The Vocabulary of GPUs for Gen AI Engineers

The conversation around GPUs in Gen AI talks often jumps straight to "just rent an H100" without explaining why.

I wrote a visual guide covering the vocabulary that actually matters:

🔹 Why GPUs over CPUs (it's not just "more cores")
🔹 HBM vs GDDR — why your RTX 4090 can't run Llama 405B
🔹 FLOPs, TFLOPS, and what those spec sheets actually mean
🔹 Precision formats: FP32 → FP16 → BF16 → FP8
🔹 The memory formula: Parameters × Bytes = VRAM needed
🔹 How inference actually works — from prompt to prediction
🔹 Temperature: the inference-time knob everyone uses but few explain

This isn't about which GPU to buy.

It's about building the mental model so you can read a spec sheet, estimate memory requirements, and have informed conversations about infrastructure.

Part 1 of a 3-part series - https://medium.com/@vinodh.thiagarajan/the-vocabulary-of-gpus-for-ml-budding-gen-ai-engineers-7a693b53b74b