Debarshi Das

Field of view10,000 km

Between a prompt and an answer, there is an entire world.

Debarshi Das is an AI infrastructure engineer in Bengaluru, India. He builds LLM inference, model serving, and GPU infrastructure for production AI systems.

Drag the planet. Scroll to descend.Choose a stop at the right to descend.

  1. OrbitA planet full of questions. This one leaves Bengaluru for Northern Virginia.
  2. RegionNorthern Virginia, home of us-east-1. The request descends toward Ashburn.
  3. HallRows of accelerators, each one waiting for work.
  4. NodeEight GPUs on one board. The fabric between them carries the prompt.
  5. PackageA die, surrounded by its memory.
  6. DieHere the prompt becomes arithmetic. Move the light.
  7. TransistorsThe arithmetic becomes current.
  8. FieldAnd somewhere in the light, an answer forms.

I work there.

Debarshi DasAI Infrastructure Engineer. Bengaluru, India.

The regions marked on the globe

Origin: Bengaluru, India, where this prompt was typed. Each AWS region sits at the approximate location of its metro area.

  • us-east-1: US East (N. Virginia), Ashburn, Virginia
  • us-east-2: US East (Ohio), Columbus, Ohio
  • us-west-1: US West (N. California), San Jose, California
  • us-west-2: US West (Oregon), Boardman, Oregon
  • ca-central-1: Canada (Central), Montreal, Quebec
  • ca-west-1: Canada West (Calgary), Calgary, Alberta
  • mx-central-1: Mexico (Central), Queretaro, Mexico
  • sa-east-1: South America (Sao Paulo), Sao Paulo, Brazil
  • eu-west-1: Europe (Ireland), Dublin, Ireland
  • eu-west-2: Europe (London), London, United Kingdom
  • eu-west-3: Europe (Paris), Paris, France
  • eu-central-1: Europe (Frankfurt), Frankfurt, Germany
  • eu-central-2: Europe (Zurich), Zurich, Switzerland
  • eu-north-1: Europe (Stockholm), Stockholm, Sweden
  • eu-south-1: Europe (Milan), Milan, Italy
  • eu-south-2: Europe (Spain), Aragon, Spain
  • il-central-1: Israel (Tel Aviv), Tel Aviv, Israel
  • me-south-1: Middle East (Bahrain), Bahrain
  • me-central-1: Middle East (UAE), United Arab Emirates
  • af-south-1: Africa (Cape Town), Cape Town, South Africa
  • ap-south-1: Asia Pacific (Mumbai), Mumbai, India
  • ap-south-2: Asia Pacific (Hyderabad), Hyderabad, India
  • ap-southeast-1: Asia Pacific (Singapore), Singapore
  • ap-southeast-2: Asia Pacific (Sydney), Sydney, Australia
  • ap-southeast-3: Asia Pacific (Jakarta), Jakarta, Indonesia
  • ap-southeast-4: Asia Pacific (Melbourne), Melbourne, Australia
  • ap-southeast-5: Asia Pacific (Malaysia), Kuala Lumpur, Malaysia
  • ap-southeast-7: Asia Pacific (Thailand), Bangkok, Thailand
  • ap-east-1: Asia Pacific (Hong Kong), Hong Kong
  • ap-east-2: Asia Pacific (Taipei), Taipei, Taiwan
  • ap-southeast-6: Asia Pacific (New Zealand), Auckland, New Zealand
  • ap-northeast-1: Asia Pacific (Tokyo), Tokyo, Japan
  • ap-northeast-2: Asia Pacific (Seoul), Seoul, South Korea
  • ap-northeast-3: Asia Pacific (Osaka), Osaka, Japan

Hello, I am Debarshi.

I am an AI infrastructure engineer in Bengaluru. I build the systems production AI runs on: LLM inference, model serving, and the GPUs beneath them.

I started on the screen, building interfaces at SOLRUF and PwC. Every screen is only as good as what stands behind it, so I kept digging: services, then servers, then GPUs and the networks between them. I still build the screen. Now I build what it stands on, too.

The path so far

  1. 2001

    Spawned on Earth

    2001

    The road starts here. The first eighteen years stay off the map.

  2. 2019

    B.Tech, Computer Science and Engineering

    Vellore Institute of Technology, Chennai, 2019 to 2023

    Computer science, and the side projects that taught me to ship, Algophi among them.

  3. 2021

    Research Intern

    Samsung India, December 2021 to September 2022

    Research on a token-based blockchain system for smart devices, with smart contracts for storing and reading data.

  4. 2022

    Frontend Developer Intern

    SOLRUF, August 2022 to September 2022

    My first frontend role: product search, faster pages through lazy loading, and responsive, accessible layouts.

  5. 2023

    Associate Consultant Intern

    PwC India, January 2023 to July 2023

    Rebuilt legacy Angular applications for a European car maker, so they loaded and changed faster.

  6. 2023

    Associate Technology Consultant

    PwC India, July 2023 to June 2024

    Modernized agreement-analysis tools in React with token-based sign-in, and built a map dashboard for enterprise clients.

  7. 2024

    Software Engineer I

    Metrum AI, July 2024 to September 2025

    Frontend and full-stack work on AI infrastructure products, where the systems behind the screens first pulled me down the stack.

  8. 2024

    Shown at SC24

    Supercomputing 2024, Atlanta, November 2024

    Our multimodal healthcare and customer-service RAG work, shown at the supercomputing conference.

  9. 2025

    Shown at AMD AI Day

    AMD, 2025

  10. 2025

    Software Engineer II

    Metrum AI, October 2025 to present

    Distributed inference and GPU clusters, and the control plane around them: routing, telemetry, access and billing. I still lead frontend work.

Workbench

Things I build in my own time, mostly to understand how something works.

Open data

gpu-specs

Cited specifications for NVIDIA and AMD GPUs. Every number carries its source document and the vendor's own words, so you can check it in one click.

121 GPUs, 166 sources, 1721 quoted values. Release v0.1.0.

curl -LO https://raw.githubusercontent.com/DarkStark9000/gpu-specs/v0.1.0/data/gpus.json
See it in use
NVIDIA H100 SXM, three of its values
Memory, as the vendor prints it
80 GB
GPU Memory | 80GB
NVIDIA H100 GPU (product page, Product Specifications table)
Memory bandwidth
3350 GB/s
GPU Memory Bandwidth | 3.35TB/s
NVIDIA H100 GPU (product page, Product Specifications table)
Memory, as the driver reports it
81 559 MiB
NVIDIA H100 80GB HBM3 ... 87MiB 81559MiB
NVIDIA Multi-Instance GPU User Guide (Supported GPUs page and PDF)
View Source of gpu-specs (opens in a new tab)

Latest

  1. Updatedgpu-specs
  2. UpdatedWhere the time in a Laya decision goes, measured on a laptop GPU
  3. Updatedsystem-one-bench
  4. PublishedWhere the time in a Laya decision goes, measured on a laptop GPU
  5. Updateddebarshidas.dev
  6. UpdatedDisaggregated prefill and decode on NVIDIA GPUs
  7. Startedgpu-specs
  8. UpdatedKV cache calculator
  9. PublishedSpeculative decoding: when it pays and when it does not
  10. Startedsystem-one-bench

Field notes

Notes I write while learning, on serving AI and building systems, with my sources.

  1. Speculative decoding: when it pays and when it does not

    When speculative decoding speeds up LLM inference: acceptance rate, draft length, batch size, long context, EAGLE-3 in vLLM and SGLang, and hybrid Mamba models.

    17 min read

  2. Effectively once: placing an order when every system answers twice

    How a mutual fund order backend applies each payment, exchange answer, and registrar row once, while every system around it retries and reorders.

    9 min read

  3. KV cache size per token: one formula does not fit every model

    The KV cache formula, computed from real configs for 12 model types, and what it means for capacity, prefix hits and handoff.

    21 min read

  4. Disaggregated prefill and decode on NVIDIA GPUs

    How disaggregated prefill and decode works on NVIDIA GPUs: what the KV-cache handoff costs over NVLink and RDMA, how Dynamo and NIXL move it, and when not to split.

    21 min read

  5. Consent is an analytics execution boundary

    A practical Next.js pattern that keeps optional analytics code and collection behind an explicit visitor decision.

    4 min read