diff --git a/Makefile b/Makefile
index 40721ed..cba1a37 100644
--- a/Makefile
+++ b/Makefile
@@ -14,7 +14,7 @@ CUDA_OUTPUT_FILE = -o $@
 # NVCC flags
 # -t=0 is short for --threads, 0 = number of CPUs on the machine
 NVCC_FLAGS = -O3 -t=0 --use_fast_math -std=c++17
-NVCC_LDFLAGS = -lcublas -lcublasLt
+NVCC_LDFLAGS = -lcublas -lcublasLt -lnvidia-ml
 NVCC_INCLUDES =
 NVCC_LDLIBS =
 NCLL_INCUDES =
diff --git a/llmc/mfu.h b/llmc/mfu.h
index 7753305..5f8934f 100644
--- a/llmc/mfu.h
+++ b/llmc/mfu.h
@@ -4,12 +4,22 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <nvml.h>
 
 // tied to enum PrecisionMode, in a future refactor make them the same
 #define MFUH_PRECISION_FP32 0
 #define MFUH_PRECISION_FP16 1
 #define MFUH_PRECISION_BF16 2
 
+inline void nvml_check(nvmlReturn_t status, const char *file, int line) {
+    if (status != NVML_SUCCESS) {
+        printf("[NVML ERROR] at file %s:%d:\n%s\n", file, line, nvmlErrorString(status));
+        exit(EXIT_FAILURE);
+    }
+};
+#define nvmlCheck(err) (nvml_check(err, __FILE__, __LINE__))
+
+
 typedef struct {
     float TF_32;       // tensor-core performance 32 bit
     float BF_16_32;    // bf16 with 32 bit accumulate
@@ -134,4 +144,63 @@ float get_flops_promised(const char* device, int precision_mode) {
     return -1.0f; // ¯\_(ツ)_/¯
 }
 
+nvmlDevice_t nvml_get_device() {
+    static bool needs_init = true;
+    static nvmlDevice_t device;
+    if(needs_init) {
+        needs_init = false;
+        nvmlCheck(nvmlInit());
+        nvmlCheck(nvmlDeviceGetHandleByIndex_v2(0, &device));
+    }
+    return device;
+}
+
+struct GPUUtilInfo {
+    unsigned int clock;
+    unsigned int max_clock;
+    unsigned int power;
+    unsigned int power_limit;
+    unsigned int fan;
+    unsigned long long throttle;
+
+    float gpu_utilization;
+    float mem_utilization;
+};
+
+GPUUtilInfo get_gpu_utilization_info() {
+    GPUUtilInfo info;
+    nvmlDevice_t device = nvml_get_device();
+    nvmlCheck(nvmlDeviceGetClockInfo(device, NVML_CLOCK_SM, &info.clock));
+    nvmlCheck(nvmlDeviceGetMaxClockInfo(device, NVML_CLOCK_SM, &info.max_clock));
+    nvmlCheck(nvmlDeviceGetPowerManagementLimit(device, &info.power_limit));
+    nvmlCheck(nvmlDeviceGetPowerUsage(device, &info.power));
+    nvmlCheck(nvmlDeviceGetCurrentClocksThrottleReasons(device, &info.throttle));
+    nvmlCheck(nvmlDeviceGetFanSpeed(device, &info.fan));
+    // other potentially interesting functions
+    // nvmlDeviceGetPcieThroughput
+    // nvmlDeviceGetPcieSpeed
+    // nvmlDeviceGetNumGpuCores
+    // nvmlDeviceGetMemoryBusWidth
+    nvmlSample_t buffer[64];
+    nvmlValueType_t v_type;
+    unsigned int sample_count = 64;
+    nvmlCheck(nvmlDeviceGetSamples(device, NVML_GPU_UTILIZATION_SAMPLES, 0, &v_type, &sample_count, buffer));
+    float gpu_utilization = 0.f;
+    for(unsigned i = 0; i < sample_count; ++i) {
+        gpu_utilization += (float)buffer[i].sampleValue.uiVal;
+    }
+    gpu_utilization /= (float)sample_count;
+    sample_count = 64;
+    nvmlCheck(nvmlDeviceGetSamples(device, NVML_MEMORY_UTILIZATION_SAMPLES, 0, &v_type, &sample_count, buffer));
+    float mem_utilization = 0.f;
+    for(unsigned i = 0; i < sample_count; ++i) {
+        mem_utilization += (float)buffer[i].sampleValue.uiVal;
+    }
+    mem_utilization /= (float)sample_count;
+
+    info.gpu_utilization = gpu_utilization;
+    info.mem_utilization = mem_utilization;
+    return info;
+}
+
 #endif // MFU_H
diff --git a/train_gpt2.cu b/train_gpt2.cu
index 8f11091..10cebcb 100644
--- a/train_gpt2.cu
+++ b/train_gpt2.cu
@@ -1829,6 +1829,12 @@ int main(int argc, char *argv[]) {
         printf0("step %4d/%d | loss %7.6f (%+.2fz)| norm %6.4f (%+.2fz)| lr %.2e | %.2f ms | %.1f%% bf16 MFU | %.0f tok/s\n",
                 step + 1, train_num_batches, model.mean_loss, zloss, grad_norm, zgrad, step_learning_rate,
                 time_elapsed_ms, 100*mfu, bias_corrected_ema_tokens_per_second);
+        if((step + 1) % 100 == 0) {
+            GPUUtilInfo gpu_info = get_gpu_utilization_info();
+            printf0("GPU Status: Compute: %2.1f%% | Memory: %2.1f%% | Fan: %2d%%\n", gpu_info.gpu_utilization, gpu_info.mem_utilization, gpu_info.fan);
+            printf0("            Clock: %4d / %4d\n", gpu_info.clock, gpu_info.max_clock);
+            printf0("            Power: %4d / %4d\n", gpu_info.power, gpu_info.power_limit);
+        }
         logger_log_train(&logger, step, model.mean_loss, step_learning_rate, grad_norm);
 
         // disable the profiler after 3 steps of optimization