Post

When FHE (fully homomorphic encryption) came up before, my first reaction was always: The technology is undoubtedly powerful, but can it handle the costs and speed requirements of actually being used in DeFi?



After all, on-chain finance is not a laboratory.

A single transfer may involve balance checks, deducting the amount, adding to the recipient’s balance, and updating the state. If all these operations are performed on encrypted data, the computational load and hardware costs become unavoidable concerns.

This time, after seeing AutoHoG released by @primus_labs, I think what’s really worth paying attention to isn’t simply the performance boost, but the smarter way it handles computation.

One pain point with FHE has been that complex logic gets broken down into large numbers of small Boolean gates, each of which requires an expensive Bootstrapping operation.

Primus’s approach is to change that process:

If a single Bootstrapping operation can handle multiple inputs and outputs, why break complex logic into one small operation at a time?

AutoHoG automates this process by identifying logic that can be combined in Verilog circuits and then generating customized homomorphic gates.

The result is up to 5.7x lower latency.

The benchmark results are even more interesting:

With 8 consumer-grade RTX 5090 GPUs and a fully encrypted 64-bit balance, a complete private transfer takes about 90ms, with a throughput of more than 210 transactions per second.

What stands out to me here is the use of consumer-grade GPUs.

That’s because whether a technology can truly become widespread depends not only on how fast it runs, but also on how much it costs to run.

If privacy computing requires expensive specialized hardware, it will likely remain a capability available only to a small number of institutions for a long time.

What Primus is demonstrating now is a path toward enabling ordinary consumer-grade GPUs to handle this kind of computation, with near-linear throughput scaling as more GPUs are added.

There’s also an easily overlooked design choice behind this: the entire process uses exact integer arithmetic, ensuring that different nodes produce exactly the same encrypted results.

For decentralized networks, this may be even more important than simply pursuing speed. Performance can be optimized, but consensus must not diverge because of floating-point differences between hardware.

I’ve always believed that truly mature on-chain privacy shouldn’t just hide transaction amounts.

It should also allow financial activities such as lending, payments, and asset management to operate normally without exposing underlying sensitive data.

AutoHoG shows me that Primus is working to solve a more fundamental problem:

Not proving that FHE can provide privacy, but making FHE capable of handling the computational workloads of real financial applications.

From GPUs to future FPGAs and ASICs, if this technical path continues to prove viable, privacy could move from being an add-on feature in DeFi to becoming default infrastructure.

That’s why I believe Primus is worth keeping an eye on.

@primus_labs #Primus
View Original
post-image
This page contains third-party content and does not constitute any advice, nor does it represent Gate's endorsement of such views. For details, please see disclaimer.

  • 2

Add a comment
Add a comment

Comment 2
ModAssistant
I still don’t quite understand how AutoHoG is actually implemented. What do you think, author?
MACDHistogramDyer
A 5.7x latency improvement looks pretty impressive, but if a consumer GPU takes 90ms per transaction, could it still fail to keep up under high concurrency? I understand the logic behind Bootstrapping merging, but in volatile markets, could this fixed-batch approach actually become a bottleneck.