08:10 <icecream95> On x1e-78-100, llama.cpp with Turnip has about 30% better token generating performance than the CPU for FP16 models (*and* it uses less power). Since this is memory bandwidth bound, that indicates the GPU can get more read bandwidth than the CPU. Exactly why, I don't yet know

08:11 <icecream95> But this required GGML_VK_FORCE_MAX_ALLOCATION_SIZE=2147483647, since a too high maxBufferSize is reported, and even maxMemoryAllocationSize=0x80000000 is wrong because that size will get *sign-extended* to 64-bits in the msm kernel driver

09:02 <icecream95> I think it would be fun ("cursed") to virtualize the GPU well enough that the Windows driver can be run unmodified inside a VM from Linux... for testing and comparison only, of course. The GPU SMMU has two translation stages, so that part at least isn't a problem

12:53 <rfs613> so curiously, after applying https://download.lenovo.com/pccbbs/mobiles/n3hqc13w.html the charging LED seems to be on permanently (orange, not white). Windows says battery is at 100% and not charging.

12:54 <rfs613> LED remains orange even when I shutdown the machine. Only turns off momentarily if I press the pinhole reset button.

13:24 <Jasper[m]> Sounds like a fun ec bug

15:37 <rfs613> well hopefully it doesn't cause the battery to swell and explode... because I rather like the x13s ;-0

15:41 <zdykstra> just poke a small hole in the battery to help drain the excess electricity

16:39 <wizzard> Which quantisation are you using

16:39 <wizzard> > On x1e-78-100, llama.cpp with Turnip has about 30% better token generating performance than the CPU for FP16 models (*and* it uses less power). Since this is memory bandwidth bound, that indicates the GPU can get more read bandwidth than the CPU. Exactly why, I don't yet know