Skip to content

Instantly share code, notes, and snippets.

@Lua12138

Lua12138/benchmark.py

Last active May 25, 2026 09:14
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save Lua12138/fc4a9807b3c2f1e7b27bf60479db877a to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save Lua12138/fc4a9807b3c2f1e7b27bf60479db877a to your computer and use it in GitHub Desktop.
Download ZIP
A script to detect GPU TFLOPS using torch
Raw
benchmark.py
import torch
import time
import platform
def get_system_info():
props = torch.cuda.get_device_properties(0)
return {
"OS": f"{platform.system()} {platform.release()}",
"CPU": platform.processor(),
"GPU": props.name,
"VRAM": f"{round(props.total_memory / (1024**3), 2)} GB",
"Arch": f"Compute Cap {props.major}.{props.minor}",
"Backend": f"Torch {torch.__version__} / {torch.version.cuda}"
}
def benchmark_compute_cell(target_vram_gb, dtype):
device = "cuda"
matrix_size = 4096
chunks, A, B = None, None, None
# 1. Physical VRAM Allocation (Fix: differentiate between float and int)
elem_size = torch.tensor([], dtype=dtype).element_size()
chunk_elements = (256 * 1024 * 1024) // elem_size
num_chunks = int((target_vram_gb * 1024**3) / (256 * 1024**2))
chunks = []
try:
for _ in range(num_chunks):
# FIX: Use randint for integer types, randn for floating points
if dtype.is_floating_point:
chunks.append(torch.randn(chunk_elements, device=device, dtype=dtype))
else:
# Fill with random integers in a safe range
chunks.append(torch.randint(-10, 10, (chunk_elements,), device=device, dtype=dtype))
except Exception:
pass
if not chunks: return 0.0
# 2. Compute Performance Measurement
try:
# Prepare operands (Fix: randint for integers)
if dtype.is_floating_point:
A = torch.randn((matrix_size, matrix_size), device=device, dtype=dtype)
B = torch.randn((matrix_size, matrix_size), device=device, dtype=dtype)
else:
A = torch.randint(-10, 10, (matrix_size, matrix_size), device=device, dtype=dtype)
B = torch.randint(-10, 10, (matrix_size, matrix_size), device=device, dtype=dtype)
res = None
torch.cuda.synchronize()
iters = 15
start = time.perf_counter()
for i in range(iters):
if dtype.is_floating_point:
if res is not None:
B[0, 0] += res.mean()
res = torch.matmul(A, B)
ops_per_iter = 2 * matrix_size**3
else:
if res is not None:
B[0, 0] += res[0, 0]
res = (A * B) + A
ops_per_iter = 2 * matrix_size**3 / 16 # Scaled for normalization
# VRAM Dependency Injection
A += chunks[i % len(chunks)][:matrix_size*matrix_size].view(matrix_size, matrix_size)
A += chunks[i % len(chunks)][:matrix_size*matrix_size].view(matrix_size, matrix_size)
torch.cuda.synchronize()
end = time.perf_counter()
duration_sec = (end - start) / iters
perf_score = (ops_per_iter / duration_sec) / 1e12
return perf_score
except Exception:
return 0.0
finally:
if chunks is not None: del chunks
if A is not None: del A
if B is not None: del B
torch.cuda.empty_cache()
# --- Execution ---
info = get_system_info()
print("="*80)
print("SYSTEM DIAGNOSTICS")
for k, v in info.items(): print(f"{k:<12}: {v}")
print("="*80)
gpu_capacity = torch.cuda.get_device_properties(0).total_memory
available_vram_gb = gpu_capacity // (1024 ** 3)
test_steps = list(range(1, available_vram_gb+1))
dtypes = {
"INT8": torch.int8,
"INT16": torch.int16,
"INT32": torch.int32,
"FP16": torch.float16,
"BF16": torch.bfloat16,
"FP32": torch.float32,
"FP64": torch.float64
}
# Run Benchmarks
results_matrix = {}
for name, dt in dtypes.items():
print(f"Benchmarking {name}...", end="\r")
col_results = {}
for gb in test_steps:
col_results[gb] = benchmark_compute_cell(gb, dt)
results_matrix[name] = col_results
print(f"All benchmarking done", end="\r")
# --- Final Table Output ---
print("\n" + "="*80)
print(f"COMPUTE PERFORMANCE REPORT (TFLOPS) - Higher is Better")
print("-" * 80)
header = f"{'VRAM GB':<10}"
for name in dtypes.keys(): header += f" | {name:<12}"
print(header)
print("-" * 80)
for gb in test_steps:
row = f"{gb:<10.1f}"
for name in dtypes.keys():
val = results_matrix[name].get(gb, 0.0)
row += f" | {val:<12.4f}"
print(row)
print("="*80)
@Lua12138

Lua12138 commented Apr 28, 2026

Copy link
Copy Markdown
Author 
================================================================================
SYSTEM DIAGNOSTICS
OS          : Linux 6.6.113+
CPU         : x86_64
GPU         : Tesla T4
VRAM        : 14.56 GB
Arch        : Compute Cap 7.5
Backend     : Torch 2.10.0+cu128 / 12.8
================================================================================

================================================================================
COMPUTE PERFORMANCE REPORT (TFLOPS) - Higher is Better
--------------------------------------------------------------------------------
VRAM GB    | INT8         | INT16        | INT32        | FP16         | BF16         | FP32         | FP64        
--------------------------------------------------------------------------------
1.0        | 2.7133       | 6.0406       | 2.9837       | 23.7107      | 1.6805       | 3.4189       | 0.2415      
2.0        | 12.6171      | 6.1250       | 3.0266       | 17.5787      | 1.6406       | 3.4680       | 0.2418      
3.0        | 14.0303      | 6.2226       | 3.0294       | 18.4580      | 1.6437       | 3.3881       | 0.2418      
4.0        | 11.9880      | 6.0711       | 3.0123       | 17.9556      | 1.6492       | 3.4096       | 0.2418      
5.0        | 12.0174      | 6.0620       | 3.0008       | 17.3214      | 1.6634       | 3.3895       | 0.2418      
6.0        | 12.3675      | 6.0301       | 2.9975       | 17.0475      | 1.6799       | 3.3856       | 0.2415      
7.0        | 12.2805      | 6.0768       | 3.0047       | 17.5876      | 1.6917       | 3.3920       | 0.2416      
8.0        | 12.2435      | 6.0752       | 3.0025       | 23.3704      | 1.7222       | 3.3840       | 0.2416      
9.0        | 10.9346      | 6.0688       | 2.9992       | 21.2667      | 1.7388       | 3.4786       | 0.2417      
10.0       | 11.2239      | 6.0820       | 2.9867       | 21.8001      | 1.7918       | 3.4283       | 0.2416      
11.0       | 11.4281      | 6.0404       | 2.9991       | 23.2962      | 1.7944       | 3.4922       | 0.2417      
12.0       | 11.3640      | 6.0961       | 3.0034       | 22.3581      | 1.8077       | 3.4947       | 0.2417      
13.0       | 11.3720      | 6.0709       | 3.1963       | 21.1299      | 1.8239       | 3.5146       | 0.2419      
14.0       | 13.9182      | 6.8030       | 3.3584       | 21.0882      | 1.8286       | 3.5004       | 0.0000      
================================================================================

@Lua12138

Lua12138 commented Apr 28, 2026

Copy link
Copy Markdown
Author 
================================================================================
SYSTEM DIAGNOSTICS
OS          : Windows 11
CPU         : AMD64 Family 25 Model 33 Stepping 2, AuthenticAMD
GPU         : AMD Radeon RX 7900 XTX
VRAM        : 23.98 GB
Arch        : Compute Cap 11.0
Backend     : Torch 2.10.0a0+rocm7.10.0a20251120 / None
================================================================================
All benchmarking done
================================================================================
COMPUTE PERFORMANCE REPORT (TFLOPS) - Higher is Better
--------------------------------------------------------------------------------
VRAM GB    | INT8         | INT16        | INT32        | FP16         | BF16         | FP32         | FP64
--------------------------------------------------------------------------------
1.0        | 1.3394       | 1.4338       | 2.5516       | 12.7965      | 8.4983       | 6.0509       | 1.0543
2.0        | 34.4811      | 4.5476       | 2.8527       | 48.1365      | 56.6924      | 19.1415      | 1.0249
3.0        | 45.0584      | 6.3482       | 2.4877       | 44.3068      | 58.1990      | 19.8741      | 1.0639
4.0        | 25.8562      | 6.4505       | 3.3395       | 52.3331      | 55.8807      | 20.3470      | 0.9779
5.0        | 27.3118      | 6.4253       | 3.2944       | 54.7542      | 59.4955      | 20.0175      | 1.0673
6.0        | 25.3211      | 6.4984       | 3.2779       | 55.7319      | 60.6414      | 19.7747      | 1.0582
7.0        | 26.0070      | 6.6061       | 3.1471       | 41.1660      | 57.5070      | 12.3532      | 1.0584
8.0        | 27.1570      | 6.3578       | 3.3463       | 56.3347      | 57.5408      | 20.4308      | 1.0619
9.0        | 22.6373      | 5.8721       | 3.2402       | 57.9432      | 57.6936      | 19.7200      | 1.0685
10.0       | 26.5400      | 6.4486       | 3.3798       | 57.1603      | 59.2644      | 20.4623      | 1.0661
11.0       | 25.4592      | 6.5189       | 3.3772       | 58.4769      | 59.9680      | 19.0941      | 1.0622
12.0       | 24.8455      | 6.3476       | 3.3154       | 57.3191      | 60.3427      | 20.4527      | 1.0644
13.0       | 26.1697      | 6.3580       | 3.3065       | 56.4202      | 57.0910      | 19.6925      | 1.0676
14.0       | 26.1580      | 6.3895       | 2.9189       | 57.9007      | 47.0302      | 20.4858      | 1.0636
15.0       | 19.9315      | 5.1354       | 3.4653       | 57.8984      | 61.4734      | 20.3091      | 1.0570
16.0       | 21.1530      | 6.8815       | 3.5538       | 57.1860      | 59.8424      | 19.7960      | 1.0664
17.0       | 23.4557      | 7.5250       | 3.4683       | 58.1642      | 61.3103      | 19.2114      | 1.0612
18.0       | 26.0285      | 7.2151       | 3.5146       | 46.8021      | 60.0445      | 20.8182      | 1.0697
19.0       | 24.6695      | 7.4401       | 3.4984       | 59.9399      | 60.7893      | 19.9111      | 1.0146
20.0       | 17.2417      | 7.3729       | 3.4683       | 49.9312      | 57.6158      | 20.8803      | 1.0548
21.0       | 12.8340      | 7.3103       | 3.5598       | 57.6464      | 50.1524      | 20.2402      | 1.0255
22.0       | 11.0178      | 7.2859       | 3.4966       | 57.1618      | 57.9797      | 19.9253      | 1.0221
23.0       | 11.5246      | 7.4358       | 3.5801       | 59.1279      | 61.5389      | 20.5735      | 1.0570
================================================================================

@Lua12138

Lua12138 commented Apr 28, 2026

Copy link
Copy Markdown
Author 
================================================================================
SYSTEM DIAGNOSTICS
OS          : Linux 6.8.0-106-generic
CPU         : x86_64
GPU         : NVIDIA A10
VRAM        : 22.06 GB
Arch        : Compute Cap 8.6
Backend     : Torch 2.11.0+cu130 / 13.0
================================================================================
All benchmarking done
================================================================================
COMPUTE PERFORMANCE REPORT (TFLOPS) - Higher is Better
--------------------------------------------------------------------------------
VRAM GB    | INT8         | INT16        | INT32        | FP16         | BF16         | FP32         | FP64
--------------------------------------------------------------------------------
1.0        | 5.4735       | 10.3598      | 5.8467       | 13.9324      | 15.5496      | 9.9438       | 0.4150
2.0        | 16.0881      | 10.3806      | 5.8485       | 63.5509      | 71.1824      | 12.0698      | 0.4171
3.0        | 17.4021      | 10.3745      | 5.8525       | 66.9665      | 69.7174      | 12.4950      | 0.4171
4.0        | 17.4651      | 8.8665       | 5.8442       | 61.3879      | 71.1452      | 12.6691      | 0.4171
5.0        | 17.9677      | 10.3492      | 5.8419       | 68.8794      | 70.6128      | 12.3076      | 0.4171
6.0        | 16.0508      | 10.3695      | 5.8528       | 69.8288      | 72.0779      | 12.3220      | 0.4155
7.0        | 18.2522      | 10.3748      | 5.8579       | 69.5522      | 71.3750      | 12.4823      | 0.4171
8.0        | 16.8158      | 10.3611      | 5.8554       | 68.6237      | 71.0050      | 12.4840      | 0.4170
9.0        | 17.4409      | 10.3813      | 5.8580       | 68.7129      | 70.2430      | 12.7346      | 0.4051
10.0       | 17.4509      | 10.3785      | 5.8604       | 70.0807      | 72.0470      | 12.5288      | 0.4171
11.0       | 16.0735      | 10.3667      | 5.6664       | 69.8102      | 71.8628      | 12.4567      | 0.4170
12.0       | 17.4099      | 10.3536      | 5.8611       | 68.6358      | 71.9821      | 12.3800      | 0.4170
13.0       | 17.4339      | 10.3730      | 5.8552       | 69.1466      | 70.2406      | 12.4161      | 0.4170
14.0       | 18.2652      | 10.3720      | 5.8499       | 68.9871      | 71.8969      | 12.3675      | 0.4158
15.0       | 17.4301      | 10.3815      | 5.8537       | 69.4430      | 71.7875      | 5.8510       | 0.4171
16.0       | 17.4106      | 10.3597      | 5.8468       | 68.8270      | 71.9325      | 12.4432      | 0.4171
17.0       | 17.5228      | 10.3727      | 5.6733       | 69.0329      | 72.2553      | 12.4842      | 0.4171
18.0       | 17.4037      | 10.3492      | 5.6680       | 69.5787      | 70.6584      | 12.5389      | 0.4055
19.0       | 25.7002      | 10.3642      | 5.8478       | 69.6702      | 72.1023      | 12.6141      | 0.4171
20.0       | 17.4559      | 10.3525      | 5.6612       | 71.8157      | 76.1790      | 12.6965      | 0.4171
21.0       | 18.1936      | 13.3459      | 6.2544       | 69.9451      | 72.6117      | 12.7249      | 0.4124
22.0       | 0.0000       | 0.0000       | 0.0000       | 0.0000       | 75.0666      | 0.0000       | 0.0000
================================================================================

@Lua12138

Lua12138 commented May 25, 2026

Copy link
Copy Markdown
Author 
================================================================================
SYSTEM DIAGNOSTICS
OS          : Linux 7.0.0-15-generic
CPU         : 
GPU         : NVIDIA GeForce RTX 5090 D v2
VRAM        : 23.41 GB
Arch        : Compute Cap 12.0
Backend     : Torch 2.11.0+cu130 / 13.0
================================================================================
All benchmarking done
================================================================================
COMPUTE PERFORMANCE REPORT (TFLOPS) - Higher is Better
--------------------------------------------------------------------------------
VRAM GB    | INT8         | INT16        | INT32        | FP16         | BF16         | FP32         | FP64        
--------------------------------------------------------------------------------
1.0        | 10.2301      | 44.7920      | 15.0907      | 21.4661      | 29.6092      | 27.6435      | 1.5918      
2.0        | 144.8313     | 49.1349      | 15.4851      | 186.2803     | 191.6524     | 58.7142      | 1.6383      
3.0        | 128.3673     | 44.6270      | 15.3233      | 185.9901     | 190.2927     | 59.5430      | 1.6383      
4.0        | 140.2818     | 50.2781      | 15.5195      | 189.5303     | 189.9085     | 59.5667      | 1.6340      
5.0        | 137.4829     | 45.4310      | 15.2576      | 184.2244     | 190.0584     | 59.5024      | 1.6395      
6.0        | 136.4229     | 49.3035      | 14.7178      | 189.7964     | 189.2351     | 59.3024      | 1.6439      
7.0        | 137.5033     | 47.8476      | 15.4067      | 188.7432     | 189.8978     | 59.4527      | 1.6436      
8.0        | 137.7749     | 49.0841      | 15.3194      | 184.3960     | 189.6070     | 59.3933      | 1.6441      
9.0        | 137.6978     | 48.2174      | 15.3767      | 189.4217     | 190.0001     | 59.3818      | 1.6391      
10.0       | 137.7415     | 49.7333      | 15.4279      | 189.6808     | 189.2064     | 59.5064      | 1.6390      
11.0       | 135.1158     | 47.8973      | 14.8607      | 188.8723     | 190.1686     | 59.4420      | 1.6440      
12.0       | 137.0954     | 49.8163      | 15.0624      | 189.8682     | 189.8156     | 59.3243      | 1.6391      
13.0       | 136.2664     | 47.4085      | 14.9469      | 189.2087     | 189.8753     | 59.4051      | 1.5912      
14.0       | 132.3044     | 50.0050      | 15.2279      | 188.6305     | 189.1356     | 59.4203      | 1.6296      
15.0       | 122.6620     | 48.0158      | 15.3992      | 188.0039     | 190.0929     | 59.5259      | 1.6441      
16.0       | 134.4475     | 49.5539      | 15.4654      | 189.2408     | 189.9341     | 59.5567      | 1.6441      
17.0       | 131.7314     | 47.8250      | 14.7391      | 189.4077     | 189.7696     | 59.4809      | 1.6437      
18.0       | 134.9713     | 49.8745      | 15.0967      | 188.5701     | 190.0609     | 59.3144      | 1.6388      
19.0       | 135.3311     | 47.4569      | 15.3348      | 188.6584     | 189.4480     | 59.4533      | 1.6392      
20.0       | 134.9556     | 49.0999      | 15.0326      | 189.2268     | 189.4351     | 58.4601      | 1.6442      
21.0       | 0.0000       | 0.0000       | 0.0000       | 189.0807     | 189.2822     | 0.0000       | 0.0000      
22.0       | 0.0000       | 0.0000       | 0.0000       | 189.0104     | 189.8917     | 0.0000       | 0.0000      
23.0       | 0.0000       | 0.0000       | 0.0000       | 189.3388     | 189.7761     | 0.0000       | 0.0000      
================================================================================

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment