QuantumResearchers built an end-to-end quantum compiler for fermionic simulation
They claim the tool reduces gate counts by up to 80% and cuts compilation times by up to 99% on large simulation instances.
Papers
Preprints and accepted papers worth a scan in plain-language titles with one sentence of context, and a straight link to the source. The images are a little weird, on purpose, for funsies. (Because science is supposed to be fun). The AI bot date parser is so bad at dates, so just rejoice that at least the papers it finds are really neat ones even if they are older.
1–24 of 819 / 819 papers
Cloud & SystemsResearchers created Thinflow to manage HPC benchmarking campaigns
Thinflow is a lightweight Bash engine that organizes benchmark runs as directory trees with automated provenance tracking and Slurm submission.
Data & I/OResearchers built a parallel ingestion pipeline for scientific data archives
The S3Importer tool uses TANGO controls and parallel workers to ingest nearly 200,000 HDF5 and FITS records into PostgreSQL and object storage.
NetworksResearchers built SLED to isolate sensor streams from edge to HPC
The framework uses radio network slicing to keep bulk sensor data from delaying latency-critical telemetry sent to HPC clusters.
AI & MLResearchers built neural operators for online bias correction in E3SM
The plug-in neural operators corrected atmospheric biases during E3SM runs while remaining stable over multi-year simulations.
AI & MLResearchers built an AI tool to find datacenter hardware specialization targets
They claim its proposed hardware extensions speed up simulated datacenter workloads by an average of 12.6% and up to 80.2%.
SimulationResearchers mapped multi-request VLM serving onto AMD Versal chips
They claim the setup decodes up to 1.3x faster than an Nvidia Tesla T4 while using roughly 39% less energy per token.
Data & I/OResearchers sped up distributed Ising machines across GPUs, FPGAs, and ASICs
They report data placement and relaxed scheduling cut time-to-solution up to 3.16x and tested a custom 28nm ASIC sustaining 28.1 Gflips/s.
AcceleratorsResearchers designed an accelerator for vision-language-action models
They claim their custom accelerator runs GR00T action heads 1.7x faster than Jetson Thor in NVFP4 while cutting energy use by 9.1x.
AcceleratorsResearchers reverse-engineered hidden Nvidia SASS controls to boost GPU speed
They found undocumented control fields in cuBLAS and built a binary optimizer that speeds up GPU kernels by up to 31.9%.
SimulationResearchers open-sourced a gem5 port supporting CHERI hardware
The tool models microarchitectural details of CHERI-enabled CPUs and can boot CheriBSD to measure memory-safety performance overheads.
AI & MLResearchers proposed ultra-low-voltage logic for energy-efficient AI chips
They claim ultra-low-voltage circuits resist failure via gate clusters, potentially cutting AI compute energy demands.
AcceleratorsResearchers built CriticNPU to speed up AI chip design exploration
They claim CriticNPU cuts hardware simulation calls in half while finding top-performing configurations for LLM inference.
AcceleratorsGPCPlace cuts crossbar port contention on partitioned GPUs
The authors claim dynamic GPC placement speeds up latency-critical GPU tasks by up to 2.38x over standard CUDA Green Contexts.
AcceleratorsResearchers converted attention models into hard-logic FPGA circuits
They claim turning transformer layers into truth-table logic runs in 1-2 µs on FPGAs, up to three orders of magnitude faster than a GPU at batch 1.
SimulationResearchers designed LeafPack to compress page-table lines in CPU caches
They claim LeafPack packs up to 16 page-table lines into one 80-byte cache slot, boosting simulated server-core performance by up to 28.4%.
SimulationResearchers proposed CAIMS for scalable in-memory similarity search
They say the multi-array design hits up to 22.3x higher throughput and 426x better energy efficiency than a server CPU on RAG retrieval workloads.
AcceleratorsResearchers used C++ templates to cut FPGA LUT usage in HLS
They claim their C++ template library cuts LUT usage on AMD Versal by up to 4.5x for matrix multiplication without modifying the HLS compiler.
AcceleratorsResearchers built a bottom-up FPGA overlay generator
They claim their bottom-up overlays run at up to 400 MHz on an AMD VCK5000, achieving a 10x speedup over OverGen with 20x less FPGA logic.
AI & MLResearchers built TripleKV to cut LLM cache stalls
They claim prefetching KV cache blocks from SSD to host DRAM cuts P95 time to first token by up to 49%.
Data & I/OResearchers built a job-aware I/O steering framework for HPC storage
They claim the method cuts OST load variance by 55% and reduces peak storage saturation duration by 87.5%.
AcceleratorsResearchers proposed ElsaGS accelerator for 3D Gaussian Splatting
They claim an algorithm-hardware co-design achieves up to a 20.6x speedup and 99.4% energy reduction over GPUs for 3DGS training.
AI & MLResearchers built an LLM framework for processor microarchitecture design
They claim KD-DSE cut ADRS by 21.7% and improved hypervolume by 4.8% over MicroEvo on an out-of-order core using just 18 simulations.
AcceleratorsResearchers built a bit-sparse accelerator for JEPA world model planning
They claim their planner-hardware co-design runs up to 1.8x faster and uses up to 52% less energy than dense accelerators on JEPA models.