diff --git a/src/app/benchmark.c b/src/app/benchmark.c
index 65f8b5a..a50a8be 100644
--- a/src/app/benchmark.c
+++ b/src/app/benchmark.c
@@ -221,15 +221,24 @@ struct background_metrics {
 };
 
 #define MAX_TELEMETRY_SAMPLES 1800
+#define MAX_TELEMETRY_PHASES 44
 
 struct telemetry_sample {
 	uint64_t elapsed_ms;
 	double temperature_c, clock_mhz;
 };
 
+struct telemetry_phase {
+	uint64_t elapsed_ms;
+	const char *suite, *mode;
+	int workload_index;
+};
+
 struct telemetry_monitor {
 	struct telemetry_sample samples[MAX_TELEMETRY_SAMPLES];
 	size_t count;
+	struct telemetry_phase phases[MAX_TELEMETRY_PHASES];
+	size_t phase_count;
 	double started;
 	volatile int stop;
 #if defined(_WIN32)
@@ -239,6 +248,19 @@ struct telemetry_monitor {
 #endif
 };
 
+static void record_telemetry_phase(struct telemetry_monitor *monitor,
+				   const char *suite, int workload_index,
+				   const char *mode)
+{
+	struct telemetry_phase *phase;
+	if (monitor->phase_count >= MAX_TELEMETRY_PHASES) return;
+	phase = &monitor->phases[monitor->phase_count++];
+	phase->elapsed_ms = (uint64_t)((now_seconds() - monitor->started) * 1000.0);
+	phase->suite = suite;
+	phase->workload_index = workload_index;
+	phase->mode = mode;
+}
+
 static int read_number_file(const char *path, double *value)
 {
 	FILE *f = fopen(path, "r");
@@ -1041,7 +1063,9 @@ int fossbench_run(int verbose, int upload_mode, int system_check)
 		fflush(stdout);
 
 		/* Run every test with all cores and one core. */
+		record_telemetry_phase(&telemetry, "asm", (int)i, "multicore");
 		raw_multi[i]  = run_test(&tests[i], (int)g_ncores);
+		record_telemetry_phase(&telemetry, "asm", (int)i, "singlecore");
 		raw_single[i] = run_test(&tests[i], 1);
 
 		printf(" %12.1f  %-11s %7.2fs\n",
@@ -1062,7 +1086,9 @@ int fossbench_run(int verbose, int upload_mode, int system_check)
 	for (i = 0; i < NTESTS; i++) {
 		printf("  %-24s", tests[i].name);
 		fflush(stdout);
+		record_telemetry_phase(&telemetry, "c", (int)i, "multicore");
 		real_multi[i] = run_test(&tests[i], (int)g_ncores);
+		record_telemetry_phase(&telemetry, "c", (int)i, "singlecore");
 		real_single[i] = run_test(&tests[i], 1);
 		printf(" %12.1f  %-11s %7.2fs\n",
 		       display_metric(&tests[i], &real_multi[i]), tests[i].unit,
diff --git a/src/app/upload.c b/src/app/upload.c
index f1098f2..17ed4b1 100644
--- a/src/app/upload.c
+++ b/src/app/upload.c
@@ -246,6 +246,21 @@ static int upload_results(const struct system_info *info,
 				used += (size_t)n;
 			}
 		}
+		n = snprintf(payload + used, sizeof payload - used, "],\"telemetry_phases\":[");
+		if (n < 0 || (size_t)n >= sizeof payload - used) return 0;
+		used += (size_t)n;
+		{
+			size_t i;
+			for (i = 0; i < telemetry->phase_count; i++) {
+				const struct telemetry_phase *phase = &telemetry->phases[i];
+				n = snprintf(payload + used, sizeof payload - used,
+					"%s{\"elapsed_ms\":%llu,\"suite\":\"%s\",\"workload_index\":%d,\"mode\":\"%s\"}",
+					i ? "," : "", (unsigned long long)phase->elapsed_ms,
+					phase->suite, phase->workload_index, phase->mode);
+				if (n < 0 || (size_t)n >= sizeof payload - used) return 0;
+				used += (size_t)n;
+			}
+		}
 		if (used + 4 >= sizeof payload) return 0;
 		memcpy(payload + used, "]}}", 4);
 		payload_len = (int)(used + 3);