George Hotz Programming Tinygrad Rdna3 Assembly Unstable Amd Drivers Crashes Part 1

Date of stream 6 May 2023.. from $1499 buy https://comma.ai/shop/comma-three. Live-stream chat added as Subtitles/CC – English (Twitch Chat) – three-dot menu icon – Show transcript. . Source files:. – https://github.com/geohot/tinygrad. Follow for notifications:. – https://twitch.tv/georgehotz. Support George:. – https://twitch.tv/subs/georgehotz. . Chapters:. 00:00:00 red vs green,rocm bugs. 00:01:30 7900xtx true flops,mojo,opencl. 00:03:00 bad assembler,v_dual_fnac_f32,ashrrev_i32_e32. 00:04:30 epyc,nice amd gpu,rocm-bandwith-test. 00:05:28 NULL pointer,reboot,cox modem . 00:07:00 not optimized,cheapETH. 00:09:05 how gpu work,warps,simd,DSP,GPU,CPU,load,not a load. 00:10:55 WGP=2CU,warp,dual issue VALU. 00:12:10 assembly,extra,rdna3,enable_early_exec,kernel. 00:13:05 staging worse,compiling empty program,binary out,stuff to compile rocm. 00:14:55 llvm-objdump -d cc.elf,asm.py,llvm-mc,obj file,lld. 00:17:04 failed to get code object ISA,llvm-readelf cc.elf. 00:18:25 shader from RDNA3,tensor vs non tensor cores float32,llvm target to amdgpu. 00:19:45 llvm target triples,failed to get metadata. 00:21:55 wavefront_size,how wide simd path is,NT_AMDGPU_METADATA,MsgPack binary. 00:23:45 amd_comgr_get_data_metadata,kernel crash,rmmod amdgpu,reseting pci. 00:28:15 metadata,wavefront_size 128,failed to read ip discovery. 00:30:35 git assembler maybe,removing the bus,reboot. 00:32:20 dkms,stock ubuntu 22.04 kernel,connecting bmc. 00:33:50 connection not private,thisisunsafe. 00:35:20 event_log,pci serr critical error. 00:36:50 debugfs to restart gpu,bmc remote kvm. 00:43:30 case,pcie errors. 00:45:22 we buy Intel+Nvidia,want to like AMD. 00:46:00 asm.py OUT_OF_HOST_MEMORY,AMD_LOG_LEVEL=4,roadata. 00:54:20 llvm,gnu export symbol,.global. 00:57:50 symbol code.kd,word gnu assembler. 01:04:40 OBJECT FUNC symbol type assembler,.type code.kd,STT_OBJECT. 01:09:00 why AMD,creating a buffer. 01:11:25 args,running custom assembly. 01:15:00 memory access fault by GPU node-1. 01:26:30 working,test number 2,commit custom asm,crashed gpu. 01:28:00 reset gpu,echoing a string,one works one hangs. 01:33:40 GLOBAL_STORE_B32,S_LOAD_B64. 01:41:50 not one crash with CUDA. 01:42:55 bad mb,suggestion from chat. 01:46:30 frustrating AMD,now on quiet,kernel 5.15.0-71,single 7900. 01:47:50 rocm to 5.5,rocm_clone.sh,new pc,graveyard of gpus. 01:49:30 anyone from AMD is watching this. 01:50:35 AMD sponsor tinycorp, compute driver. 02:02:50 userspace should never pagefault in kernel. 02:03:20 dropping 0days on AMD, flags. 02:04:40 money, can’t search twitter login why Elon. 02:04:55 I will fix drivers,LBTQ. 02:07:00 nvidia datacenter vs gaming,libamdoci-orca64. 02:13:35 AMD care,we love you, nvidia-nvlink. 02:19:40 bad pcie,not enough pci lanes. 02:23:20 product for tinycorp,home DGX, writing a fuzzer,fuzz.py. 02:31:55 same crash,rocm-smi gpureset. 02:36:30 map_queues_cpsch,dqm lock has to be locked. 02:45:40 gpu P2P disabled,cuda ioctl SM_ISSUE_RATE,FP16 vs FP32 accumulate. 02:53:00 new crash,using pyopencl,normal ubuntu,linux 6.3,kernel crashes. 02:57:00 build something better than they have,companies respond to fear,still use an iPhone. 03:04:45 Android spyware,write AMD driver,Jim Keller,tenstorrent. 03:07:00 marketcap show driver stability, apple value brand. 03:10:10 apple shader code clean. 03:12:55 advice to nvidia,,4080 stupid,Stevey’s Google Platforms Rant, deal per TFLOP. 03:16:30 Alex home,qualcomm, kernel 6.3. 03:26:30 crashed,website driver. 03:36:15 rebuild initramfs,stable until it crashed. 03:43:46 normal install,RDNA2 card test,rocm 5.5. 03:49:30 Steve Jobs nvidia. 04:02:25 gpu-burn. 04:09:00 build kernel from source,menuconfig,Linus let this in kernel. 04:15:50 just building kernels and llvm. 04:19:30 last chance,opencl context, building kernel. 04:28:25 geohot rapping,tom delonge. 04:37:35 release documention,legacy,amd gpu psp. 04:49:15 alive,clinfo,pyopencl,kernel build. 04:57:18 crash again,failed to add hardware queue to MES. 05:07:00 swaping the GPUs,no crash, bandwidth test. 05:23:55 it’s not the GPU crash reproduced on both. 05:26:10 going to NY,the benefit of the doubt. 05:30:10 appreciate open source,vbios 7900xtx,rocm smi reset gpu. 05:47:00 amd cards beloved by ML enthusiasts. 05:54:40 AMD MES, virtualisation enabled bios. 06:03:00 sending them back. 06:13:41 hsviewer keyboard. 06:16:51 vast.ai,buying 7900xtx stupid. 06:19:10 bios,boot delay,config.sys. 06:22:15 wait for bmc,clean driver for DL. 06:27:50 61 vs 34 FLOPS,AMD to mlperf,intel arc. 06:30:10 sign from AMD before R_DAY. 06:33:15 Jensen wins. . Official George Hotz communication channels:. – https://geohot.com. – https://twitter.com/realGeorgeHotz. – https://instagram.com/georgehotz. – https://tinygrad.org. – https://geohot.github.io/blog. – https://twitch.tv/georgehotz. – https://github.com/geohot. . We archive George Hotz and comma.ai videos for fun.. Follow for notifications: . – https://twitter.com/geohotarchive. . Thank you for reading and using the SHOW MORE button.. We hope you enjoy watching George’s videos as much as we do.. See you at the next video.