Sandeep Kumar

Research Scientist at Intel, Bangalore, India

Post Series

Curated reading paths through the technical posts. Each series is an ordered sequence — follow it top to bottom, or jump to any part.

LLM Inference from the Ground Up

From the end-to-end inference flow and vector math to serving optimizations and advanced model architectures.

  1. 1 LLM Inference Introduction: the end-to-end flow and vector math
  2. 2 Two Papers That Changed LLM Serving: FlashAttention and PagedAttention (deep dive)
  3. 3 Agentic LLMs and Mixture-of-Experts (MoE)

CPU Instructions for ML (Intel ISA Deep Dives)

How numeric precision, VNNI, and AMX map ML math onto Intel CPU instructions, capped by tensor-processing primitives for LLM inference.

  1. 1 Everything on Precision: Numeric Formats in Modern ML and HPC (deep dive)
  2. 2 VNNI (Vector Neural Network Instructions): A Tutorial (deep dive)
  3. 3 From Scalar to Tiles: Understanding Intel AMX Through a Hands-On GEMM Benchmark on Granite Rapids (deep dive)
  4. 4 Intel TPP for LLM Inference: How Tensor Processing Primitives Accelerate Every Transformer Block on CPU

CPU Microarchitecture & Memory

How modern CPUs move and share data: cache coherence, prefetching, and the IOMMU.

  1. 1 Cache Coherence Demystified: The MESI Protocol on Intel Granite Rapids (deep dive)
  2. 2 Prefetching Demystified: Hardware and Software Prefetchers on Intel Granite Rapids (deep dive)
  3. 3 IOMMU: The Gatekeeper Between Devices and Memory (deep dive)

Performance Profiling

Measuring where time and memory go, from Linux perf and /proc counters to GPU memory attribution.

  1. 1 Understanding Linux perf: stat and record
  2. 2 Adding a counter to the *proc* interface
  3. 3 GPU Memory Profiling Tools (NVIDIA and Intel)

Binaries, Libraries & Linking

How C programs call into libraries at the binary level, how to swap and instrument them, and how that surface is attacked.

  1. 1 Library Calling in C
  2. 2 Installing custom LibC (in progress)
  3. 3 Dyninst (in progress)
  4. 4 Attacks on the control flow integrity of applications

Intel SGX

Installing Intel SGX, understanding its internals, and running unmodified apps inside enclaves with Graphene.

  1. 1 Installing Intel SGX on Ubuntu 18.04
  2. 2 Intel SGX Internals
  3. 3 Install Graphene SGX

Filesystem Internals

How on-disk filesystems lay out data, from FAT to inode-based designs.

  1. 1 Fat File System
  2. 2 Inode File system (in progress)