"Has anyone else hit the physical TGP ceiling on the Mobile 4060? I wrote a PyTorch cold-boot script that sustains 26.3+ TFLOPS across massive 163K x

Accelerated Computing CUDA CUDA Programming and Performance
1

(venv) PS C:\Users\watch\OneDrive\Desktop\veilsight\burst_expirements> python apex_hunter.py

🌌 VEILSIGHT APEX HUNTER: THE 28 TFLOPS BARRIER

Target: Absolute Silicon Saturation (Cold-Boot Burst)

[✓] Hardware: NVIDIA GeForce RTX 4060 Laptop GPU
[] Forging 81,920 x 8192 Tensor…
[] Pre-warming the CUDA cores…
🚀 IGNITING TRUE BURST (Fixed 50 Iterations)…

==================================================

🏆 APEX BURST COMPLETE

Iterations: 50

Time: 20.743 seconds
COMPUTE: 26.50 TFLOPS
BANDWIDTH: 6.79 GB/s

(venv) PS C:\Users\watch\OneDrive\Desktop\veilsight\burst_expirements> python apex_hunter.py

🌌 VEILSIGHT APEX HUNTER: THE 28 TFLOPS BARRIER

Target: Absolute Silicon Saturation (Cold-Boot Burst)

[✓] Hardware: NVIDIA GeForce RTX 4060 Laptop GPU
[] Forging 97,280 x 9728 Tensor…
[] Pre-warming the CUDA cores…
🚀 IGNITING TRUE BURST (Fixed 50 Iterations)…

==================================================

🏆 APEX BURST COMPLETE

Iterations: 50

Time: 35.001 seconds
COMPUTE: 26.30 TFLOPS
BANDWIDTH: 5.68 GB/s

(venv) PS C:\Users\watch\OneDrive\Desktop\veilsight\burst_expirements> python apex_hunter.py

🌌 VEILSIGHT APEX HUNTER: THE 28 TFLOPS BARRIER

Target: Absolute Silicon Saturation (Cold-Boot Burst)

[✓] Hardware: NVIDIA GeForce RTX 4060 Laptop GPU
[] Forging 102,400 x 9728 Tensor…
[] Pre-warming the CUDA cores…
🚀 IGNITING TRUE BURST (Fixed 50 Iterations)…

==================================================

🏆 APEX BURST COMPLETE

Iterations: 50

Time: 36.843 seconds
COMPUTE: 26.30 TFLOPS
BANDWIDTH: 5.66 GB/s

(venv) PS C:\Users\watch\OneDrive\Desktop\veilsight\burst_expirements> python apex_hunter.py

🌌 VEILSIGHT APEX HUNTER: THE 28 TFLOPS BARRIER

Target: Absolute Silicon Saturation (Cold-Boot Burst)

[✓] Hardware: NVIDIA GeForce RTX 4060 Laptop GPU
[] Forging 163,840 x 8192 Tensor…
[] Pre-warming the CUDA cores…
🚀 IGNITING TRUE BURST (Fixed 50 Iterations)…

==================================================

🏆 APEX BURST COMPLETE

Iterations: 50

Time: 41.800 seconds
COMPUTE: 26.30 TFLOPS
BANDWIDTH: 6.58 GB/s

(venv) PS C:\Users\watch\OneDrive\Desktop\veilsight\burst_expirements> python apex_hunter.py

🌌 VEILSIGHT APEX HUNTER: THE 28 TFLOPS BARRIER

Target: Absolute Silicon Saturation (Cold-Boot Burst)

[✓] Hardware: NVIDIA GeForce RTX 4060 Laptop GPU
[] Forging 163,840 x 8192 Tensor…
[] Pre-warming the CUDA cores…
🚀 IGNITING TRUE BURST (Fixed 50 Iterations)…

==================================================

🏆 APEX BURST COMPLETE

Iterations: 100

Time: 83.689 seconds
COMPUTE: 26.28 TFLOPS
BANDWIDTH: 6.58 GB/s

2

APEX_HUNTER.py

import torch
import time
import gc

print(“==================================================”)
print(“🌌 VEILSIGHT APEX HUNTER: THE 28 TFLOPS BARRIER”)
print(“Target: Absolute Silicon Saturation (Cold-Boot Burst)”)
print(“==================================================”)

device = torch.device(“cuda” if torch.cuda.is_available() else “cpu”)
if device.type != ‘cuda’:
print(“❌ CRITICAL: CUDA not detected.”)
exit()

1. UNLOCKING THE NVIDIA DRIVER

This tells the GPU to physically benchmark algorithmic pathways before executing

torch.backends.cudnn.benchmark = True

Allow Tensor Cores to use reduced precision internally for massive speedups

torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = True

2. THE GOLDEN RATIO (Base-2 Aligned for L2 Cache)

dim = 8192
rows = 163840

print(f"[✓] Hardware: {torch.cuda.get_device_name(0)}“)
print(f”[*] Forging {rows:,} x {dim} Tensor…")

try:

3. EXTREME INFERENCE MODE

Deeper than no_grad(), this disables all PyTorch version tracking

with torch.inference_mode():
A = torch.randn(rows, dim, dtype=torch.float16, device=device)
B = torch.randn(dim, dim, dtype=torch.float16, device=device)

    print("[*] Pre-warming the CUDA cores...")
    for _ in range(10):
        _ = torch.matmul(A, B)
    torch.cuda.synchronize()
    
    print("🚀 IGNITING TRUE BURST (Fixed 50 Iterations)...")
    iters = 100
    
    # Native CUDA Hardware Timers
    start_event = torch.cuda.Event(enable_timing=True)
    end_event = torch.cuda.Event(enable_timing=True)
    
    start_event.record()
    for _ in range(iters):
        _ = torch.matmul(A, B)
    end_event.record()
    
    # Wait for the specific batch to finish
    torch.cuda.synchronize()
    
    # elapsed_time returns milliseconds, convert to seconds
    burn_time = start_event.elapsed_time(end_event) / 1000.0 
    
    # The Math
    ops_per_iter = 2 * rows * dim * dim 
    total_ops = ops_per_iter * iters
    tflops = (total_ops / burn_time) / 1e12
    
    bytes_moved = ( (rows * dim) + (dim * dim) + (rows * dim) ) * 2
    bandwidth = ((bytes_moved * iters) / burn_time) / 1e9
    
    print("\n==================================================")
    print("🏆 APEX BURST COMPLETE")
    print("==================================================")
    print(f"Iterations:  {iters}")
    print(f"Time:        {burn_time:.3f} seconds")
    print(f"COMPUTE:     {tflops:.2f} TFLOPS")
    print(f"BANDWIDTH:   {bandwidth:.2f} GB/s")
    print("==================================================")

except torch.cuda.OutOfMemoryError:
print(“\n❌ VRAM WALL HIT. The cache overflowed.”)
except Exception as e:
print(f"\n❌ ERROR: {e}")
finally:
if ‘A’ in locals(): del A
if ‘B’ in locals(): del B
torch.cuda.empty_cache()
gc.collect()