Many FPGA designs hit a power wall before they hit a logic wall. The enclosure is sealed, the rack has a fixed power envelope, or the board controller has to run every hour the system is on. In each case, the question is the same: how much work can you fit within the available power budget? The answer is mostly set long before the first power estimate. It’s set by the fabric architecture: how signals are registered, how data moves, and how densely logic packs. That’s where Agilex® 2X does its power work, and it does it differently at each end of the portfolio.
Why does power set the limit now?
Power used to be a line item checked late in a design. In many systems it’s now the first constraint.
Data center line cards and accelerator platforms work inside rack-level power and cooling budgets. Radios and satellite payloads live with thermal limits that don’t move. At the other end, board controllers, motor drives, and root-of-trust devices never switch off, so every milliwatt they draw, they draw all day.
Those are two different power problems. A bandwidth-heavy system needs more work per watt. An always-on control system needs fewer watts. A portfolio that solves only one of them sends designers to a new architecture when the requirement changes. Agilex 2X addresses both through the HyperFlex® fabric architecture: 3rd Gen HyperFlex fabric targets performance per watt, while HyperFlex-LP targets low power.
How does HyperFlex reduce fabric power at the same frequency?
3rd Gen HyperFlex fabric architecture places registers throughout the routing fabric, enabling efficient pipelining and retiming. As congestion and utilization increase, routing paths can become longer and slower; HyperFlex registers help maintain higher performance under these challenging conditions. Each generation has further refined this approach, as reflected in the performance curve.
Relative fabric Fmax and relative fabric power at iso-frequency across three generations of the HyperFlex fabric architecture, normalized to Stratix 10 devices.At the same frequency, Agilex 27 fabric uses approximately 40 percent less power than Agilex 7 fabric.¹ As the chart shows, Agilex 7 made a step of similar size over Stratix 10. Agilex 27 also delivers 1.8X performance per watt versus Agilex 7. Separately, it runs at more than 15 percent higher Fmax, on average, versus Agilex 7.¹
For systems where cooling and power delivery set the limit, the power reduction often matters more than the speed gain. It’s headroom you can spend on the workload.
Where else does the power go?
3rd Gen HyperFlex fabric can move large amounts of data across the device. But in bandwidth-heavy designs, implementing wide data movement in soft interconnect consumes logic and routing resources and can increase compile time, leaving fewer fabric resources available for the application workload.
Agilex 27 hardens that job in the HyperFlow data network. Data moves across the device at multi-terabit-per-second rates without consuming soft logic or adding routing congestion. Logic packing helps too. Recompile an existing design for Agilex 27 FPGAs in Quartus® Prime Pro Edition Software and it packs 5 to 10 percent more densely, using fewer adaptive logic modules (ALMs).
Together, those changes cut the power spent on design infrastructure and leave more of the budget for the application. That shows up three ways:
- Reaching target performance within rack or enclosure power budgets
- Reducing cooling and operating costs
- Giving more of the thermal budget to the workload instead of to moving data
What does low power mean for I/O-centric systems?
At the other end of the portfolio, the problem changes. A board and rack controller, a motor drive, or an interface bridge rarely needs peak fabric performance. It needs plenty of I/O across a broad voltage range, a small footprint, and a power draw low enough to disappear into the system budget, because it never switches off.
That’s the job Agilex 22 is built for, and it’s why Agilex 22 uses HyperFlex-LP, a low-power implementation of the HyperFlex fabric architecture. HyperFlex-LP optimizes programmable logic for lower area, lower power, and higher packing density. It uses the familiar Quartus Prime design flow while tuning the fabric to minimize power. Agilex 22 FPGAs are right-sized for I/O-centric systems, offering high I/O density for board and rack control, motor control, root-of-trust, and interface-bridging applications.
Why does one architecture matter for power?
Power savings are easy to lose in a migration. A new architecture means new IP, new timing behavior, and new tools. Each change is a chance for the power estimate to drift from what ships.
Agilex 2X keeps one architecture and one software experience across every family. Every device is supported by Quartus Prime software, extended by QuartusIQ, the AI expert companion for Quartus Prime software. Design workflows, IP, interfaces, and engineering knowledge carry forward from family to family.
The takeaway
Power efficiency is decided by the architecture, long before the board is built. Agilex 2X tunes one fabric for two different power problems: more work per watt where bandwidth rules, and lower power where the FPGA is always on.
Notes and disclaimers
¹ Based on Altera’s modelling of Agilex 27 versus Agilex 7 device fabric performance and power. Results may vary. Agilex 2X device architecture and feature descriptions are preliminary and subject to change. Interface support, device options, and performance figures are Altera targets and vary by device.
Capabilities and availability vary by family and series.
Contributor