One of the authors (of one of the two models, not this particular paper) here. Just a clarification, these models are *not* burned into silicon. They are trained with brutal QAT but are put onto fpgas. For axol1tl, the weights are burned in the sense that the weights are hard-wired in the fabric (i.e., shift-add instead of conventional read-muk-add cycle), but not on the raw silicon so the chip can be reprogrammed. Though, for projects like smartpixel or HG-Cal readout, there are similar ones targeting silicon (google something like "smartpixel cern", "HGCAL autoencoder" and you will find them), and I thought it was one of them when viewing the title.
Some slides with more info: https://indico.cern.ch/event/1496673/contributions/6637931/a... The approval process for a full paper is quite lengthy in the collaboration, but a more comprehensive one is coming in the following months, if everything went smoothly.
Regarding the exact algorithm: there are a few versions of the models deployed. Before v4 (when this article was written), they are slides 9-10. The model was trained as a plain VAE that is essentially a small MLP. In inference time, the decoder was stripped and the mu^2 term from the KL div was used as the loss (contributions from terms containing sigma was found to be having negliable impact on signal efficiency). In v5 we added a VICREG block before that and used the reconstruction loss instead. Everything runs in =2 clock cycles at 40MHz clock. Since v5, hls4ml-da4ml flow (https://arxiv.org/abs/2512.01463, https://arxiv.org/abs/2507.04535) was used for putting the model on FPGAs.
For CICADA, the models was trained as a VAE again, but this time distilled with supervised loss on the anomaly score on a calibration dataset. Some slides: https://indico.global/event/8004/contributions/72149/attachm... (not up-to-date, but don't know if there other newer open ones). Both student and teacher was a conventional conv-dense models, can be found in slides 14-15.
Just sell some of my works for running qat (high-granularity quantization) and doing deployment (distributed arithmetic) of NNs in the context of such applications (i.e., FPGA deployment for <1us latency), if you are interested: https://arxiv.org/abs/2405.00645 https://arxiv.org/abs/2507.04535
Happy to take any questions.
Very cool to see you work! Early in my PhD I did some work with GNN accelerators on FPGAs (which I think later ended up in some form as a colab with some CERN or Fermilab folks) and have chatted a bit in the past with the FastML, HLS4ML, and HEP folks.
I have since pivoted a lot of my PhD work (still related the HLS and EDA). But I wonder what is the current main limitation/challenges of building these trigger systems in hardware today. For example, in my mind it seems like the EDA and tooling can be a big limitation such as reliance on commercial HLS tools which can be buggy, hard to use, and hard to debug. From experience, this makes it harder to build different optimized architectures in hardware or build co-design frameworks without having high HLS expertise or putting in a lot of extra engineering/tooling effort. Also tool runtimes make the design and debug cycle longer, especially if you are trying to DSE on post-implementation metrics since you bring in implementation tools as well.
But I might be way off here and the real challenges are with other aspects beyond the tools.
Thank you for the comment, and the questions are great.
The problems you described here are pretty much precise. In the past, and mostly now, we are replying on the commercial Vivado/Vitis HLS toolchains for the deployment of these networks through hls4ml, a template based compiler of the quantized models to the HLS projects. For this class of fully parallel (II=1) models, the tools usually give fine results, but indeed can be wrong sometimes (great recent example from our college's post: https://sioni.web.cern.ch/2026/03/24/debugging-fastml).
Tool runtime is another issue. For the models discussed in this post, they are not larger than ~30K LUTs, and with the low complexity (~dense only), synthesis time was fine. But for larger ones, like the ones here (https://arxiv.org/abs/2510.24784), it can take up to... a week for one HLS compilation while eating ~80G ram. Can get worse if time multiplex is in place things like #pragma HLS dataflow is used...
Personally, I do not usually DSE on post implementation/HLS results, since for the unrolled logic blocks, ok-ish performance model can be derived obtained w/o doing the synthesis (via ebops defined in HGQ, or better if using heuristics based on the rough cost of low level operations the design will translate to). But there are works doing DSE based on post HLS results (https://arxiv.org/pdf/2502.05850, real vitis synth), or using some other surrogate to get over the problem (e.g., https://arxiv.org/abs/2501.05515, using bops). High-level surrogate models are also being developed (https://arxiv.org/pdf/2511.05615).
We are also trying to get alternatives to the commercial HLS toolflows. For instance, I'm working on the direct to RTL codegen (da4ml) way (optionally via XLS), and the current work-in-progress is at https://github.com/calad0i/da4ml/tree/dev, if you are interested: all combinational or fully pipelined things are supported with reasonable performance model (~10% err in LUTs and ~20% err in latency), but multicycle, or stateful design generations still need a lot of manual intervention (not automated), which are to be implemented in the future. Since at some stages of the trigger chain, the system is/will be time-multiplexed, such functionality will be needed in the future.
Other works on this direction includes adding new backends to hls4ml that are oos (e.g., openhls/XLS), or other alternatives like chisel4ml (https://github.com/cs-jsi/chisel4ml). Hopefully, we will be no-longer reliant on the commercial tools till RTL for the incoming upgrade. That being said, Vivado still appears to be the only choice for the post RTL stages for us.
