feat: GPU Fingerprinting — Channel 8 for PPA (RIP-0308)

node/gpu_attestation.py

@@ -0,0 +1,86 @@
+import hashlib
+import json
+import os
+import subprocess
+from typing import Dict, Tuple
+
+def validate_gpu_fingerprint(fingerprint_data: Dict) -> Tuple[bool, str]:
+    """
+    Server-side validation of GPU fingerprint data (RIP-0308).
+
+    Checks for:
+    1. Silicon identity consistency (Compute Capability vs VRAM vs Arch)
+    2. Physical property variance (Jitter, Thermal Ramp, Latency Spread)
+    3. Cross-channel correlation (Fabric ratio for iGPUs)
+    """
+    try:
+        # 1. Identity Consistency Check
+        gpu_name = fingerprint_data.get("gpu_name", "").lower()
+        vram = fingerprint_data.get("vram_mb", 0)
+        cap = fingerprint_data.get("compute_capability", "0.0")
+
+        # known H100 SXM5 signature
+        if "h100" in gpu_name and "sxm" in gpu_name:
+            if cap != "9.0" or vram < 80000:
+                return False, "identity_mismatch: H100 SXM requires sm_9.0 and 80GB VRAM"
+
+        # 2. Channel 8c: Warp Jitter (Scheduling Variance)
+        # Real hardware has jitter (CV > 0.005). Perfect 0 is a sign of emulation.
+        ch8c = next((ch for ch in fingerprint_data.get("channels", []) if "8c" in ch["name"]), None)
+        if ch8c:
+            cv = ch8c.get("data", {}).get("cv", 0)
+            if cv < 0.005:
+                return False, "synthetic_timing: Warp jitter too low (likely emulated)"
+            if cv > 0.8:
+                return False, "high_latency_noise: Warp jitter too high (likely high-overhead VM)"
+
+        # 3. Channel 8b: Compute Asymmetry (FP16:FP32 Ratio)
+        # Physical ALU layout determines this ratio. It's generation-specific.
+        ch8b = next((ch for ch in fingerprint_data.get("channels", []) if "8b" in ch["name"]), None)
+        if ch8b:
+            ratios = ch8b.get("data", {}).get("asymmetry_ratios", {})
+            fp16_ratio = ratios.get("fp16_to_fp32", 0)
+            # Ampere/Ada Tensor Core ratio is typically > 2.0
+            if "rtx" in gpu_name and fp16_ratio < 1.1:
+                 return False, "alu_mismatch: Tensor core throughput not detected"
+
+        # 4. Channel 8i: iGPU Coherence (if applicable)
+        ch8i = next((ch for ch in fingerprint_data.get("channels", []) if "8i" in ch["name"]), None)
+        if ch8i:
+            fabric_ratio = ch8i.get("data", {}).get("fabric_ratio", 0)
+            # iGPUs have low fabric ratio (shared bus). dGPUs over PCIe would be much higher.
+            if fabric_ratio > 50:
+                return False, "fabric_latency_too_high: Likely discrete GPU masquerading as iGPU"
+
+        return True, "valid"
+    except Exception as e:
+        return False, f"validation_error: {str(e)}"
+
+def get_gpu_attestation_payload(device_index=0) -> Dict:
+    """
+    Run appropriate GPU fingerprinting and return the structured payload.
+    """
+    import torch
+    from miners.gpu_fingerprint import run_gpu_fingerprint
+    from miners.tensor_core_fingerprint import run_tensor_core_fingerprint
+
+    # Run main multi-channel fingerprint
+    fp = run_gpu_fingerprint(device_index=device_index)
+    payload = fp.to_dict()
+
+    # Add Channel 8f (Tensor Core LSB Signature) for deep silicon identity
+    try:
+        if fp.compute_capability >= "7.0":
+            tc_fp = run_tensor_core_fingerprint(device_index=device_index)
+            payload["channels"].append({
+                "name": "8f: Tensor Core Precision Drift",
+                "passed": tc_fp.all_passed,
+                "data": {
+                    "precision_hash": tc_fp.precision_hash,
+                    "lsb_signature": tc_fp.precision_hash # Use LSB composite
+                }
+            })
+    except Exception:
+        pass
+
+    return payload