github/tfhe-rs
1
1
Fork 0
mirror of https://github.com/zama-ai/tfhe-rs.git synced 2026-01-09 14:47:56 -05:00
Code Issues Packages Projects Releases Wiki Activity

chore(hpu): adds a few words on HPU ERC20 based on GPU words

Browse Source
This commit is contained in:
pgardratzama
2026-01-02 15:38:51 +01:00
parent b365aa5a71
commit 501d44dbda
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tfhe/docs/getting-started/benchmarks/hpu/hpu-erc20.md
View File

@@ -0,0 +1,8 @@
Similarly to the [CPU benchmarks](../cpu/cpu-erc20.md), the latency and throughput of a confidential ERC20 token transfer can be measured.
![](../../../.gitbook/assets/hpu-hlapi-erc20-benchmark-hpux1-latency-throughput.svg)
The whitepaper version of the ERC20 is the same implementation than on CPU or GPU using a selection between the amount to be transferred and 0 based on result of comparison between this amount and from_amount followed by an addition and a subtraction. The hpu_optim version is very much the same processing but with a single HPU instruction (IOp) doing the complete ERC20 processing.
The hpu_simd version is a different measure which uses another HPU IOp called ERC20_SIMD (for Single Instruction Multiple Data). This ERC20_SIMD takes an input of 12 triplets (from, to, amount) and returns 12 pairs (new_from, new_to) and is particulartly efficient when having many independent transfers to execute.
The throughput shown here is the maximum that can be achieved with TFHE-rs on a single FPGA node running the HPU, in an ideal scenario where all transactions are independent.
In a blockchain protocol, the throughput would be limited by the latency of the network and the necessity to apply other operations (compression, decompression, rerandomization).