How Much Memory Bandwidth Does My System Have?
By Devin Chua · Data checked by Owen Nakamura · Updated Every channel moves 64 bits, which is 8 bytes, per transfer. A DDR5 DIMM splits internally into two 32-bit sub-channels: that improves efficiency but does not change the total, so the arithmetic is the same for DDR4 and DDR5. Populating a second DIMM per channel adds capacity, not bandwidth, and on most desktop platforms it forces a lower stable speed.
Theoretical memory bandwidth is channels times transfer rate times eight bytes. A dual-channel DDR5-6000 system peaks at 96 GB/s, and real benchmarks land around 70 to 85% of that. The arithmetic is identical for DDR4 and DDR5, because a channel moves 64 bits per transfer in both. Channels set bandwidth, not the number of modules you fit.
The formula
GB/s = channels x MT/s x 8 bytes / 1000
every channel moves 64 bits = 8 bytes per transfer
a DDR5 DIMM is two 32-bit sub-channels internally, which does not change the totalWorked examples
| Case | Inputs | Working | Result |
|---|---|---|---|
| Desktop DDR5 | 2 channels, 6000 MT/s | 2 x 6000 x 8 / 1000 | 96 GB/s |
| Desktop DDR4 | 2 channels, 3200 MT/s | 2 x 3200 x 8 / 1000 | 51.2 GB/s |
| 8-channel server | 8 channels, 4800 MT/s | 8 x 4800 x 8 / 1000 | 307.2 GB/s |
Where the formula comes from
Every memory channel moves 64 bits, which is 8 bytes, per transfer. Multiply by the transfer rate and by the number of channels and you have the theoretical peak. A dual-channel DDR5-6000 system moves two channels times six billion transfers times eight bytes, which is 96 GB/s.
The arithmetic is identical for DDR4 and DDR5, which surprises people who have read that DDR5 DIMMs are internally split into two 32-bit sub-channels. That split improves access efficiency, because two smaller requests can be in flight independently, but it does not change the total width. A DDR5 DIMM still presents 64 bits per transfer.
Why your benchmark reports less
Refresh cycles take the memory offline periodically. Bank conflicts stall requests that happen to target the same internal structure. Command overhead consumes cycles that carry no data. The memory controller itself has limits independent of the modules. Together these put real-world results somewhere around 70 to 85% of theoretical peak, and a benchmark landing in that band is behaving normally rather than indicating a problem.
Channels, not sticks
The most expensive misconception here is that adding modules adds bandwidth. It does not. Bandwidth comes from channels, which are a property of the processor and board, and mainstream desktop platforms are dual-channel regardless of whether the board has two slots or four. Populating four slots adds capacity, and on most desktop platforms it forces a lower stable transfer rate, so four sticks frequently run slower than two.
Channel counts worth knowing: mainstream desktop is two, Threadripper is four or eight depending on the line, and current EPYC and Xeon Scalable are eight or twelve. If bandwidth genuinely is your constraint, a platform change does far more than any kit upgrade. Whether bandwidth is your constraint depends on the workload: integrated graphics, large in-memory datasets and many-core compute are bandwidth-bound, while gaming on a discrete card usually is not.
Questions
Why is my measured bandwidth lower than this?
Refresh cycles, bank conflicts, command overhead and the memory controller's own limits all take a share. Seventy to eighty-five percent of theoretical peak is a normal result, and the figure the benchmark reports is the one that matters for real work.
Does adding a second DIMM per channel double bandwidth?
No. Bandwidth comes from channels, not from the number of modules. A second DIMM per channel adds capacity, and on most desktop platforms it forces a lower stable transfer rate, so four sticks often run slower than two.
Is the calculation different for DDR5?
No. A DDR5 DIMM splits into two independent 32-bit sub-channels, which improves access efficiency, but the total per DIMM is still 64 bits per transfer. The formula is the same for both generations.
How many channels does my system have?
Mainstream desktop platforms are dual-channel regardless of how many slots the board has. Threadripper is four or eight depending on the line, and current EPYC and Xeon Scalable are eight or twelve. Slot count is not channel count.
Does more bandwidth make my machine faster?
Only for workloads that are memory-bound: integrated graphics, large in-memory datasets, video encoding and many-core compute. For gaming on a discrete card and ordinary desktop use, latency and capacity usually matter more than peak bandwidth.