About the corecollective.dev news, some context:

maybe I'll never be actually happy...

tolerance for a game of telephone with "AI"

Came across one unholy hack in QEMU today... https://gitlab.com/qemu-project/qemu/-/blob/master/hw/i386/vapic.c The APIC TPR register is accessed very often by Windows XP/Server 2003 and the way to make that bearable for virtualization pre-FlexPriority is to... patch the guest at runtime...

> When FEAT_MOPS is implemented, for a synchronous Data Abort on a Memory Copy and Memory Set instruction, ISV is 0. Ugh

oh my, in Jetson documentation.... > The HWVNs for MB2 and later components are stored in MB1-BCT, instead of in burnable fuses, so they cannot be ratchet-updated at runtime. To complete a ratchet-update, the boot software has to patch the binary for MB1-BCT, and because the software is […]

After seeing a non-sensical AI submission somewhere... Please don't do things like this: > Signed-off-by: GitHub Copilot <[email protected]> A bot cannot certify a DCO, so just don't

RE: https://mastodon.social/@fesshole/116075082915120087 oh no

DGX Spark power is down 18W in the OS updates this month when the ConnectX-7 is unused. A corresponding firmware update is also neccessary. For the Ascent GX10 the corresponding firmware update is in lvfs-testing so I also had to do: sudo fwupdmgr enable-remote lvfs-testing

x86 makes me miss having MMIO vmexits being decodable solely through the syndrome register instead of having to decode the instruction (which is generally the case for arm64)...

200b capex for the year, stock down 15% in a day and layoffs

LLM-generated code and the LKML...

Due to the huge compute needs of AI, it became overbearing in this cloud computing industry with lots of focus on it. And that seems to be at odds with what's the right thing to do for other workloads...

nvidia tegra t238

what the https://github.com/qualcomm/llama.cpp/blob/eb275c7f03f5411d88c6f5c56ea8926b25763199/docs/backend/snapdragon/windows.md#enable-npu-driver-test-signatures Needing signing is... not acceptable, Qualcomm

Oh wow a SYCL backend for macOS is well under way: https://github.com/AdaptiveCpp/AdaptiveCpp/issues/864#issuecomment-3829274089

hypervisor: an OS kernel scheduling particularly heavyweight processes with more state

JetPack/L4T Don't think I need to say more

The shutdowns will continue until Congress improves

There's no performance number less truthful than AI accelerator TOPS numbers

both the MacBook Pro and the Asus Ascent GX10/GB10 system here have coil whine under big load :-/

Alright NVIDIA Vera has screwed up PCIe with unaligned accesses to PCIe BARs being broken: https://bugs.launchpad.net/ubuntu/+source/linux-nvidia/+bug/2138952 and https://github.com/NVIDIA/NV-Kernels/commit/688e754768522e3d6993a18b20a810769bdf642a Normal-NC demoted to Device-nGnRE is a bit […]

Oddball food item (pic taken yesterday)

Not for human consumption

DGX Spark was a bold way to pre-launch a laptop chip before the software is fully ready

btw NVIDIA N1/N1X laptops will boot Linux ISOs with ACPI out of the box and the regular NVIDIA GPU driver package will work too

https://developer.nvidia.com/vulkan-driver New NVIDIA Vulkan dev driver, supporting Maxwell and later With https://docs.vulkan.org/features/latest/features/proposals/VK_EXT_descriptor_heap.html Will be interesting to see the vkd3d-proton commits, this should finally address D3D12 perf layered […]

nvhpc compilers with nvpl

the Realtek Linux upstream driver for their 10GbE NICs is really bad :/ and the downstream driver works as expected but is a completely separate codebase...

GB10 passes the largely obsolete "functional big endian" test (in a VM)

Customary AMI warning on the DGX Spark when enabling the TPM Also looks like the AMI setup was rendered at 4K lmao

So after playing around with an NVIDIA GB10/Spark system a bit: - yes this actually boots standard OSes ootb, just put Fedora to a USB stick and boot - ships with 3 out of tree drivers out of the box Realtek NIC uses a downstream driver instead of the upstream one An NV NIC power management […]

Apple's Hypervisor.framework isn't that buggy, but just narrowed down what I think is a bug in there that I've took some time to care enough about :D

Today, Windows Server 2008 and 2008 R2 (based on Windows Vista and 7) finally, truly reach end of support. Windows Vista and Server 2008 is the longest supported version of Windows, at 19 years. Behind it is Windows XP at 17 years, and Windows for Workgroups 3.11 at 16 years […]

Ampere's Ampere1C is an ARMv9.2 core with SVE2. *But* it doesn't have FEAT_SVE_BitPerm, like Qualcomm's 3rd-gen Oryon https://github.com/llvm/llvm-project/pull/175442/files

Openly available image generation models and preventing doctored photos are quite contradictory - as you can just remove any guardrails if it runs on your machine. And the cat is out of the bag since quite a while now... Mandating to use centrally managed services isn't that effective either […]

btw ignore any Jetson Thor performance numbers with llama.cpp llama.cpp doesn't support tcgen05 usage so the performance numbers are _way_ off...

DDoS by LLM-generated pull requests

tbh I don't know what I'm doing... I feel unhappy, have nothing tying me to where I am, and don't really have a purpose either

exceeding your social interaction quota for the day despite having almost none

may you be given the confidence of an average Hacker News poster oh wait is that a blessing or a curse

:-(

> Through this collaboration, automakers can gain dramatically simplified Android updates, streamlined integration, lower engineering costs, and faster time-to-market, supported by a robust 10-year plan for critical software updates only 10 years? That's pretty small for cars, Qualcomm

Pile of C? That's ok Trying to use PHP? That's banned (yes that's corporate policy at a certain place)

half precision (FP4) single precision (FP8) double precision (FP16) /s

:/

unified data plane and console/IPMI network because BMC security such wow

May you live in interesting times. Happy new year!