T5000 - Sample to use full performance

Robotics & Edge Computing Jetson Systems Jetson Thor
,
1

Dear Nvidia Team

We are testing the T5000 currently with “gpu_burn” and the tool “stress”. With these tools however, we are not able to get a higher power draw than ~120W. Do you have samples to put the full load that is possible on the module to reach the maximum power? Is this maximum power 130 W or higher?

According to the datasheet, the total power can reach 20Vx22A + 5Vx6A = 470W. How can these high currents be provoked?

Thank you for your help.

Kind regards

2

sudo nvpmodel --help

cat /etc/nvpmodel.conf

3

We know about nvpmodel, but that does not answer how we can stress the system to a maximum.

4

Set maximums.

sudo nvpmodel -m 0
sudo jetson_clocks

Monitoring

# Run in another terminal to watch temperature
watch -n 1 'cat /sys/devices/system/cpu/online; echo; grep . /sys/class/thermal/thermal_zone*/temp'

# Or install jtop
sudo -v
curl -LsSf https://raw.githubusercontent.com/rbonghi/jetson_stats/master/scripts/install_jtop_torun_without_sudo.sh | bash
# Or upgrade jtop
sudo -v
curl -LsSf https://raw.githubusercontent.com/rbonghi/jetson_stats/master/scripts/upgrade-jtop.sh | bash

CPU / memory stress

sudo apt install stress-ng 

#Control-C kills stress-ng

cat <<'EOF' > stress_ng.sh
#!/usr/bin/env bash
set -euo pipefail

cleanup() {
    echo
    echo "Stopping stress-ng..."
    sudo pkill -TERM stress-ng || true
    sleep 2
    sudo pkill -KILL stress-ng || true
}

trap cleanup INT TERM EXIT

stress-ng \
  --cpu 14 \
  --cpu-method matrixprod \
  --vm 4 \
  --vm-bytes 80% \
  --matrix 4 \
  --matrix-size 4096 \
  --timeout 0 &

STRESS_PID=$!

wait "$STRESS_PID"
EOF

chmod +x stress_ng.sh

GPU stress: CUDA matrix multiply

cat > thor_cuda_stress.cu <<'EOF'
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <stdio.h>
#include <stdlib.h>

#define CHECK_CUDA(x) do { \
    cudaError_t err = (x); \
    if (err != cudaSuccess) { \
        fprintf(stderr, "CUDA error %s:%d: %s\n", __FILE__, __LINE__, cudaGetErrorString(err)); \
        exit(1); \
    } \
} while (0)

#define CHECK_CUBLAS(x) do { \
    cublasStatus_t err = (x); \
    if (err != CUBLAS_STATUS_SUCCESS) { \
        fprintf(stderr, "cuBLAS error %s:%d: %d\n", __FILE__, __LINE__, err); \
        exit(1); \
    } \
} while (0)

int main(int argc, char **argv) {
    int n = 8192;
    if (argc > 1) n = atoi(argv[1]);

    size_t bytes = (size_t)n * n * sizeof(float);
    printf("Allocating three %.2f GiB matrices, n=%d\n",
           (3.0 * bytes) / (1024.0 * 1024.0 * 1024.0), n);

    float *A, *B, *C;
    CHECK_CUDA(cudaMalloc(&A, bytes));
    CHECK_CUDA(cudaMalloc(&B, bytes));
    CHECK_CUDA(cudaMalloc(&C, bytes));

    CHECK_CUDA(cudaMemset(A, 1, bytes));
    CHECK_CUDA(cudaMemset(B, 2, bytes));
    CHECK_CUDA(cudaMemset(C, 0, bytes));

    cublasHandle_t h;
    CHECK_CUBLAS(cublasCreate(&h));

    const float alpha = 1.0f;
    const float beta = 0.0f;

    unsigned long long iter = 0;
    while (1) {
        CHECK_CUBLAS(cublasSgemm(
            h,
            CUBLAS_OP_N,
            CUBLAS_OP_N,
            n, n, n,
            &alpha,
            A, n,
            B, n,
            &beta,
            C, n));

        CHECK_CUDA(cudaDeviceSynchronize());

        if ((++iter % 10) == 0) {
            printf("completed %llu GEMMs\n", iter);
            fflush(stdout);
        }
    }

    return 0;
}
EOF

# Compile
nvcc -O3 thor_cuda_stress.cu -lcublas -o thor_cuda_stress

# Run
./thor_cuda_stress 8192
# Or
./thor_cuda_stress 12288