Sample results
This is what a Spartan Server Labs engagement produces: measured results for every BIOS profile, application benchmarks at increasing load, and an interactive dashboard.
1. Sample test results
Each profile is tested at 25, 50, 75 and 100% load, with power measured at every step, then run through a burn-in. The table compares three profiles on the same demonstration server.
| BIOS profile | CPU (blocks/s) | Memory (GB/s) | Disk (MB/s) | System power at full CPU load (W) | CPU blocks/s per watt | Burn-in |
|---|---|---|---|---|---|---|
| Power Efficient | 326.8 | 24.0 | 1,900 | 215 | 1.52 | PASS, 4 h, max 72.9 °C |
| Low Latency | 380.0 | 29.0 | 1,900 | 275 | 1.38 | PASS, 4 h, max 74.0 °C |
| HPC | 425.6 | 33.0 | 1,900 | 308 | 1.38 | PASS, 4 h, max 75.0 °C |
Network, GPU and DPU tests were skipped in this sample because the demonstration server has none of those devices. A real report says plainly what was and was not tested.
Power draw rises with load
Work done at each load level
In this sample the HPC profile does the most work but draws the most power, while Power Efficient does the least work for the least power. The right choice depends on your workload, which is exactly what the comparison is for.
2. Sample application benchmark
Application workloads are run at increasing numbers of concurrent connections, to show where throughput peaks, where latency starts to climb and how much power each result costs. This sample shows a database benchmark (pgbench), then a summary of four workload types.
Database throughput against connections
Tail latency (99th percentile)
| Connections | p99 latency, Power Efficient | p99 latency, Low Latency | p99 latency, HPC |
|---|---|---|---|
| 100 | 8.0 ms | 8.0 ms | 8.0 ms |
| 500 | 37.1 ms | 33.0 ms | 30.3 ms |
| 1,000 | 124.3 ms | 108.0 ms | 97.3 ms |
| 2,000 | 269.6 ms | 233.0 ms | 208.9 ms |
| 5,000 | 1,868.5 ms | 1,608.0 ms | 1,436.6 ms |
Summary of four workload types
| Workload | Tool | BIOS profile | Peak throughput | Highest stable level | Power at peak | Throughput per watt |
|---|---|---|---|---|---|---|
| Database | pgbench | Power Efficient | 3,870 transactions/s | 2,000 | 215 W | 18.0 |
| Low Latency | 4,500 transactions/s | 2,000 | 275 W | 16.4 | ||
| HPC | 5,040 transactions/s | 2,000 | 308 W | 16.4 | ||
| Web serving | HTTP load | Power Efficient | 22,360 requests/s | 5,000 | 215 W | 104.0 |
| Low Latency | 26,000 requests/s | 5,000 | 275 W | 94.5 | ||
| HPC | 29,120 requests/s | 5,000 | 308 W | 94.5 | ||
| Storage | fio random read | Power Efficient | 206,400 IOPS | 256 | 215 W | 960.0 |
| Low Latency | 240,000 IOPS | 256 | 275 W | 872.7 | ||
| HPC | 268,800 IOPS | 256 | 308 W | 872.7 | ||
| AI inference | PyTorch MLP | Power Efficient | 38,700 samples/s | 1,024 | 215 W | 180.0 |
| Low Latency | 45,000 samples/s | 1,024 | 275 W | 163.6 | ||
| HPC | 50,400 samples/s | 1,024 | 308 W | 163.6 |
Application workload tests are a new extension offered after a pilot, as described on the Services page. The tools shown are examples; we can run your own benchmark command instead.
3. Sample dashboard
The interactive dashboard works offline and lets you explore every result: inventory, power, performance, burn-in and application benchmarks, by profile. Try the tabs below, or open it full screen.
Synthetic data from the fictional server DEMO-SRV. The dashboard runs entirely in your browser. It makes no network requests and sends nothing anywhere.
Not sure which profile your servers should run?
Tell us about your servers. We will explain how a pilot would work, what it would take, and what you would receive.
Request a pilot