Thread 106630455

How important is RAM speed in real world scenarios?

>>106630455 (OP)
depends on use case

it depends. for most a percentage point doesnt matter at all but for me even a hypothetical 0.01% is enough to take action.

>>106630455 (OP)
Use case dependent.

>>106630455 (OP)
It's free performance.

>>106630455 (OP)
Depends on your other bottlenecks. For gaming, CPU-bound situations tend to the only ones to benefit.

>>106630455 (OP)
Less you have the faster you need it

Tighter timings often matter more than raw speed for gaming. Relational databases need low latency to function because queries often involve chains of lookups. The more you have to hit RAM (when stuff isn't in L2/L3 cache), the more important that becomes because every nanosecond you don't have that data is time the processor is idling.

Gaming is a bit weird because you often get higher average framerates with fast ram, but the 1% and .1% lows get better with tight timings. The .1% lows in particular are usually related to having to go fetch some data that points to other data that points to other data that lets you calculate something that tells what what other data you need to fetch. Even 1 nanonsecond delays on each step can add up to appreciable microstutters when you need to perform hundreds or thousands of lookups as part of a single frame render or core game logic loop. That's similar to relational databases, but modern games do a lot of loading and unloading large textures, so raw speed is still important. Not all games are going to care about this though, and even for the ones that do you run into pretty hardcore diminishing returns. In general it's not worth the price premium for the absolute top shelf memory unless you're spending many thousands of dollars on the system. Save the 100-150 bucks and buy a better case, a bigger SSD, go up a tier on the processor, or just pocket it.

AI workloads seem to largely care about speed more than latency because they're shuffling huge amounts of shit around, and aren't doing enormous amounts of random lookups and table sleuthing. Even then though, some part of AI training does behave like a database, so it's not like latency isn't important, it's just not as do or die as it is on a production database that needs 2+TB of RAM to function.

>>106631118
In practice ram bandwidth can be more important than latency for 1% lows etc. It depend. One thing fast ram is better at than average ram is keeping the i/o queue size small leading to lower latency ram accesses even under memory load. Otherwise generally low ram latency is ofc nice for performance.

https://old.chipsandcheese.com/2025/08/20/skymont-in-gaming-workloads/

>>106630455 (OP)
Mostly irrelevant. It is just epenis and min-max types they are fretting over the marginal gains from getting factory overclocked SKUs.
The only avenue where it is actually matters is in HPC scene not silly gayming rigs.

>>106630537
reducing timings is free performance, assuming stability. Raising clocks costs power / heat and can create subtle problems but is generally easier.