ML Engineers
Training models that are hitting memory walls or speed limits. You need to go below the framework and understand what's actually happening on the hardware.
Cohort 1 starts July 1, 2026 — Limited seats
The GPU Engineering
Course for the AI Era
From GPU architecture to CUDA kernels to distributed training at scale. Six hands-on modules built for engineers who need to operate at the hardware layer.
6
Modules
30+
Lab exercises
100%
Hands-on
Limited
Seats available
Who it's for
Built for people who work close to the metal.
Platform & DevOps Engineers
Managing GPU clusters, Kubernetes nodes, or bare-metal servers for AI teams. You want to monitor, debug, and operate GPU workloads with confidence.
CS Students & Researchers
Learning GPU programming for research or a career in AI infrastructure. You want practical experience, not just theory — with labs that mirror real production environments.
Full Curriculum
Six modules. Zero fluff.
Built for practitioners.
01
GPU Systems Fundamentals
Understand every layer of the NVIDIA software stack — from silicon to container. The mental model that makes everything else click.
GPU Architecture basics
SMs, warps, thread blocks, memory hierarchy
CUDA ecosystem
How NVIDIA driver, CUDA toolkit, and cuDNN fit together
Docker + NVIDIA Container Toolkit
Containerized GPU environments from scratch
GPU Architecture
CUDA Ecosystem
NVIDIA Driver
CUDA Toolkit
cuDNN
Docker
NVIDIA Container Toolkit
02
PyTorch Performance
Move from "it runs" to "I know why it's slow." Profile memory, fix bottlenecks, and squeeze real throughput from your training loop.
Tensors & the training loop
GPU tensor lifecycle, grad computation, data loading
Mixed precision
FP16/BF16, autocast, gradient scaling
Profiling
PyTorch Profiler, Nsight, reading traces
Tensors
GPU Training Loop
Mixed Precision
Memory Usage
Profiling
03
CUDA for Engineers
Write your first CUDA kernels in Python. Understand memory transfer, shared memory, and the performance traps that trip up real engineers.
Kernel programming
Thread indexing, parallelism, occupancy
Memory transfer
Host↔device, pinned memory, PCIe bandwidth
Numba / CUDA Python
Write real kernels without leaving Python
CUDA Kernels
Memory Transfer
Numba
CUDA Python
Performance Pitfalls
04
Distributed Training
Train across multiple GPUs and nodes. Understand the communication layer — NCCL, NVLink, NVSwitch — and why your multi-GPU job is stalling.
DDP & FSDP
PyTorch distributed, gradient sync, bucket sizing
NCCL internals
AllReduce, ring topology, NCCL tuning
NVLink / NVSwitch
GPU-to-GPU interconnect, bandwidth vs PCIe
DDP
NCCL
NVLink
NVSwitch
Data Parallelism
Model Parallelism
05
GPU Operations
Operate shared GPU environments like a pro. Schedule jobs, monitor utilization, detect failures before they cascade, and read logs that actually tell you something.
Slurm scheduler
sbatch, squeue, GPU partitions, preemption
GPU monitoring
nvidia-smi, DCGM, utilization vs saturation
Failure detection
XID errors, ECC, health checks, alert patterns
Slurm
Job Scheduling
GPU Monitoring
Failure Detection
Health Checks
06
Optimize a Real Training Workload
Apply everything. Take a real training job, profile it end-to-end, identify and fix bottlenecks, and produce a GPU performance report that explains the tradeoffs.
Optimize throughput
Apply mixed precision, profiling, and kernel-level fixes to a real model
GPU performance report
Document findings, show evidence, make data-driven recommendations
Portfolio artifact
A completed report you can show in interviews or use internally
Course Format
Live cohorts. Real labs. GPU access included.
The first cohort will be live instructor-led sessions with lab time after each module. We are measuring demand to finalize schedule, depth, and pricing.
Students get access to GPU lab environments — no cloud account required for labs.
Reserve your spot
Weekend live cohort
Saturday sessions, recorded for async replay
~24 hours of content
Spread across 6 modules with lab time
GPU lab access
Hands-on environments provided, no setup required
Certificate of completion
With capstone performance report artifact
We are calibrating based on
- Schedule preference (weekend vs weeknight)
- Current experience level and role
- Interest in team / corporate seats
- Budget signals for first cohort pricing
Join the Waitlist
Shape the first NullHz cohort.
Starts July 1, 2026
Pleasanton, CA
Limited seats
Submit your interest below. Your responses directly inform how the first cohort is structured — schedule, depth, pricing, and lab environment.
Early access pricing before public launch
Priority enrollment — seats are limited
In-person in Pleasanton, CA with GPU labs