Senior Principal Architect · Supermicro

Gowhar Reshi

Building AI Factories · Designing GPU Clusters · Scaling AI Infrastructure

I architect the systems that train and serve large-scale AI workloads — from RoCEv2 fabrics and lossless network design to distributed storage, GPU cluster operations, and reference architectures for next-generation AI factories.

AI FactoriesGPU ClustersRoCEv2 / RDMAAI NetworkingStorage ArchitectureDC NetworkingSONiCPFC / ECN / DCQCN
View ProjectsGet in Touch
scroll
Domains

Focus Areas

Where I spend my time — designing, operating, and writing about infrastructure at the intersection of AI compute, high-performance networking, and scalable storage.

⚡

AI Infrastructure & GPU Clusters

End-to-end architecture of AI factories: server selection, GPU interconnect topology, cluster operations, and scale-out design for training and inference workloads.

H100 / H200NVLinkInfiniBandCluster Ops
🔗

AI Networking & RoCEv2

Lossless Ethernet fabric design for RDMA workloads. PFC configuration, ECN tuning, DCQCN behavior, and operational readiness for large-scale AI fabrics.

RoCEv2PFCECNDCQCNSONiC
💾

AI Storage Architecture

High-throughput storage for AI pipelines: parallel file systems, NVMe over Fabrics, checkpoint storage design, and data pipeline architecture.

NVMe-oFGPFS / LustreCheckpointing
🌐

Data Center Networking

Spine-leaf fabric design, BGP in the data center, SONiC deployments, multi-tenant segmentation, and operational realities of large-scale switching.

BGPSpine-LeafSONiC400G / 800G
🔐

Security & Cryptography

Infrastructure security architecture, zero trust design, and post-quantum cryptography — practical application in enterprise and cloud environments.

Zero TrustPost-QuantumHybrid Encryption
📐

Reference Architectures

Building validated, reproducible reference designs for AI infrastructure deployments — bridging product capabilities with real-world operational requirements.

Validated DesignsBOMOps Runbooks
Publications

Technical Writing

Long-form articles on AI infrastructure, networking, and cryptography. Architecture-first, operationally grounded.

01
AI Infrastructure
Designing Next-Generation AI Infrastructure: Balancing Compute, Network, Memory, and Scale
A systems-level look at the architectural decisions that define modern AI factory design — GPU selection, fabric topology, memory hierarchy, and the operational tradeoffs that determine cluster efficiency.
Apr 2026
02
Cybersecurity
The Future of Cryptography: Why Hybrid Encryption Will Define the Post-Quantum Era
NIST's post-quantum standards are finalized. What that means for infrastructure architects — and why hybrid encryption is the only viable migration path for systems that need to outlast quantum threats.
Mar 2026
03
Cybersecurity
The Evolution of Cryptography: From War Rooms to Algorithms
Tracing the arc from Caesar ciphers and Enigma machines to AES, RSA, and the looming quantum threat — and what history tells us about the next transition.
Feb 2026
Deep Dives

Architecture Notes

Focused technical notes on specific design decisions, protocol behaviors, and operational patterns in AI infrastructure.

RoCEv2 / RDMA
PFC Storm Prevention in Large-Scale RoCEv2 Fabrics
Why PFC pause frames cascade, how to design watchdog timers and deadlock detection, and the fabric topology choices that reduce blast radius.
SONiC / DC Networking
BGP EVPN in AI Factory Networks: Operational Reality vs. Marketing Claims
What EVPN actually delivers in a GPU cluster environment, where it introduces operational complexity, and when simpler L3 fabrics are the right call.
AI Storage
Checkpoint Storage Design for 10,000-GPU Training Clusters
Failure probability math, checkpoint frequency tradeoffs, tiered storage architecture, and why most teams under-spec their checkpoint storage by 10×.
GPU Clusters
Rail-Optimized vs. Full-Bisection Bandwidth: When Each Topology Wins
The communication pattern analysis behind choosing rail-optimized fabrics for LLM training, and the workloads where full-bisection still makes sense.

Let's Talk Infrastructure

Open to technical conversations, architecture reviews, speaking invitations, or collaboration on research and reference architecture work.

Get in Touch