RTX 5060 Ti 16GB Sanity Test Script

After provisioning a new RTX 5060 Ti 16GB server on our dedicated hosting, run a sanity test before putting a workload on it. Fifteen minutes of validation catches hardware or driver issues before they surprise you mid-deployment.

Contents

• Hardware visible
• CUDA and PyTorch
• Stress test
• Baseline inference
• VRAM ceiling

Hardware

nvidia-smi
# Expect: RTX 5060 Ti, 16 GB, driver 565+
nvidia-smi --query-gpu=name,memory.total,memory.used,temperature.gpu,power.draw --format=csv
# Baseline: ~15 MB used, ~35C, ~15 W idle

If nvidia-smi returns “No devices found”, driver is not loaded – reboot or reinstall driver.

CUDA and PyTorch

python3 -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"
# True RTX 5060 Ti

python3 -c "import torch; a = torch.randn(1024, 1024).cuda(); b = torch.randn(1024, 1024).cuda(); print((a @ b).sum().item())"
# Any finite number; tests matmul on GPU

Stress Test

Load the card briefly to check thermals:

python3 -c "
import torch, time
a = torch.randn(8192, 8192).cuda().half()
b = torch.randn(8192, 8192).cuda().half()
start = time.time()
for _ in range(200):
    c = a @ b
torch.cuda.synchronize()
print(f'200 iters: {time.time()-start:.1f}s')
"

Expected: 200 iterations in ~8-12 seconds. Monitor nvidia-smi during the run – temperature should stay under 80°C, power near 170 W.

Baseline Inference

Quick vLLM smoke test:

pip install vllm
python -m vllm.entrypoints.openai.api_server \
  --model neuralmagic/Llama-3.1-8B-Instruct-FP8 \
  --quantization fp8 --max-model-len 4096 &

sleep 120  # wait for model load

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"neuralmagic/Llama-3.1-8B-Instruct-FP8","messages":[{"role":"user","content":"Hello, say GigaGPU."}],"max_tokens":50}'

Expected: response with generated text in under 3 seconds.

VRAM Ceiling

python3 -c "
import torch
total = torch.cuda.get_device_properties(0).total_memory / 1e9
reserved = torch.cuda.memory_reserved(0) / 1e9
print(f'Total VRAM: {total:.1f} GB, Reserved: {reserved:.1f} GB')
"
# Expect Total ~16.0 GB

If total VRAM reported is under 15.5 GB, check that you did not get a different card variant. Contact support.

Once sanity passes, run the full benchmark script to establish your baseline throughput numbers.