🏆AMD MI300X Hackathon · lablab.ai

3-Agent AI Tumour Board
Powered by AMD MI300X

The only GPU with 192 GB unified memory — running GigaPath ViT-Giant + Llama 3.3 70B simultaneously, with multi-round agent debate grounded in NCCN 2024 guidelines. No model swapping. No OOM. Just results.

📖 The Story🧬 Specialists📊 Benchmark

▶ Run a Demo Case — one click, no setup

VRAM · AMD MI300X
live
MI300X — 192 GB HBM30.0 / 192 GB(192 GB free)
H100 SXM5 — 80 GB limit✗ OOM
needs multi-GPU sharding
40 GB
Llama 3.3 70B Q4_K_S
21% of 192 GB
20 GB
KV Cache (full debate)
10% of 192 GB
15 GB
Qwen-VL 7B (BF16)
8% of 192 GB
22 GB
LoRA ×3 (8B base)
11% of 192 GB
3 GB
GigaPath ViT-Giant
2% of 192 GB
9 GB
Qdrant + Overhead
5% of 192 GB
AMD MI300X · 192 GB HBM3
~109 GB used · 83 GB headroom · No model swapping
192 GB
HBM3 Unified Memory
3
AI Agents
≤3
Debate Rounds
~60s
Full Pipeline

Why Only AMD Makes This Possible

All five AI specialists live simultaneously in the MI300X's 192 GB unified memory. Toggle components to see when an H100 runs out of room.

🧮 Interactive VRAM Simulator

Toggle components to see which GPU can hold the full AOB stack. Required components are locked.

AMD MI300X · 192 GB HBM3109 / 192 GB
192 GB
NVIDIA H100 · 80 GB limit💥 OOM +29 GB
← 80 GB OOM threshold
H100 cannot run this configuration. Requires 109 GB — 29GB over the H100's 80 GB limit. The MI300X has 83 GB headroom remaining.

How the Board Works

🔬
Pathologist
GigaPath ViT-Giant 1.1B

Analyses histopathology patches with GigaPath ViT-Giant. Runs MC Dropout uncertainty quantification. Extracts 8 biomarker scores.

📚
Researcher
RAG + NCCN 2024

Retrieves evidence from NCCN guidelines. Challenges draft plan. Flags missing molecular tests and contraindicated regimens.

👨‍⚕️
Oncologist
Llama 3.3 70B

Synthesises pathology + evidence into a full management plan. Revises under critique. Finalises with consensus score ≥70/100.

🔬 Pathologist→📚 Researcher→👨‍⚕️ Oncologist (draft)→⚖️ Debate Loop (×3)→✅ Consensus ≥70→📊 Final Report
Upload Your Own Patches
Optional: submit 224x224 histology patches (PNG/JPG). Demo cases work best for judges.