---
title: "Discussion on \"Blockwise FP8 on Ada GPUs: why your L4 falls back to Marlin, measured\" | Hashnode"
description: "Discussion on \"Blockwise FP8 on Ada GPUs: why your L4 falls back to Marlin, measured\". What 128×128 block scaling is, why it needs its own kernel, and what the missing kernel costs on a real L4"
url: "https://hashnode.com/posts/blockwise-fp8-ada-gpus-marlin-fallback/6abfad8204a9ca58d66a76ac"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Blockwise FP8 on Ada GPUs: why your L4 falls back to Marlin, measured](https://amankarki.hashnode.dev/blockwise-fp8-ada-gpus-marlin-fallback?utm_source=hashnode&utm_medium=feed)

What 128×128 block scaling is, why it needs its own kernel, and what the missing kernel costs on a real L4

Oct 2 · 8 min read · amankarki.hashnode.dev  
Read the article

-   [#cuda](https://hashnode.com/tag/cuda)
-   [#gpu](https://hashnode.com/tag/gpu)
-   [#llm](https://hashnode.com/tag/llm)
-   [#vllm](https://hashnode.com/tag/vllm)
-   [#quantization](https://hashnode.com/tag/quantization)

## Written by

[Aman Karki](https://hashnode.com/@amankarki)

@amankarki · India

  
[View profile](https://hashnode.com/@amankarki)

C++/CUDA Engineer · LLM Inference & Systems

C++/CUDA engineer building LLM inference and GPU code from scratch: verbum.cpp, Lattice, RAAG. 2 CUDA PRs merged into llama.cpp. I write about inference, GPU performance, and measuring things honestly.

## Responses

Markdown supported · ⌘↵ to post

No responses yet.