Forum Discussion
New ContributorIs Spatial IP ready for LLM / transformer inference?
I am using FPGA AI Suite 2026.1.1 (with the new spatial compiler). Most of the FPGA AI Suite handbook examples I see are classical CNN / vision flows (ResNet-style) on PCIe, hostless JTAG, and SoC.
Is transformer / LLM inference (attention layers, variable sequence lengths, large KV-cache activations, etc.) something we can target today with dla_compiler + Spatial IP, or is Spatial still aimed primarily at CNN-like graphs, or is custom RTL expected?
And if yes, are there any LLM examples, guides, recommended flows, or known limitations?
Thanks,
3 Replies
DomiP_AlteraLLM inference is not supported across any FPGA Ai suite releases as of now. It's in the works.
The standard FPGA AI Suite workflow is not supported for LLM / Transformer models on any platform.- DomiP_Altera
Currently FPGA AI Suite 2026.1.1 (with the new spatial compiler) doesn't support LLM / transformer inference.
raghadCould you please clarify:
- Is LLM inference unsupported across all FPGA AI Suite releases, or is this something being actively worked on for a future release?
- Can I confirm that the standard FPGA AI Suite workflow is not supported for LLM / Transformer models on the Agilex 5 E-Series 065B Modular Development Kit?
The workflow I am referring to is:
Hugging Face model(optimum-intel) → OpenVINO IR (.xml + .bin) → architecture_optimizer → dla_compiler (Sequential flow) or Spatial Compiler → Integrate the generated FPGA AI Suite IP into a Quartus Prime project → Generate the bitstream and program it onto the Agilex 5 E-Series 065B board → Run inference using the FPGA AI Suite runtime (host application).
I understand this may not be a push-button process and could require significant modifications to the generated RTL — but is this workflow still considered a viable starting point for implementing LLM / Transformer inference on the Agilex 5 E-Series 065B?
Thank you very much.
