What's interesting here imho, is the staggering difference it makes on an CPU integrated GPU (which makes sense since the GPU is using the comptuer RAM instead of its own RAM).
Also, we've been told DDR5 had built-in dual channel on a single stick so I was not sure this would make any difference at all.
> Also, we've been told DDR5 had built-in dual channel on a single stick so I was not sure this would make any difference at all.
It's really necessary to start thinking in terms of the total width of the memory bus. DIMMs have a 64-bit wide connection. Dual-channel used to only mean half of your DIMM slots were on channel A and half of them were on channel B, for a total of a 128-bit wide memory bus if you populated both channels.
DDR5 splits the DIMM's 64-bit connection into two sub-channels of 32 bits each. But mainstream CPUs still have the same 128-bit total width for their DRAM controllers, so you still need more than one DIMM installed to use all the (sub)channels provided by the CPU's memory controller.
Comparing the 128-bit wide memory bus against the bus widths used by discrete GPUs (and comparing the memory clock speeds) makes it much less surprising that running integrated graphics in a crippled 64-bit wide memory configuration would be problematic. Graphics is a very bandwidth-hungry task.
https://greatpcreview.com/guides/cas-latency-vs-ram-speed/
If you learn about latency and RAM timings, and also how the processor/CPU interacts with memory via the controllers and kernel then you get a much deeper and valuable understanding.
Most people confuse “correct configuration” with “more powaaa!”
By analogy: you have now tuned your car and it is running properly. It’s not that it has “more horsepowaaaa!” But rather it can utilize the power it does have effectively.
Even if both channels are populated, there can also still be additional benefits from bank interleaving - where you have two memory ranks per channel - particularly on AMDs Zen architecture family (but Intel as well).
Gamers Nexus backed it up with a 27 minute video in late 2020, with plenty of benchmarks to demonstrate that there can be measurable improvement from bank interleaving. https://youtu.be/-UkGu6A-6sQ
Most people will probably think you're talking about rams as in the animal, or maybe the truck.
Computer literacy is on its way down, not that it was particularly high in the first place, so it's foolhardy to just assume some random Joe would know what a computer is, let alone RAM, let alone channels.
Well, I really meant "HN readers" when I said "people". Totally agree with your point though. Kids don't think technology is cool anymore. When I was growing up even the jocks knew how to burn a CD or put a file on a flash drive.
I recall reading some post on a programmer's blog a couple years ago. He lost all his data because he didn't know how to run backups properly; he was the kind of guy who would call over a friend to help him fix something as trivial as badly seated RAM (he wouldn't know it was something trivial, obviously).
Yes. That is a programmer with literally no knowledge of what a computer is or how it works, even at a very rudimentary level.
Ever since I've been careful to never assume anyone knows anything about computers until they demonstrate otherwise, even in a seemingly techy place like HN.
People seem generally unaware of the difference between number of channels, width of channels, and number of dimms.
In fact this question is wrong, or at least misleading. The OP went from one DDR5 dimm (with two 32 bit channels) to 2 DDR5 dimms with a total of 4 32 bit channels. Thus doubled the bandwidth, but also doubled the number of memory quests you can have in flight.
Given that GPUs in general are quite bandwidth intensive, not surprising that graphics performance significant increased. This is the main reason why the M1 pro -> M1 max doubles the memory bandwidth, it doubles the iGPU performance in many cases.
