Sardor Razikov commited on
Commit
53f2b06
·
1 Parent(s): 5e74789

docs: conservative VRAM-accounting wording + add amd-hackathon-2026 tag

Browse files
Files changed (1) hide show
  1. README.md +2 -1
README.md CHANGED
@@ -11,6 +11,7 @@ pinned: false
11
  license: mit
12
  short_description: Repo-scale coding agent — 256K context on a single MI300X
13
  tags:
 
14
  - amd-developer-hackathon
15
  - agents
16
  - coding-agent
@@ -32,7 +33,7 @@ tags:
32
  - Qwen3-Coder-Next-FP8 weights ≈ 80 GB
33
  - 256K KV cache @ FP8 ≈ 38 GB
34
  - activations ≈ 25 GB → **~143 GB total on a single GPU**
35
- - NVIDIA H100 80GB physically OOMs. AMD MI300X 192GB just runs it.
36
 
37
  This is a memory-architecture story, not a CUDA-vs-ROCm one.
38
 
 
11
  license: mit
12
  short_description: Repo-scale coding agent — 256K context on a single MI300X
13
  tags:
14
+ - amd-hackathon-2026
15
  - amd-developer-hackathon
16
  - agents
17
  - coding-agent
 
33
  - Qwen3-Coder-Next-FP8 weights ≈ 80 GB
34
  - 256K KV cache @ FP8 ≈ 38 GB
35
  - activations ≈ 25 GB → **~143 GB total on a single GPU**
36
+ - NVIDIA H100 80GB cannot accommodate this configuration on a single card by VRAM accounting (~143 GB > 80 GB). AMD MI300X 192 GB has the headroom.
37
 
38
  This is a memory-architecture story, not a CUDA-vs-ROCm one.
39