Ephrem Wu

Advanced Micro Devices (United States)

Unclaimed · AI-structured profileOpenAlex

AMD researcher mapping FlashAttention and LLM decoding workloads onto NPUs via open-source MLIR compilers

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

No output score yet: this snapshot provides no publication counts, venue list, or per-year output totals for Ephrem Wu. The axis would measure cumulative and annual paper volume across venues once a fuller bibliographic record is linked.

Impact—

Impact is unscored: no citation counts, h-index, or referencing works are available in the input data. This axis would measure citations to the two cataloged arXiv papers and their uptake once citation coverage exists.

Focus100/100

Focus scores a perfect 100 because every field-classified paper in the catalog sits in a single field — computer architecture and hardware-accelerated ML systems (field distribution: 1 entry, count 1). Both 2026 papers reinforce that read: one maps FlashAttention onto AMD XDNA NPUs via compiler tooling, the other analyzes speculative decoding for block-diffusion models. The caveat is a thin base: only one paper carries a formal field label, so the concentration is clear but measured on minimal evidence.

Collaboration39/100

Collaboration scores 39, reflecting a compact working circle: 6 distinct coauthors are on record — Erwei Wang, Victor J. B. Jung, Jiajie Li, Andre Rosti, Joseph Melber, and Samuel Bayliss — consistent with small, focused industry teams at AMD rather than large academic consortia. The pattern splits across two modes: a middle-author position on the multi-author NPU compiler paper and first-plus-corresponding authorship on the speculative-decoding paper, suggesting both team contributor and lead-analyst roles.

Activity—

Activity cannot be scored: with only a single publication year (2026) in scope, there is no multi-year cadence to evaluate. This axis would measure publication frequency and recent-output velocity across consecutive years.

Reproducibility55/100

Reproducibility scores 55, computed as a mean over just one assessed paper (reproMean 55, reproPapers 1). That paper — the XDNA NPU FlashAttention study — carries a medium repro-readiness rating, helped by its reliance on the open-source MLIR-AIR and IRON toolchain, yet Evolve's verdict flags thin measurement rigor and weak baselines alongside the 3.62 TFLOP/s fused-kernel result. With a single-paper base, this score is indicative rather than settled.

Research summary

Ephrem Wu is a researcher at Advanced Micro Devices (AMD) in the United States working at the intersection of computer architecture and hardware-accelerated machine-learning systems. The Evolve catalog holds two 2026 arXiv papers. The first, a middle-author collaboration titled 'Programming AMD XDNA NPUs with Open-source Compiler Tools: A FlashAttention Case Study', maps FlashAttention onto AMD XDNA NPUs through the open-source MLIR-AIR and IRON compiler stack, characterizing four mappings with roofline analysis; the fused kernel reaches 3.62 TFLOP/s — 2× the IRON mapping and 6.5× better energy than an iGPU — though Evolve's review flags thin measurement rigor and baselines. The second, a first-author and corresponding-author paper, 'Ceiling-Clipped Acceptance Histograms Indicate Stranded Speed-up in Block-Diffusion Speculative Decoding', argues that ceiling-clipped acceptance histograms leave achievable speed-up unrealized; it awaits full catalog review. The bibliometric picture is narrow but sharp: focus scores 100, with all classified output in one field; collaboration scores 39 across 6 distinct coauthors; reproducibility scores 55 from a single assessed paper with medium readiness. Output, impact, and activity axes currently lack data.

Collaboration pattern

Ephrem Wu's recorded collaborations are compact: 6 distinct coauthors, all tied to the AMD orbit — Erwei Wang, Victor J. B. Jung, Jiajie Li, Andre Rosti, Joseph Melber, and Samuel Bayliss. The two 2026 papers show two modes of participation: a middle-author role on the team-built NPU compiler case study, and first-plus-corresponding authorship on the speculative-decoding analysis, i.e., contributing inside an engineering group while also driving independent analyses.

Research keywords

FlashAttentionAMD XDNA NPUNPU programmingMLIR-AIRIRON compilerspatial architecture mappingroofline analysiskernel fusionspeculative decodingblock diffusion modelsLLM inference accelerationperformance benchmarking

Highlights

Fused FlashAttention kernel reaches 3.62 TFLOP/s on AMD XDNA NPUs
In a 2026 arXiv case study, the team maps FlashAttention onto AMD XDNA NPUs using the open-source MLIR-AIR and IRON compiler stack, characterizing four distinct mappings with roofline analysis; the fused kernel delivers 3.62 TFLOP/s — 2× the IRON mapping and 6.5× better energy than an iGPU — though Evolve's review notes measurement rigor and baselines remain thin.
An open-source compiler path into NPU spatial architectures
The same study demonstrates that spatial-architecture mapping on AMD's XDNA NPUs can be done entirely through open-source tools (MLIR-AIR, IRON) rather than proprietary flows, with roofline analysis used to justify placement choices — an accessible entry point for kernel engineers targeting NPUs.
First-author analysis: stranded speed-up in block-diffusion speculative decoding
A 2026 first-author, corresponding-author arXiv paper argues that ceiling-clipped acceptance histograms in block-diffusion speculative decoding indicate stranded speed-up — acceptance rates pinned at a ceiling leave achievable acceleration unrealized. This paper is not yet fully reviewed in the Evolve catalog.

Timeline

2026Co-authored the arXiv case study 'Programming AMD XDNA NPUs with Open-source Compiler Tools: A FlashAttention Case Study' (middle authorship), mapping attention kernels onto XDNA NPUs via open-source compilers.
2026Led and published the first-author, corresponding-author arXiv paper 'Ceiling-Clipped Acceptance Histograms Indicate Stranded Speed-up in Block-Diffusion Speculative Decoding'.
2026Within a single year, output spans both sides of the AI hardware stack: compiler-level NPU kernel mapping on the systems side and acceptance-dynamics analysis for block-diffusion decoding on the algorithm side.

Papers in the Evolve catalog

2026-09-18
Programming AMD XDNA NPUs with Open-source Compiler Tools: A FlashAttention Case Study

Frequent co-authors

Erwei Wang×1 Victor J. B. Jung×1 Jiajie Li×1 Andre Rosti×1 Joseph Melber×1 Samuel Bayliss×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric scores are computed from a minimal snapshot: only two cataloged papers, both 2026 arXiv preprints, feed the radar.
Focus (100) and reproducibility (55) each rest on a single classified or assessed paper (fieldDist count 1; reproPapers 1), so both scores are thin-based.
Output, impact, and activity are unscored — no citation counts, publication totals, or multi-year coverage exist in the input data.
No ORCID or country data was provided; identity is name-resolved from public bibliographic data, and the second paper lacks Evolve review metadata (no verdict, keywords, or slug).