Measures EC2 instance performance using CPU Passmark and makes the results available.
Add regions and/or families. With regions selected you get every instance type's score per region (families narrow the rows); with only families selected you get the classic size-by-family comparison. Toggle the metric (CPU score or value per $). Cells are heat-coloured.
Instance  |
|---|
| Instance |
|---|
| Instance |
Rank |
Score |
vCPU |
RAM |
On Demand Price (Hourly)** |
Value Score |
Added |
Last Updated |
|
|---|---|---|---|---|---|---|---|---|---|
| Region |
Instance |
Rank |
Score |
vCPU |
RAM |
On Demand Price (Hourly)** |
Value Score |
Added |
Last Updated |
|
|---|---|---|---|---|---|---|---|---|---|---|
NVIDIA GPU instances benchmarked with PyTorch — matrix-multiply throughput (TFLOPS), HBM memory bandwidth, and ResNet-50 inference. Heat-coloured per column.
| Instance |
Region |
Rank |
GPU |
Count |
GPU Mem (GB) |
FP32 TFLOPS |
TF32 TFLOPS |
FP16 TFLOPS |
Mem BW (GB/s) |
ResNet-50 (img/s) |
On Demand Price (Hourly) |
Value Score |
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Every benchmarked instance type plotted by hourly price (x, log scale) against CPU Passmark score (y, log scale). Up-and-left is better value. Colour by family group; hover a point for detail. Diagonal guide lines mark equal value-for-money.
How much of each region's offered EC2 fleet has been benchmarked, and how fresh the data is. Offered counts are live from EC2 (excluding mac/a1). Unreachable regions are flagged.
| Region | Offered | Benchmarked | Coverage | Last swept |
|---|
Instance types by when they first appeared in our benchmarks (FirstSeen) — a rough feed of what AWS has been rolling out. Newest first; one row per instance type (earliest region seen).
| Instance |
Family |
Score |
vCPU |
First seen |
Region |
|---|
Every benchmarked result grouped by the actual CPU silicon (Intel / AMD / AWS Graviton), not instance family. Ranked by median CPU Passmark score and by best value (score per $/hr). The Graviton-vs-x86 question, with real numbers.
| Processor |
Vendor |
Samples |
Median score |
Peak score |
Best value |
|---|
Rank the whole fleet by the sub-benchmark that matches your workload — not the blended score. Pick a metric; results re-rank by raw throughput and by throughput per $/hr.
| # | Instance |
Region |
Metric |
Price/hr |
Per $/hr |
|---|
The other dimension: RAM bandwidth and latency from the memory benchmark (SUMM_ME). For in-memory databases, caches, and analytics where memory, not CPU, is the bottleneck. Sorted by memory score.
| Instance |
Region |
Mem score |
Latency (ns) |
Read (large) |
Write |
Threaded |
RAM (GB) |
|---|
Instance families (rows) across every region (columns), median per cell. In value mode each family is coloured against its own cross-region range — green where that family is best value-for-money, red where it's dearest — so geographic arbitrage jumps out. In CPU score mode colour is a global scale. Hover for detail, click to open that region.
Scroll to zoom · drag to pan