Measurement
MEM Score
One number on a log scale. The reference machine scores 50; ten times the reference scores 100.
Formula
Score = 50 + 50 · log₁₀( Π (xᵢ ÷ refᵢ)^wᵢ )
Each metric is divided by a fixed reference machine, weighted, and combined as a weighted geometric mean. Latency enters inverted because lower is better. The server recomputes the score from the submitted metrics; the result is rounded to one decimal. Score version: mem-1.
Reference and weights
| Metric | Reference | Weight |
|---|---|---|
| Read | 100 GB/s | 0.25 |
| Write | 100 GB/s | 0.15 |
| Copy | 100 GB/s | 0.20 |
| Random | 20 GB/s | 0.15 |
| Sustain | 100 GB/s | 0.20 |
| Latency (inverted) | 800 ns | 0.05 |
Reading the scale
| Throughput vs reference | 0.1x | 0.5x | 1x | 2x | 4x | 10x | 30x |
|---|---|---|---|---|---|---|---|
| MEM Score | 0.0 | 34.9 | 50.0 | 65.1 | 80.1 | 100.0 | 123.9 |
Score and percentile
The score says how much memory throughput a machine delivers. It is not a verdict on the machine: that is the job of the class percentile, and the reason every board ranks inside a device class.
AI decode ceiling
Single-stream token generation reads every weight once per token, so tokens per second cannot exceed read throughput divided by the model's size in bytes. The result screen shows this upper bound; real runtimes land below it.
| Read throughput | 8B at 4-bit (4 GB) | 70B at 4-bit (35 GB) |
|---|---|---|
| 100 GB/s | 25 tokens/s | 2.9 tokens/s |
| 430 GB/s | 108 tokens/s | 12.3 tokens/s |
| 1000 GB/s | 250 tokens/s | 28.6 tokens/s |