How to use this test
- Close other applications. Memory bandwidth is shared across everything running. A busy machine measures lower.
- Run the benchmark. It allocates 64 MB and runs three separate measurements, each several times.
- Compare sequential against random. The ratio between them reveals how well your cache hierarchy is coping — often more informative than either figure alone.
- Read the latency figure. Dependent reads cannot overlap, so that number reflects real memory latency rather than throughput.
Three different questions
"How fast is my memory" is really three questions, and they have different answers. This test measures all three because a single number would hide the interesting part.
| Measurement | What it exposes | What real work it resembles |
|---|---|---|
| Sequential | Peak streaming bandwidth | Copying large files, video editing, loading assets |
| Random | Bandwidth once prefetching fails | Databases, compilers, most real programs |
| Latency | Cost of one dependent access | Pointer-heavy code, linked structures, game logic |
Why sequential access looks so fast
Processors are excellent at predicting sequential access. Walk through memory in order and the prefetcher fetches the next block before you ask for it, so most reads are served from cache rather than from memory. The figure you measure is closer to cache bandwidth than to RAM bandwidth.
The random test uses a large prime stride specifically to defeat that prediction. Each access lands somewhere the prefetcher did not anticipate, so the measurement reflects what happens when caching stops helping — which is closer to how most software actually behaves.
Why latency is measured separately
Bandwidth tests let many requests overlap: modern processors keep multiple memory accesses in flight at once, which hides latency behind throughput. To measure latency you have to prevent that overlapping, and the standard method is pointer chasing — each read returns the address of the next, so nothing can begin until the previous one finishes.
That is what the latency figure represents: the real cost of one dependent memory access, in nanoseconds, with no help from parallelism.
What the numbers suggest
- Sequential to random ratio of 2–5×. Normal. Your cache hierarchy is behaving as designed.
- Ratio above 6×. Unusually wide. Often means other software is evicting your data from cache — close things and re-run before concluding anything.
- Latency above roughly 180 ns. High. Background load and battery power profiles both inflate this considerably.
- Low bandwidth on a laptop. Check whether you have one memory stick or two. Single-channel halves your bandwidth and is extremely common in budget machines.
Related tests
The CPU benchmark includes memory access in its workload, so a machine with slow memory scores lower there too. System information shows how much memory your browser is willing to admit to, and explains why that figure is deliberately vague.
Frequently asked questions
Why can this not tell me my RAM speed or timings?
Because no browser exposes them. There is no API for memory frequency, timings, channel configuration or manufacturer — those live far below the sandbox boundary. What a browser can do is allocate memory and time how fast it can be accessed, which measures the practical result of all those specifications combined.
Why is sequential access so much faster than random?
Cache and prefetching. Your processor predicts sequential access and fetches data before it is requested, so streaming through memory in order hits cache almost every time. Strided access with a large prime stride defeats that prediction, forcing real memory accesses. A gap of 2 to 5 times is normal and healthy; a much larger one suggests something is competing for cache.
What does the latency figure actually measure?
Pointer chasing. Each read determines the address of the next read, so nothing can be fetched in parallel or predicted ahead. That serialises the accesses and exposes true latency, which throughput tests hide. It is the classic method for measuring memory latency and it is why the number is in nanoseconds rather than gigabytes per second.
Is more bandwidth better?
Up to a point, and less than marketing suggests. Most everyday software is limited by latency and cache behaviour rather than raw bandwidth. Integrated graphics are a genuine exception, because they share system memory with the CPU and are frequently bandwidth-starved — which is why dual-channel memory improves integrated GPU performance dramatically while barely affecting a desktop with a discrete card.
Would adding a second memory stick help?
Very likely, if you currently have one. A single stick runs in single-channel mode, halving available bandwidth. Adding a matched second stick enables dual-channel and typically improves memory throughput substantially — and on a laptop with integrated graphics it can improve gaming performance by a large margin for very little money. This is one of the best value upgrades available.
Why does my score change between runs?
Memory is shared. Other applications, the browser itself, and garbage collection all compete for bandwidth and cache. The test runs each measurement five times and takes the median to reduce this, but a genuinely busy machine will still measure lower. Close things and re-run for a cleaner figure.