round trip over Ethernet
Encoded in, decoded out, over 1 Gbps, without loss between a remote client and the fabric.
OUTCOME
What resulted.
An H.264 decoder running as an FPGA cloud instance
Provisioned through OpenStack, decoding a stream and returning it to a remote client.
A decoder built in Vivado HLS
The codec expressed in high-level synthesis rather than hand-written RTL.
Cloud-driven reconfiguration
The design loaded through ICAP under OpenStack control.
A pattern for elastic acceleration
An approach for treating FPGA accelerators as provisionable cloud capacity.
Confidentiality
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture and confidential deliverables are not disclosed. Where a figure would identify a customer or a design, it is omitted rather than approximated. More detail is available under a non-disclosure agreement, within the limits each customer has agreed.
CAPABILITY USED
What this was built from.
Every item links to its own page.
PLATFORM
Faststream Silicon
FPGA and SoC design.
SERVICEFPGA Design
Architecture, HDL, timing closure.
CAPABILITYCloud, OTA & Device Management
Orchestration and fleet control.
SOLUTIONImage Processing & FPGA Vision
The wider body of imaging work.
CASE STUDYJPEG Decoder
Another codec in hardware.
CAPABILITYAI Silicon & Acceleration
Datapaths and acceleration.
WHERE THIS APPLIES
Industries this serves.
COMMON QUESTIONS
Questions this programme gets asked.
What is ICAP?
The Internal Configuration Access Port — the FPGA's own port for reconfiguring the fabric. It was used, under cloud control, to load the decoder design.
Why implement the decoder in Vivado HLS?
High-level synthesis raises the design abstraction above hand-written RTL. Applying it to H.264 decode — a stateful, control-heavy codec — was part of what the programme set out to do.
How does video reach the FPGA?
A remote client sends the H.264 stream over 1 Gbps Ethernet; the FPGA decodes it and returns the decoded video over the same link.
KEEP READING
Related work.
BUILD WITH FASTSTREAM
Bring us the difficult part.
Tell us the codec, the device and how it has to be provisioned. Turning accelerators into elastic capacity is core Faststream work.