Nov 2026 – Present · In progress
Scalable DLRM Inference across GPUs and CPU Memory
Recommendation models have embedding tables too big for a single GPU. This project spreads inference across multiple GPUs and host CPU memory, so models that would normally run out of memory can still be served at scale.
Website → · Source →
Oct 2026 – Present · In progress
FPGA DRAM Bandwidth Study on the Ultra96-V2
LLM decoding is limited by memory bandwidth, not compute. This project pushes an Ultra96-V2 FPGA toward its DRAM limit on a large fully connected layer and measures what actually moves the needle: memory access settings, clock speed, port configuration and weight precision.
Feb 2026 – May 2026 · Completed
DAG Operator Fusion Scheduler
In deep neural networks, shuffling data between fast memory and DRAM often costs more than the math itself. This scheduler decides which operators to fuse and which tensors to keep close, slashing data movement across large DNN graphs.
Details
- Scaled schedule search with graph-cut decomposition, reducing solver wall time by 88% (43s to 5s) with no loss in schedule quality.
- Fused MatMul→Pointwise chains with in-situ output activation reuse, keeping intermediates in fast memory and admitting fusion candidates the baseline wrongly pruned.
- Designed a retain-aware DP that jointly chooses fusion boundaries and which tensors stay resident, contributing ~25% latency reduction on memory-bound benchmarks within a combined 35.8% end-to-end improvement.
Poster (PDF) → · Source →
Jan 2026 – Apr 2026 · Completed
Power-Efficient RV32I Pipelined Processor
A 7-stage pipelined RISC-V CPU in SystemVerilog, optimized for both performance and performance-per-watt.
Details
- Architected the pipeline in RTL with hazard detection, data forwarding and branch resolution.
- Collaborated on a GShare branch predictor and write-back D-cache, contributing to a 26% geomean IPC improvement.
- Designed a speculative Return Address Stack, cutting JALR mispredictions on FFT from 99.8% to 35%.
- Reduced the critical path and tuned the cache for a 45% MIPS/Watt gain and 60% higher clock frequency over baseline.
Oct 2025 – Nov 2025 · Completed
Dynamic Memory Allocator
A from-scratch 64-bit malloc/free in C, tuned for memory utilization and throughput.
Details
- Built segregated free lists with best-fit search, reaching 74.5% utilization at 6,705 Kops throughput.
- Designed 16-byte mini-blocks with bit-packed header metadata, improving utilization by 24% over the implicit-list baseline.
- Implemented boundary-tag coalescing for adjacent free blocks, plus a runtime heap checker validating alignment, free-list consistency and block invariants.