Advanced

AMD Instinct and ROCm

AMD's Instinct line competes on memory: 192 GB on the MI300X, 256 on the MI325X, 288 on the MI355X, with dense fp8 peaks at or above NVIDIA's at each generation. The catch is software: ROCm and HIP run most PyTorch and the main serving engines, but the attention, MoE and quantization kernels arrive on CUDA first and measured throughput has trailed the spec sheet. Interviewers want the numbers, the wavefront-of-64 difference, and a clear-eyed statement of when the memory argument wins.

Unlock the full curriculum — ₹2,000 / $25every concept + every answer · 6 months · no auto-renew

RELATED CONCEPTS

GPU Generations: A100 to Blackwell GPU Execution Model Numerics: FP32, BF16, FP8 and FP4 Serving Engines: vLLM, SGLang and TensorRT-LLM Tensor Cores and Matrix Units TCO: Buy vs Rent

PRACTICE THIS IN REAL QUESTIONS

GPU & Accelerator ArchitectureThe MI300X has 192 GB per card. When does AMD actually win against an H100 for LLM serving, and what would you check before betting on it?→CUDA, Triton & Kernel EngineeringPort a hand-written CUDA kernel to MI300X. What translates mechanically, and what silently computes the wrong answer?→Napkin Math, Cost & CapacityRank the H100, MI300X and Trainium2 by cost per token for decode→Open-Weights Models & Serving EnginesYou must serve a frontier open-weights model on non-NVIDIA accelerators. Plan it.→Hardware, Cabling & Cluster Build-OutA vendor claims their accelerator beats an H100 at half the price. How do you evaluate that?→GPU & Accelerator ArchitectureWalk me through the CUDA execution model: what are grids, blocks and warps, and what does the hardware actually schedule?→