Memory gets judged by one number on the box, usually capacity. Yet a desktop with 32GB of DDR5 can finish many jobs faster than one with 64GB of DDR4, and a server with 512GB of DDR5 RDIMM can leave both far behind on enterprise workloads. Each result is correct, because each system wins on a different property of memory. This guide separates the four that matter: capacity, speed, bandwidth and architecture.
The four properties of memory
- Capacity is how much data fits in memory at once, measured in gigabytes. It decides whether the workload fits or spills to storage.
- Speed is the transfer rate of each module, measured in megatransfers per second (MT/s), such as DDR4-3200 or DDR5-6000. Latency, the delay before data starts arriving, is part of this picture.
- Bandwidth is how much data the whole memory system moves per second. It equals speed times bus width times the number of channels.
- Architecture covers how the memory and platform are built: DDR4 or DDR5, unbuffered or registered, how many channels the processor has, and how errors are handled.
No single property predicts performance. The workload decides which one becomes the bottleneck.
Three systems side by side
Take three typical configurations. The bandwidth figures are theoretical peaks, calculated as transfer rate times 8 bytes per channel times the number of channels.
| System A: desktop | System B: desktop | System C: dual-socket server | |
|---|---|---|---|
| Memory | 64GB DDR4-3200 (2 x 32GB) | 32GB DDR5-6000 (2 x 16GB) | 512GB DDR5-4800 RDIMM (16 x 32GB) |
| Memory channels | 2 | 2 | 16 (8 per processor) |
| Theoretical peak bandwidth | 51.2 GB/s | 96.0 GB/s | 614.4 GB/s |
| Typical latency in nanoseconds | 10 ns (CL16) | 10 ns (CL30) | About 16.7 ns (CL40) |
| Error correction | None on standard modules | On-die ECC inside each chip | Full ECC across the memory bus |
| Wins when | The job needs 33 to 60GB | The job fits in 32GB and moves a lot of data | The job needs hundreds of gigabytes or many cores fed at once |
Why 32GB of DDR5 can beat 64GB of DDR4
System B has half the capacity of System A but nearly twice the bandwidth: 96 GB/s against 51.2 GB/s. When the workload fits in 32GB, capacity beyond that sits idle, and bandwidth decides how quickly the processor gets its data. Games, integrated graphics (which use system memory as video memory), code compiles, compression and many creative tools move large amounts of data through memory continuously. On those jobs, the faster bus shows up directly as shorter run times and higher frame rates.
Latency does not cancel the gain. DDR5 has higher CAS latency numbers, but its clock is faster, so the delay in nanoseconds is close. True latency equals CAS latency times 2,000 divided by the transfer rate: CL16 at DDR4-3200 is 10 ns, and CL30 at DDR5-6000 is also 10 ns.
DDR5 also changes the internal design. Each DDR5 module splits its 64-bit bus into two independent 32-bit subchannels, doubles the burst length to 16, and has 32 banks against 16 on DDR4. More banks and more independent channels let the memory controller keep more requests in flight, which helps when many cores read memory at once. For the full history, see from magnetic core to DDR5.
One fair caveat: a DDR5 desktop also has a newer processor, since DDR4 and DDR5 use different slots and platforms. Part of System B's lead comes from the CPU. Memory bandwidth is still a large share of it on data-heavy work.
When 64GB of DDR4 wins
Capacity matters at a sharp threshold. While the working set fits in memory, extra gigabytes add nothing. Once it exceeds memory, the operating system pages data to the SSD. DRAM answers in roughly a hundred nanoseconds; even a fast NVMe SSD takes tens of microseconds, hundreds of times longer. Performance falls off a cliff rather than slowing gradually.
That makes System A the better machine for an 8K video timeline, a 40GB dataset in memory, a large Photoshop composite or several virtual machines at once. If your workload pushes past 32GB, slower memory that holds everything beats faster memory that cannot. To see how much you use today, check RAM type, speed and capacity in Windows 11 and watch memory use while your heaviest job runs.
Why 512GB of DDR5 RDIMM outperforms both
System C wins on every axis that enterprise software stresses at once. It has 16 times the capacity of System B, and it has 16 memory channels across two processors, for a theoretical 614.4 GB/s. That is about six times System B and twelve times System A. Server processors are built around this: 4th and 5th Gen Intel Xeon Scalable processors have 8 DDR5 channels per socket, and AMD EPYC 9004 and 9005 processors have 12.
The architecture is what makes that capacity possible. A registered DIMM (RDIMM) has a registering clock driver that buffers command and address signals, so each channel can drive larger modules and more of them. DDR5 RDIMMs also carry full error-correcting code across the memory bus, plus the reliability features servers need to run for years. The on-die ECC built into every DDR5 chip only protects data inside the chip, so it is not a substitute. See our ECC vs non-ECC guide and memory types explained for more.
Applications that benefit most:
- In-memory databases and caches keep entire datasets in RAM, so capacity sets the dataset size and bandwidth sets query speed.
- Virtualization hosts divide memory among virtual machines. More capacity means more VMs per host and fewer hosts to buy.
- Analytics and simulation run dozens of cores that all read memory at once. Without enough channels, cores sit idle waiting for data.
- AI inference on the CPU reads the full model weights for every token it generates, which makes it one of the clearest bandwidth tests there is.
A server does not win everything. Server memory usually runs at lower speeds and slightly higher latency than enthusiast desktop memory, and a single thread on a fast desktop chip can finish first. The server advantage is aggregate: total capacity and total bandwidth across many cores. Our article on CPU bottlenecks vs memory bottlenecks shows how to tell which one limits you.
A worked example: running a language model on the CPU
Local language models show all four properties at once. Generating each token requires reading the model's weights from memory, so the model must fit (capacity), and the maximum token rate is roughly bandwidth divided by model size (bandwidth). These are rough theoretical ceilings; real results are lower.
| Model (4-bit, approximate size) | System A: 51.2 GB/s, 64GB | System B: 96 GB/s, 32GB | System C: 614.4 GB/s, 512GB |
|---|---|---|---|
| 8B parameters (about 5GB) | About 10 tokens/s ceiling | About 19 tokens/s ceiling | About 120 tokens/s ceiling |
| 70B parameters (about 40GB) | Fits; about 1.3 tokens/s ceiling | Does not fit in memory | Fits; about 15 tokens/s ceiling |
| 400B+ parameters (200GB or more) | Does not fit | Does not fit | Fits |
With a small model, System B beats System A, purely on bandwidth. With a 70B model, System B cannot run it at all, and System A can, slowly. System C runs both, faster, and handles models neither desktop can load. For sizing guidance, see how much RAM an AI workstation needs and do AI servers need more system RAM.
Which matters for your workload
| Workload | Usually limited by | Best of the three |
|---|---|---|
| Gaming, integrated graphics | Bandwidth and latency | System B (32GB DDR5) |
| Code compiles, compression, photo editing | Bandwidth, with modest capacity | System B, if the project fits in 32GB |
| 8K video timelines, large Photoshop files, several local VMs | Capacity | System A (64GB DDR4) |
| In-memory databases, caching layers | Capacity, then bandwidth | System C (512GB RDIMM) |
| Virtualization hosts with dozens of VMs | Capacity per host | System C |
| Analytics, simulation, CPU inference on large models | Bandwidth across many cores | System C |
A simple rule: first buy enough capacity that your largest job fits with room to spare. After that, spend on bandwidth: faster modules on desktops, and every memory channel filled on servers and workstations. A server with half its channels empty gives up half its bandwidth, whatever the capacity. Our guide to dual, quad and eight-channel memory explains channel population, and why RAM runs slower when all slots are filled covers the speed trade-off of adding more modules per channel.
Matching memory to the job
NEMIX RAM carries memory for all three cases. For desktops, browse desktop memory and DDR5 UDIMM. For servers and workstations, see DDR5 RDIMM, DDR5-6400 RDIMM and DDR4 RDIMM, or start from our data center server memory page. Planning a large build? Read how much 1TB of RAM costs.
Frequently asked questions
Is 32GB of DDR5 faster than 64GB of DDR4?
Often, as long as the workload fits in 32GB. A dual-channel DDR5-6000 system has a theoretical peak of 96 GB/s against 51.2 GB/s for dual-channel DDR4-3200. Once the working set grows past 32GB, the 64GB DDR4 system wins because the DDR5 system has to page to storage.
Does more RAM make a computer faster?
Only when the current amount is too small. Extra capacity that the workload never touches adds no speed. Capacity prevents slowdowns; bandwidth and latency decide how fast the system runs once everything fits.
Is DDR5 latency worse than DDR4?
CAS latency numbers are higher on DDR5, but the clock is faster, so the time in nanoseconds is similar. DDR4-3200 at CL16 and DDR5-6000 at CL30 both work out to 10 nanoseconds.
Why do servers use RDIMMs instead of desktop memory?
A register on each RDIMM buffers command and address signals, which lets a server run more and larger modules per channel with full ECC. That is how servers reach hundreds of gigabytes or terabytes of memory. Our guide on why server motherboards use RDIMM covers the details.
Is a server faster than a desktop for everything?
No. A server wins on total capacity and total bandwidth across many cores. A single thread on a fast desktop processor can finish sooner than on a server core, and desktop memory often has slightly lower latency.
How do I calculate memory bandwidth?
Multiply the transfer rate in MT/s by 8 bytes per 64-bit channel, then by the number of channels. DDR5-4800 across 16 channels is 4,800 x 8 x 16 = 614.4 GB/s. That is a theoretical peak; measured results are lower.
Buying for more than one system
IT teams and integrators that maintain desktops, workstations and servers with different memory needs can standardize on one kit per model. Request a quote with your models and quantities, or send us your system model and we will confirm the right module before you order.
Bandwidth and latency figures are theoretical values calculated from published memory specifications. Measured performance depends on the processor, platform, memory population and workload.