Hi all!
I’m looking for some input on AI 9 HX 370 mainboard performance.
There are quite a few specs on the desktop, but clearly from various presentations Framework and AMD acknowledge there is a performance difference between soldered and non-soldered memory, especially when it comes to parallel (“GPU”, “NPU”, “AI”) performance.
I also understand that there are other boards out there with a similar layout [ex. 1, 2, 3], but I am very specifically interested in this exact configuration – chassis dynamics, phase change tech, and all.
Has Framework released any information related to what actual performance one could expect out of the Ryzen AI 9 HX 370 for example? Verbiage like “up to 50 tops” doesn’t exactly inspire confidence.
How much memory is actually available to the NPU in an unsoldered configuration? What is the actual performance hit that caused them to go strictly soldered RAM on the desktop version? It must be pretty significant.
I may be interested in coughing up the $1000 for this board, but I want to see some numbers first.
The quantity of memory available is not impacted by that.
In this case the performance difference is more about LPDDR5X vs DDR5.
LPDDR5X has around a third higher bandwidth (good), but it also has around a third higher latency (bad) than DDR5. The impact varies depending on the task, but typically in GPU or NPU loads the bandwidth is more important (meaning LPDDR5X is better) but in most CPU loads the latency is more important (meaning DDR5 is slightly better). Overall LPDDR5X is typically considered better on average.
However the impact will vary depending on the workload and on the CPU model. Afaik there are not yet any public benchmark results of the HX 370 with DDR5 so I do not know exact numbers.
LPDDR5X however requires better signaling integrity and shorter traces, meaning it is typically soldered. Although there is a new type of memory module (LPCAMM2) with improved integrity to support LPDDR5X.
The CPU in the Framework Desktop only supports LPDDR5X (not DDR5). Additionally it uses a 256-bit wide memory bus (as opposed to 128-bit like most CPUs). The 256-bit wide memory bus made it so that using LPCAMM2 would require a complicated board layout that would cripple memory performance much more severely, which is why they went soldered.
2 Likes