# Amir Yazdanbakhsh

> Research Scientist at Google DeepMind · Co-founder & Co-lead, Machine Learning for Computer Architecture · Co-founder, MLArchSys (co-located with ISCA since 2020).

## Biography

Amir Yazdanbakhsh is a Research Scientist at Google DeepMind, where he co-founded and co-leads the Machine Learning for Computer Architecture team. His research bridges machine learning, computer architecture, and systems—spanning autonomous AI agents that write, optimize, and formally verify performance-critical software and hardware, ML for systems, automated code generation, and developer efficiency. Much of his recent work centers on agentic code and system optimization. ECO, an LLM-driven optimizer deployed across Google's warehouse-scale production fleet, has landed 6,400+ commits and reclaimed the equivalent of hundreds of thousands of CPU cores. He also leverages AlphaEvolve to discover novel compiler heuristics, interprocedural code layouts, and FHE kernels for TPUs. He conceived and led the multi-agent effort that built a cycle-accurate RISC-V CPU simulator from scratch capable of booting a full operating system. He co-authored Self-Refine, one of the earliest works demonstrating that large language models can iteratively improve their own outputs through self-feedback, cited 7,000+ times. Prior to DeepMind, he completed the Google Brain AI Residency, where he developed scalable reinforcement learning systems for complex real-world tasks and applied machine learning to accelerator design. His work in hardware-software co-design has directly shaped multiple headline hardware features across several generations of Google's Tensor Processing Units (TPUs). His research has been recognized with induction into the ISCA Hall of Fame, multiple IEEE Micro Top Picks awards, CACM Research Highlights nominations, Distinguished Paper Awards, and elevation to Senior Member of IEEE and ACM. He co-founded MLArchSys, the flagship workshop on ML for architecture and systems, co-located with ISCA since 2020. He earned his Ph.D. in Computer Science from Georgia Tech, supported by the Microsoft Research PhD Fellowship and the Qualcomm Innovation Fellowship. The Beyond Research section of his site features Under the Hood (زیر کاپوت), a Farsi podcast on how AI actually works, from chips and GPUs to inference, for a general audience. Public contact email: ayazdan@google.com.

## Research Thrusts

- Agents & System Optimization
- Efficient LLMs
- ML for Architecture
- Hardware for AI
- LLM Reasoning

## Selected Publications

- [ECO: An AI-Driven Code Efficiency Optimizer for Warehouse Scale Computers](https://ayazdan.com/papers/eco) — OSDI 2026 (Deployed across Google's fleet) — [PDF](https://www.usenix.org/system/files/osdi26-lin-hannah.pdf) · [Project](https://www.usenix.org/conference/osdi26/presentation/lin-hannah)
  A production system that uses LLMs to optimize C++ code across Google's fleet: it localizes opportunities with fleet-wide profiling and embedding-based search over mined performance anti-patterns, then verifies edits with testing, LLM self-review and post-deployment monitoring. It has landed 6,400+ commits (99.5% without rollback), saving the equivalent of several hundred thousand normalized CPU cores.
- [Self-Refine: Iterative Refinement with Self-Feedback](https://ayazdan.com/papers/self-refine) — NeurIPS 2023 (7,000+ citations) — [PDF](https://proceedings.neurips.cc/paper_files/paper/2023/file/91edff07232fb1b55a505a9e9f6c0ff3-Paper-Conference.pdf) · [Project](https://selfrefine.info/)
  Self-Refine lets a single LLM improve its own outputs at test time by iteratively writing feedback on its draft and refining it — no extra training data, fine-tuning or reinforcement learning. Across 7 tasks from dialog to math reasoning with GPT-3.5, ChatGPT and GPT-4, it improves task performance by about 20% absolute on average.
- [Learning Performance-Improving Code Edits](https://ayazdan.com/papers/learning-performance-improving-code-edits) — ICLR 2024 (Spotlight; IEEE Micro Top Picks) — [PDF](https://arxiv.org/pdf/2302.07867) · [Project](https://pie4perf.com/)
  Introduces a dataset of 77,000+ human performance-improving C++ edit pairs with a gem5-based environment for reliable measurement, and adapts LLMs to high-level program optimization via retrieval-based prompting, chain-of-thought, performance-conditioned fine-tuning and self-play. The best combination reaches a 6.86× mean speedup, above the average human optimization (3.66×).
- [Concorde: Fast and Accurate CPU Performance Modeling with Compositional Analytical-ML Fusion](https://ayazdan.com/papers/concorde) — ISCA 2025 (IEEE Micro Top Picks) — [PDF](https://arxiv.org/pdf/2503.23076)
  Learns fast, accurate CPU performance models by combining simple analytical models — which bound the performance impact of each microarchitectural component — with ML over compact performance distributions, instead of simulating every instruction. It is more than five orders of magnitude faster than a cycle-level simulator with about 2% average CPI error, enabling design-space and sensitivity analyses that were previously infeasible.
- [LIA: A Single-GPU LLM Inference Acceleration with Cooperative AMX-Enabled CPU-GPU Computation and CXL Offloading](https://ayazdan.com/papers/lia) — ISCA 2025 (IEEE Computer Society Best Paper Award) — [PDF](https://dl.acm.org/doi/pdf/10.1145/3695053.3731092) · [Code](https://github.com/ece-fast-lab/ISCA-2025-LIA)
  A single-GPU LLM inference framework that offloads computation to Intel AMX-enabled Xeon CPUs (measured at 20–40 TFLOPS) and pairs CXL memory with DDR for throughput-oriented serving. With one H100 GPU it achieves up to 19× lower latency and 5.1× higher throughput than the latest single-GPU offloading framework, and CXL offloading adds a further 1.5× throughput.
- [MESA: Microarchitecture Extensions for Spatial Architecture Generation](https://ayazdan.com/papers/mesa) — ISCA 2023 (ISCA Hall of Fame) — [PDF](https://dl.acm.org/doi/pdf/10.1145/3579371.3589084)
  A CPU-side hardware block that monitors running threads, translates their machine code into configurations for an on-chip spatial accelerator, and reconfigures it dynamically — enabling transparent offloading without specialized compilers. On the Rodinia benchmarks it averages a 1.3× speedup and 1.8× better energy efficiency over a multicore CPU baseline.

## Additional Publications

- [SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads?](https://ayazdan.com/papers/swe-fficiency) — ICML 2026 — [PDF](https://arxiv.org/pdf/2511.06090) · [Project](https://swefficiency.com/)
  A repository-level code optimization benchmark of 498 real-workload tasks across nine widely used Python data-science, ML, and HPC repositories (including NumPy, pandas, and SciPy), built via an automated pipeline that validates expert PR speedups and unit tests. State-of-the-art agents achieve less than 0.23× the expert speedup on average, struggling with bottleneck localization, cross-function reasoning, and preserving correctness.
- [ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution](https://ayazdan.com/papers/parevo) — ICML 2026 — [PDF](https://arxiv.org/pdf/2603.02510) · [Code](https://github.com/WildAlg/ParEVO)
  An agentic evolutionary framework for synthesizing high-performance parallel algorithms over irregular data structures (sparse graphs, unbalanced trees, non-uniform meshes), combining the 13,820-task Parlay-Instruct corpus, fine-tuned models aligned with ParlayLib, and an evolutionary coding agent guided by compilers, race detectors, and profilers. On ParEval it achieves 106× average speedup (13.6× on irregular graph problems) and matches or beats human baselines by up to 4.1× on irregular kernels.
- [AI-Propeller: Warehouse-Scale Interprocedural Code Layout Optimization with AlphaEvolve](https://ayazdan.com/papers/ai-propeller) — CAIS 2026 Workshop — [PDF](https://arxiv.org/pdf/2606.00131)
  Uses the Magellan/AlphaEvolve agentic workflow to evolve Propeller's compiler heuristic into a fine-grained interprocedural code layout optimizer, paired with Vizier hyperparameter tuning and real hardware performance counters as the evolutionary reward. Evaluated on large warehouse-scale binaries already optimized with FDO and PLO, it delivers 0.23% to 1.6% performance improvements.
- [Adapting AlphaEvolve to Optimize Fully Homomorphic Encryption on TPUs](https://ayazdan.com/papers/adapting-alphaevolve-to-optimize-fully-homomorphic) — MLArchSys @ISCA 2026 — [PDF](https://arxiv.org/pdf/2605.14718)
  Applies AlphaEvolve's closed-loop LLM code evolution with real hardware execution feedback and correctness testing to co-optimize Fully Homomorphic Encryption kernels across TPU matrix (MXU) and vector (VPU) units. Within 24 hours on Google Cloud TPUv5e, it improves TFHE (Jaxite) bootstrap latency by 2.5× and CKKS (CROSS) rotation and multiplication latency by 1.31× and 1.18× over human-engineered baselines.
- [Magellan: Autonomous Discovery of Novel Compiler Optimization Heuristics with AlphaEvolve](https://ayazdan.com/papers/magellan) — C4ML @CGO 2026 — [PDF](https://arxiv.org/pdf/2601.21096)
  An agentic framework coupling an LLM coding agent, evolutionary search, and autotuning to synthesize compact, executable C++ compiler heuristics evaluated directly on macro-benchmarks. In LLVM function inlining it outperforms decades of hand-crafted heuristics for both binary size and performance, matches expert register-allocation priority rules on large workloads, and demonstrates portability to XLA.
- [DSTBench: Distributed Systems Traces-Driven Agentic Performance Debugging in Microservices](https://ayazdan.com/papers/dstbench) — IC2E 2026 — [PDF](https://www.bu.edu/peaclab/files/2026/10/DSTBench-camera_ready_final_accessible.pdf) · [Code](https://github.com/ai4cloudops/DSTBench)
  A live-cluster benchmark and harness for evaluating LLM agents on operational microservice performance debugging, combining DSTFaults (eight reproducible gray-failure categories across four microservice apps with pre-fault baselines) and DSToon (compact causal-DAG trace serialization plus telemetry pagination). The harness raises AIOpsLab detection accuracy from 18% to 80% and Flash detection from 62% to 88%, with up to 27% localization accuracy.
- [CodeRosetta: Pushing the Boundaries of Unsupervised Code Translation for Parallel Programming](https://ayazdan.com/papers/coderosetta) — NeurIPS 2024 — [PDF](https://arxiv.org/pdf/2410.20527) · [Project](https://coderosetta.com/) · [Code](https://github.com/tehranixyz/CodeRosetta)
  An unsupervised encoder–decoder transformer trained with tailored pretraining and code-semantics objectives to translate bidirectionally between programming languages and their parallel HPC extensions (C++ ↔ CUDA and Fortran ↔ parallel C++). It improves C++→CUDA compilation accuracy by 6.05% over prior state of the art (and +22.08 BLEU / +14.39 CodeBLEU over closed-source LLMs) and improves Fortran→C++ CodeBLEU by at least 4.63 points.
- [OPTIMA: Optimal One-Shot Pruning for LLMs via Quadratic Programming Reconstruction](https://ayazdan.com/papers/optima) — ACML 2026 — [PDF](https://arxiv.org/pdf/2512.13886) · [Code](https://github.com/paramathic/optima)
  A scalable one-shot post-training LLM pruning method that formulates layer-wise weight reconstruction after mask selection as independent row-wise Quadratic Programs sharing a single layer Hessian, solved in parallel on accelerators without fine-tuning. Integrating with existing mask selectors, it improves zero-shot accuracy by up to 3.97% absolute and prunes an 8B transformer on a single H100 within 60 GB peak memory.
- [Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding](https://ayazdan.com/papers/learning-to-keep-a-promise) — ICML 2025 — [PDF](https://arxiv.org/pdf/2502.11517)
  Introduces PASTA, a learning-based asynchronous decoding system where LLMs are fine-tuned to emit PASTA-LANG annotations marking semantically independent parts of their own responses so an inference-time interpreter can decode chunks in parallel. On AlpacaEval, it Pareto-dominates existing parallel decoding baselines with 1.21×–1.93× geometric-mean speedups.
- [SLiM: One-Shot Quantization and Sparsity with Low-Rank Approximation for LLM Weight Compression](https://ayazdan.com/papers/slim) — ICML 2025 — [PDF](https://arxiv.org/pdf/2410.09615) · [Code](https://github.com/Mohammad-Mozaffari/slim)
  A unified one-shot LLM compression framework that combines probabilistic uniform weight quantization (SLiM-Quant), semi-structured 2:4 pruning, and mathematically computed saliency-based low-rank adapters to offset compression error without retraining. It improves 2:4 + 4-bit accuracy on LLaMA-2-7B by up to 5.66%, cuts end-to-end memory to 0.23×, and reaches up to 4.3× / 3.8× speedups on RTX 3060 and A100 GPUs.
- [SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity](https://ayazdan.com/papers/sparselora) — ICML 2025 — [PDF](https://arxiv.org/pdf/2506.16500) · [Project](https://z-lab.ai/projects/sparselora/) · [Code](https://github.com/z-lab/sparselora)
  Accelerates LLM fine-tuning through contextual sparsity using a lightweight, training-free SVD sparsity estimator that dynamically selects sparse weight subsets for loss and gradient computation while accounting for layer, token, and step sensitivity. Across reasoning, coding, and instruction-following tasks, it cuts compute by up to 2.2× and achieves up to 1.6× measured fine-tuning speedup with preserved accuracy.
- [Effective Interplay between Sparsity and Quantization: From Theory to Practice](https://ayazdan.com/papers/effective-interplay-between-sparsity-and-quantization) — ICLR 2025 (Spotlight) — [PDF](https://arxiv.org/pdf/2405.20935)
  Gives the first mathematical proof that sparsity and quantization are not orthogonal: combining them compounds error, and order matters — quantizing before sparsifying can remove important weights. Experiments on OPT and LLaMA (125M–8B), ViT and ResNet confirm the analysis and yield best practices for efficient deployment.
- [The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws](https://ayazdan.com/papers/the-journey-matters) — ICLR 2025 — [PDF](https://arxiv.org/pdf/2501.12486) · [Project](https://sparsellm.com/)
  Systematically explores 80 sparse pre-training schedules across sparsity levels and compute budgets, finding that pruning between 25% and 75% of training compute achieves near-optimal loss. Proposes a unified scaling law replacing parameter count in Chinchilla with average parameter count over pre-training, accurately predicting loss for both sparse and dense models while enabling smaller, cheaper-to-serve models.
- [SLOPE: Double-Pruned Sparse Plus Lazy Low-Rank Adapter Pretraining of LLMs](https://ayazdan.com/papers/slope) — ICLR 2025 — [PDF](https://arxiv.org/pdf/2405.16325) · [Code](https://github.com/Paramathic/slope)
  A sparse LLM pretraining method that introduces double-pruned N:M sparsity in the backward pass (pruning transposed weights for faster sparse gradients) and adds lazy low-rank adapters only in the final 1% of pretraining iterations to recover accuracy. On OPT-33B/66B, it accelerates training by up to 1.25× and inference by up to 1.54× while reducing training and inference memory footprints to 0.63× and 0.61×.
- [Progressive Gradient Flow for Robust N:M Sparsity Training in Transformers](https://ayazdan.com/papers/progressive-gradient-flow-for-robust) — CPAL 2025 (Oral) — [PDF](https://arxiv.org/pdf/2402.04744) · [Code](https://github.com/abhibambhaniya/progressive_gradient_flow_nm_sparsity)
  Shows that standard N:M structured-sparsity training recipes lose quality at high sparsity (>80%) because of noisy gradients, and progressively restricts gradient flow to pruned weights with decay mechanisms. This improves model quality by up to 2% (vision) and 5% (language), and up to 2% at equal training FLOPs.
- [QuArch: A Benchmark for Evaluating LLM Reasoning in Computer Architecture](https://ayazdan.com/papers/quarch) — ICML 2026 (Best of CAL at HPCA 2026) — [PDF](https://arxiv.org/pdf/2510.22087) · [Project](https://quarch.ai/)
  The first benchmark for LLM knowledge and reasoning in computer architecture: 2,671 expert-validated question–answer pairs spanning processor design, memory systems and interconnection networks. Frontier models score 34–73% on advanced questions and struggle most with higher-order architectural reasoning; the benchmark and leaderboard are public at quarch.ai.
- [ArchGym: An Open-Source Gymnasium for Machine Learning Assisted Architecture Design](https://ayazdan.com/papers/archgym) — ISCA 2023 — [PDF](https://arxiv.org/pdf/2306.08888) · [Code](https://github.com/srivatsankrishnan/oss-arch-gym) · [Blog](https://research.google/blog/an-open-source-gymnasium-for-machine-learning-assisted-computer-architecture-design/)
  An open-source gymnasium connecting diverse ML search algorithms to computer architecture simulators for fair, reproducible design-space exploration across memory controllers, DNN accelerators, and AR/VR SoCs (21,000+ experiments). Shows that with tuned hyperparameters ('hyperparameter lottery') diverse search algorithms perform comparably, and demonstrates an ArchGym-trained proxy cost model with 0.61% RMSE and 2,000× faster simulation.
- [GRANITE: A Graph Neural Network Model for Basic Block Throughput Estimation](https://ayazdan.com/papers/granite) — IISWC 2022 — [PDF](https://arxiv.org/pdf/2210.03894) · [Code](https://github.com/google/gematria)
  A graph neural network model that estimates x86-64 basic-block throughput by modeling instruction structural and data dependencies directly as graphs, paired with multi-task decoders across microarchitectures. It achieves 6.9% average test error (1.7% lower than prior work) while running training and inference roughly 3.0× faster.
- [An Evaluation of Edge TPU Accelerators for Convolutional Neural Networks](https://ayazdan.com/papers/an-evaluation-of-edge-tpu-accelerators) — IISWC 2022 — [PDF](https://arxiv.org/pdf/2102.10423)
  Presents microarchitectural details and a large-scale characterization of three classes of Google Edge TPU accelerators (deployed in Coral/Pixel and pipeline designs) across 423,000 unique convolutional neural networks. Also introduces millisecond-latency learned ML models for estimating accelerator latency and energy as a fast alternative to cycle-accurate simulation for hardware–software co-design.
- [RAGO: Systematic Performance Optimization for Retrieval-Augmented Generation Serving](https://ayazdan.com/papers/rago) — ISCA 2025 — [PDF](https://arxiv.org/pdf/2503.14649) · [Code](https://github.com/google/rago)
  Introduces RAGSchema, a structured abstraction capturing diverse retrieval-augmented generation pipelines, characterizes wide workload variability across RAG variants, and presents RAGO, a systematic performance optimization framework for RAG serving. RAGO delivers up to a 2× increase in QPS per chip and a 55% reduction in time-to-first-token latency over LLM-serving extensions.
- [DaCapo: Accelerating Continuous Learning in Autonomous Systems for Video Analytics](https://ayazdan.com/papers/dacapo) — ISCA 2024 (Distinguished Artifact Award) — [PDF](https://arxiv.org/pdf/2403.14353)
  A hardware–algorithm co-design for on-device continuous learning in video analytics that runs inference, labeling and retraining concurrently on a spatially partitionable, precision-flexible accelerator with spatiotemporal resource allocation. It achieves 6.5% and 5.5% higher accuracy than the GPU-based systems Ekya and EOMU while consuming 254× less power.
- [Tandem Processor: Grappling with Emerging Operators in Neural Networks](https://ayazdan.com/papers/tandem-processor) — ASPLOS 2024 (IEEE Micro Top Picks Honorable Mention) — [PDF](https://dl.acm.org/doi/pdf/10.1145/3620665.3640365) · [Code](https://actlab-genesys.github.io/)
  An on-chip companion processor to the GEMM unit of neural accelerators that efficiently executes the growing variety of non-GEMM operators, using a specialized ISA and microarchitecture that streamline memory access and express layers as primitive vector operations. Its synthesizable RTL and compiler are open source as part of the GeneSys project.
- [In-Storage Domain-Specific Acceleration for Serverless Computing](https://ayazdan.com/papers/in-storage-domain-specific-acceleration) — ASPLOS 2024 — [PDF](https://actlab-genesys.github.io/assets/pdf/dscs.pdf)
  Shows that remote disaggregated storage overhead (>55% of latency) caps conventional serverless acceleration, and proposes DSCS-Serverless—embedding a compact 25 W programmable domain-specific accelerator directly inside storage drives with peer-to-peer flash-to-DRAM transfers. Across eight serverless applications, it outperforms a 250 W disaggregated RTX 2080 Ti setup and delivers 3.7×/1.7× speedup and 4.3×/1.9× energy savings over ARM- and FPGA-based computational storage.
- [Sparse Attention Acceleration with Synergistic In-Memory Pruning and On-Chip Recomputation](https://ayazdan.com/papers/sparse-attention-acceleration-with-synergistic) — MICRO 2022 — [PDF](https://arxiv.org/pdf/2209.00606)
  Architects SPRINT, a hybrid accelerator that computes approximate self-attention scores in parallel inside ReRAM crossbars, prunes low scores via lightweight analog thresholding before fetching, and recomputes surviving scores digitally on chip to turn quadratic attention into near-linear work. Across transformer models with 16 KB on-chip memory, SPRINT achieves 7.5× average speedup and 19.6× energy reduction with only 0.36% average accuracy degradation.
- [Accelerating Attention through Gradient-Based Learned Runtime Pruning](https://ayazdan.com/papers/accelerating-attention-through-gradient-based-learned) — ISCA 2022 — [PDF](https://arxiv.org/pdf/2204.03227)
  Formulates discrete runtime attention-pruning thresholds as a soft differentiable regularizer co-optimized with model weights during backpropagation, paired with LeOPArd, a bit-serial transformer accelerator supporting bit-level early termination. Across 43 tasks (MemN2N, BERT, ALBERT, GPT-2, ViT), post-layout results show 1.9× average speedup and 3.9× energy reduction with under 0.2% accuracy degradation.
- [SnaPEA: Predictive Early Activation for Reducing Computation in Deep Convolutional Neural Networks](https://ayazdan.com/papers/snapea) — ISCA 2018 — [PDF](https://cseweb.ucsd.edu/~vakhlagh/ISCA18-SnaPEA.pdf)
  Exploits ReLU's zeroing of negative convolution outputs via weight sign-reordering and runtime partial-sum sign checks to terminate CNN convolutions early in exact (lossless) or threshold-controlled predictive modes. Over Eyeriss, SnaPEA achieves 28% average speedup and 16% energy reduction with zero accuracy loss in exact mode, and 2.02× average speedup (up to 3.59×) and 1.89× energy savings (up to 3.14×) in predictive mode within 3% accuracy loss.
- [General-Purpose Code Acceleration with Limited-Precision Analog Computation](https://ayazdan.com/papers/general-purpose-code-acceleration-with-limited) — ISCA 2014 (IEEE Micro Top Picks Honorable Mention) — [PDF](https://cseweb.ucsd.edu/~hadi/doc/paper/2014-isca-anpu.pdf) · [Retrospective](https://bpb-us-w2.wpmucdn.com/sites.coecis.cornell.edu/dist/7/587/files/2023/06/St_Amant_2014_General-Purpose.pdf)
  A circuit-to-compiler approach that converts approximable code regions into analog neural networks, tackling limited precision, noise and range constraints with circuit techniques, a hardware/software interface, training methods and compiler support. It delivers a 3.7× whole-application speedup and 6.3× energy savings with under 10% quality loss on all but one benchmark.
- [What Makes Chain-of-Thought Prompting Effective? A Counterfactual Study](https://ayazdan.com/papers/what-makes-chain-of-thought-prompting) — EMNLP 2023 — [PDF](https://aclanthology.org/2023.findings-emnlp.101.pdf) · [Code](https://github.com/reasoning-machines/prompt-lib)
  Uses counterfactual prompting to systematically ablate symbols, patterns, and text in few-shot Chain-of-Thought prompts across three LLMs. Finds that specific symbols matter little while consistent structural patterns and natural web-style text are critical—and that intermediate steps primarily communicate what the task is and help models fill in missing commonsense knowledge rather than teaching how to solve it.

## Selected Writing, Talks & Press

- [ArchGym: An Open-Source Gymnasium for ML-Assisted Computer Architecture Design](https://research.google/blog/an-open-source-gymnasium-for-machine-learning-assisted-computer-architecture-design/) — Google Research Blog
- [Offline Optimization for Architecting Hardware Accelerators](https://research.google/blog/offline-optimization-for-architecting-hardware-accelerators/) — Google Research Blog
- [Machine Learning for Computer Architecture](https://research.google/blog/machine-learning-for-computer-architecture/) — Google Research Blog
- [Massively Large-Scale Distributed Reinforcement Learning with Menger](https://research.google/blog/massively-large-scale-distributed-reinforcement-learning-with-menger/) — Google Research Blog
- [Teamwork: When AI Becomes a Research Partner (RISC-V CPU Simulator)](https://antigravity.google/blog/teamwork-when-ai-becomes-a-research-partner) — Google Antigravity Blog
- [OCP Keynote Talk](https://www.youtube.com/watch?v=F5Eieaz7u1I) — Open Compute Project
- [MLArchSys Workshop](https://sites.google.com/view/mlarchsys) — Co-founder (ISCA since 2020)
- Press coverage: [WIRED](https://www.wired.com/story/fit-billions-transistors-chip-let-ai-do/) · [ZDNet](https://www.zdnet.com/article/googles-deep-learning-finds-a-critical-path-in-ai-chips/) · [InfoQ](https://www.infoq.com/news/2021/03/google-ai-chip-design/)

## Podcast — Under the Hood (زیر کاپوت)

- Farsi podcast explaining how AI works from chips and GPUs to inference for a general audience.
- [Listen on Spotify](https://open.spotify.com/show/6dTQAH75geR7sAfPMBsow1) · [Instagram](https://www.instagram.com/zire.kapoot/) · [Telegram](https://t.me/zirekapoot)

## Profiles & Contact

- Website: https://ayazdan.com/
- Public email: ayazdan@google.com
- Office Hours (students, collaborations, media): https://ayazdan.com/#office-hours
- Google Scholar: https://scholar.google.com/citations?user=Vdu_sqwAAAAJ&hl=en
- ORCID: https://orcid.org/0000-0001-8199-7671
- OpenReview: https://openreview.net/profile?id=~Amir_Yazdanbakhsh1
- GitHub: https://github.com/yazdanbakhsh
- LinkedIn: https://www.linkedin.com/in/ayazdanb
- X: https://x.com/ayzddzya
