My FW13 7640u 2x16GB 5600 LPDDR5 RAM, clocking at 40GB/s as expected, it’s 2X 64 bit bus dual channel for a total of 128 bit bus.
My 358H with 32GB LPCAMM2 7466 is clocking at 100GB/s. It’s 128 bit bus.
This is WAY faster than it should be? I would have expected around 60GB/s.
Am I missing something here?
I am indeed getting 2.5X performance, that is way more than I expected I would get.
DDR5 5600 dual channel should have theoretical bandwith of 90GB/s, and LPDDR5 ~7500 should have 120GB/s. Effective bandwidth would be ~76GB/s for the former and ~100GB/s for the latter. Which leads me to believe for some reason in your prior configuration you weren’t getting dual channel speeds.
Yeah, i do think there was something wrong with the 7640u memory system.
I happened to make a benchmark when I was running single channel and it was half speed 22GB/s, so it was using both stick at 44GB/s. it was just losing lots of performance for some reasons. Weird. There would be so much performance to unlock with that 7640u MOBO as bandwidth is the bottlneck for lots of workloads.
| Ram Type | Speed [MHz] | Bus Size [b] | Bandwidth [GB/s] | Userbenchmark [GB/s] | Efficiency [%] |
|---|---|---|---|---|---|
| LPDDR5 SODIMM | 5600 | 64 | 44.8 | 22.1 | 49.3% |
| LPDDR5 SODIMM | 5600 | 128 | 89.6 | 43.3 | 48.3% |
| LPDDR5X LPCAMM2 | 7466 | 128 | 119.5 | 100.2 | 83.9% |
Have you tried some other form of memory bandwidth benchmark, aida64 for example?
Userbenchmark is known to have a pretty extreme anti amd bias, not sure that is what is going on but your numbers are way low.
I didnd’t just use syntetic, I tried applications, and across the board there is a 2.5X uplift, in line with the bandwidth uplift.
Biggest giveaway is the LM Studio inference. Prefill is compute bound and had an enormous uplift, that’s mostly the stronger B390 vs 760m compute and vastly superior driver. But the inference speed, that’s horribly bandwidth bound, that being 2.5X is a giveaway that the real bandwidth is 2.5X.
What I do not know, is if it was my 7640u 760m that had something wrong with it, or if it’s an issue with that APU architecture.
Here some more benchmarks
There may have been an issue with yours? Are you sure both your sticks were working?
2.5x is pretty much exactly what you’d expect going from single channel 5600 to 7466lpcamm.
I’m sure it was dual channel because single stick recorded 22GB/s and dual stick recorded 44GB/s,
7640U 96GB Ram Ubuntu
Using LocalLLM
Qwen coder 32GB - 2 token / sec Qwen coder 14B - 4 token / sec Qwen coder 3Gb - 14 token / sec DeepSeek R1 distill Qwen 7B - 9 token / sec
I found a post, and it’s in line with the results I got. It’s likely some limitation of the 7640u memory architecture.
I just ran the userbenchmark thing on my 7840u fw13 and I am getting way higher results than you. Auda64 also reports a bit over 61GB/s read on a dirty windows 11 to go install. I don’t think the 2 extra cores can explain that whole difference.
Still faster ram + better memory controller do give phantherlake quite an edge here. Very curious how the new memory interconnect on zen6 works out but looking at strix halo probably quite well.
