Sardor Razikov commited on
Commit ·
53f2b06
1
Parent(s): 5e74789
docs: conservative VRAM-accounting wording + add amd-hackathon-2026 tag
Browse files
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
|
| 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 |
|