I build memory systems for AI.

My current work is on high bandwidth flash (HBF), an emerging 3D-stacked NAND flash memory that puts terabytes of capacity next to the accelerator. With HBF, a single accelerator or a small node can serve larger language models.

Before Huawei, I did my Ph.D. and a postdoc in Onur Mutlu’s SAFARI Research Group at ETH Zürich. I worked on processing-in-memory (PIM) across the stack, from computing inside DRAM chips to the programming models and compilers that make it usable.

Recent papers

All publications
  1. HPCA 2027To appear

    FLINT

    Efficiently Leveraging High Bandwidth Flash for Capacity-Scalable LLM Inference Acceleration

    FLINT holds LLM weights in HBF, next to HBM. It reads the flash at high bandwidth, keeps refresh off the inference path, and replaces SSD-style support for arbitrary writes with a compact read-only translation table.

    PaperLetter, IEEE CAL 2026

  2. MICRO 2026

    PipeDRAM

    A Data-Transposition-Free Processing-Using-DRAM Architecture with Hardware/Software Pipelining

    PaperCode

  3. ICCD 2026

    DaPPA

    A Unified Data-Parallel Programming Framework for Efficient Processing-in-Memory Execution

    PaperCode

  4. ICS 2025

    Proteus

    Achieving High-Performance Processing-Using-DRAM via Dynamic Precision Bit-Serial Arithmetic

    PaperCodeTalk

Research

Moving data between processors and memory limits both performance and energy efficiency. My work has three threads: memory systems for AI, understanding data movement, and computing inside memory. With colleagues, I also work on DRAM reliability and security, near-data processing for applications, and programming and system support for PIM.

Memory systems for AI

LLM inference is increasingly limited by the memory capacity of the accelerator rather than by its compute. At Huawei, I work on adding that capacity with high bandwidth flash.

  1. FLINTHPCA 2027

    Stores LLM weights in HBF, next to HBM. A hardware burst-buffer controller coalesces and pipelines HBF reads, using the existing NAND flash buffers, to keep bandwidth high. A phantom-plane refresh mechanism moves refresh off the critical inference path, and a read-only flash translation layer maps the weights to HBF locations with a compact table.

    PaperLetter, IEEE CAL 2026

Understanding data movement

Before changing the hardware, I measure where time and energy go, in real applications and on real systems.

  1. DAMOVIEEE Access 2021

    A methodology and benchmark suite for data movement bottlenecks. We characterized 77K functions across 345 applications, classified them by the sources of their data movement, and selected 144 of them as the first open-source benchmark suite for main-memory data movement studies. A case study with near-data processing shows which fix suits each kind of bottleneck.

    PaperCodeTalk

Computing inside memory

Processing-using-DRAM computes with the DRAM arrays themselves, so the data never leaves the memory chip. SIMDRAM, MIMDRAM, Proteus, and PipeDRAM each remove a limitation of the one before.

1 in A or B 1 in the carry or S 0 Activated rows

Each column stores two 4-bit numbers, A and B, one bit per row. Activating the rows of one bit position computes that bit of the sum S, and the new carry, in every column at once. A subarray has thousands of columns, so it adds thousands of pairs without moving the data off the chip.

SIMDRAM, MIMDRAM, and Proteus compute on this vertical layout. PipeDRAM computes on the horizontal layout that CPUs use, so data is never transposed.

  1. SIMDRAMASPLOS 2021

    Builds arbitrary operations from in-DRAM majority and NOT, and runs them one bit at a time across thousands of DRAM columns at once. With 16 DRAM banks, it provides 88× the throughput of a CPU and 5.8× that of a high-end GPU, on average across 16 operations.

    PaperLecture

  2. MIMDRAMHPCA 2024

    Lets different segments of a DRAM row run different operations at the same time, using fine-grained DRAM access: a multiple-instruction multiple-data model inside each subarray. Compiler passes map ordinary loops onto it. It provides 34× the performance and 14.3× the energy efficiency of SIMDRAM, on average.

    PaperCode

  3. ProteusICS 2025

    Lowers the bit precision of each operation at run time to what its values need. It also runs independent steps across several DRAM arrays at once, and picks the data representation and arithmetic algorithm for each instruction. With a single DRAM bank, it provides 10.2× the performance per area of SIMDRAM and 8.1× lower energy, on average.

    PaperCodeTalk

  4. PipeDRAMMICRO 2026

    Computes directly on data in the horizontal layout that CPUs use, so data is never transposed. The memory controller reorganizes the bits inside each memory request, and a pipelined execution model overlaps bit-dependent and bit-independent operations across the DRAM array. It provides 11.8× the performance of both SIMDRAM and MIMDRAM, and 80.4× that of Proteus, on average.

    PaperCode

  1. DaPPAICCD 2026

    A programming framework for these systems. Programmers express a computation once, with data-parallel patterns and a dataflow interface, and DaPPA’s back-ends generate the data orchestration for a real near-memory system (UPMEM) and the multi-level tiling for processing-using-DRAM. On UPMEM, it is 1.6× faster than hand-tuned code, with 95% fewer lines of code, on average.

    PaperCode

With colleagues

  1. DRAM reliability and security

    Read disturbance (RowHammer) in real DRAM chips and defenses against it: Chronus, Variable Read Disturbance, Svärd, ABACuS, and RowHammer under reduced wordline voltage; and error profiling for on-die ECC with HARP. Chronus (HPCA 2025) received the 2026 Intel Hardware Security Academic Award.

  2. Near-data processing for applications

    Hybrid transactional and analytical databases (Polynesia), raw-signal genome analysis inside storage (MARS), stencil computations near the cache (Casper), and an overview of the field, A Modern Primer on Processing in Memory.

  3. Programming and system support

    Besides DaPPA: transcendental functions for PIM (TransPimLib), machine learning workloads on a real PIM system (ISPASS 2023), CPU-memory data movement in memory-centric systems (ALP), and performance prediction for near-memory computing (NAPEL).

Open-source software

I release the code and simulators behind my main systems.

DAMOV
Benchmark suite and methodology for data movement bottlenecks
Ramulator-PIM
Simulator for processing-in-memory
MIMDRAM
gem5 model and LLVM compiler passes
Proteus
Code for dynamic-precision arithmetic in DRAM
DaPPA
The programming framework, with its UPMEM and processing-using-DRAM back-ends
PipeDRAM
Code for transposition-free processing-using-DRAM

Publications

* Co-first authors.

Talks and tutorials

Memory-Centric Computing Systems tutorials

Co-organizer, with Onur Mutlu and the SAFARI Research Group.

  1. ISCA 2025

    Tokyo, 21 June 2025

  2. ICS 2025

    Salt Lake City, 8 June 2025

  3. ASPLOS 2025

    Rotterdam, 30 March 2025

  4. PPoPP, HPCA, and CGO 2025

    Las Vegas, 1 March 2025

  5. MICRO 2024

    Austin, 2 November 2024

  6. ISCA 2024

    Buenos Aires, 29 June 2024

  7. HEART 2024

    Porto, 21 June 2024

Invited talks

2025
IBM Research Zurich
2024
Apple, AMD, Arm (Austin and Cambridge), SK hynix, Rambus, Stanford University, UT Austin, Texas State University, ACCESS Seminar, EFCL Summer School, DOT-PIM Workshop at ESWEEK, ABUMPIMP Minisymposium in Madrid (video), and Memory & Storage Summit in Bengaluru (slides)
2023
University of Toronto, Huawei Toronto

Conference talks

  1. ICS 2025

    Proteus: Achieving High-Performance Processing-Using-DRAM via Dynamic Precision Bit-Serial Arithmetic

    Salt Lake City

    Video

  2. HPCA 2024

    MIMDRAM: An End-to-End Processing-Using-DRAM System for High-Throughput, Energy-Efficient and Programmer-Transparent Multiple-Instruction Multiple-Data Processing

    Edinburgh; also at SRC TECHCON 2024, Austin

  3. FMS 2024

    Software Frameworks for Productive and Effective Processing-in-Memory Architectures

    Santa Clara

  4. FMS 2024

    Methodologies, Workloads, and Tools for Processing-in-Memory: Enabling the Adoption of Data-Centric Architectures

    Santa Clara; also at ISVLSI 2022, Nicosia

  5. SRC TECHCON

    pLUTo: Enabling Massively Parallel Computation in DRAM via Lookup Tables

    Austin

    Video

  6. MICRO 2023

    PUMA: Efficient and Low-Cost Memory Allocation and Alignment Support for Processing-Using-Memory Architecture

    Student Research Competition, Toronto

  7. ISVLSI 2022

    Heterogeneous Data-Centric Architectures for Modern Data-Intensive Applications: Case Studies in Machine Learning and Databases

    Nicosia

    Video

  8. ICDE 2022

    Polynesia: Enabling High-Performance and Energy-Efficient Hybrid Transactional/Analytical Databases with Hardware/Software Co-Design

    Virtual

    Video

  9. PACT 2021

    Google Neural Network Models for Edge Devices: Analyzing and Mitigating Machine Learning Inference Bottlenecks

    Virtual

  10. SRC TECHCON 2019

    Understanding Data Movement Bottlenecks in Modern Data-Intensive Workloads

    Austin

Experience and education

  1. 2026 to now

    Senior Researcher

    Huawei Zurich Research Center

  2. 2025 to 2026

    Postdoctoral Researcher

    SAFARI Research Group, ETH Zürich, with Onur Mutlu

  3. 2018 to 2025

    Ph.D. in Computer Science

    ETH Zürich, advised by Onur Mutlu. Thesis: New Tools, Programming Models, and System Support for Processing-in-Memory Architectures.

  4. 2019

    Hardware Engineering Intern

    Google, San Francisco Bay Area

  5. 2017 to 2018

    Research Assistant

    SAFARI Research Group, ETH Zürich

  6. 2016 to 2017

    M.S. in Computer Science

    Federal University of Rio Grande do Sul, Brazil, advised by Luigi Carro. Thesis: A Generic Processing in Memory Cycle Accurate Simulator under Hybrid Memory Cube Architecture.

  7. 2014 to 2015

    Exchange student in Computer Engineering

    Portland State University, USA, with a Brazil Scientific Mobility Program scholarship

  8. 2012 to 2015

    B.S. in Computer Science

    Federal University of Viçosa, Brazil

Awards
2026 Intel Hardware Security Academic Award (Recent Publication Award), for Chronus, HPCA 2025.
Service
Reviewer for IEEE Transactions on Computers. Subreviewer for ISCA, MICRO, ASPLOS, HPCA, PACT, DSN, EuroSys, USENIX ATC, CCS, PLDI, ISPASS, ISCAS, MSST, NVMW, IEEE CAL, TCAD, TVLSI, TCSI, TED, and Nature Electronics.
Tools
Languages: C/C++, Python, Verilog, and SystemC. Simulators: gem5, ZSim, and Ramulator. Compilers and profilers: LLVM, perf, and VTune.