> well that is fortunate

I think that was one of the arguments of the Anthropic principle [1], that there doesn't appear to be any reason why there are 3 spatial dimensions and 1 time dimension, or why the fundamental constants are what they are - but if they weren't then there wouldn't be anyone to exist to say "well that is fortunate".

[1] https://en.wikipedia.org/wiki/Anthropic_principle#Dimensions...

yes, exactly. It didn't have to be this way, but it had to be this way to observe it. Survivorship bias.

You're right that the ECC validation is very robust, but that only validates one small part - that the drive is reading what it has previously written, not that the data was correct when it came in to the drive, correctly handled by the firmware, or even written in the correct place (LBA) on the drive.

There's been times when some features of entire models of drives have been disabled in the Linux kernel because of buggy firmware that silently writes bad data (with correct ECC), so reading it back is successful from both the drive's and the OS's block driver views.

I was hit by this myself with the queued TRIM command firmware bug that affected all Samsung EVO 840 SSDs (Linux kernel commit 9a9324d3969678d44b330e1230ad2c8ae67acf81 if you want to look into the history) - the drive didn't report any errors, but ZFS kept reporting corruption, and kept on fixing it in the background.

Please can you enlighten the ignorant that are also gamers? :)

downloading a game on steam can fully utilize gigabit speeds, and modern games can be 100+gb

I play games but I suppose I don't download a 100GB game frequently enough for it to matter to myself.

Is that 80-100GB of unique data multiple times per day? Is it encrypted with a different key every time? Sounds very interesting.

Yes. Our build system produces complete packages of the game, and I'm a platform engineer - so within one day I'll often download a PC package, one for PS5, Xbox, Switch, Android and iOS, depending on what kind of thing I'm working on. So yes all of them are completely unique. Or I'm looking at builds made for different backends. Or I pushed a preflight for overnight cooking/building, it takes about 8 hours - so I need to download the output to see if it works across different systems.

Not OP, but modern asymmetric cryptography on the Internet negotiates a new encryption key on every connection.

True, but that is transport (TLS). I was asking about the underlying data.

I am curious why you are interested in that.

Because I would have expected for game development that it would be rare for majority of the large assets (art, textures, 3d meshes, video, sound, music) to change multiple times per day, so - to me - sounds like would perfectly suit a differential/incremental sync protocol, rather than treating everything as a large packed blob. At least that was the case for the game development I did.

And generally - you are correct. But builds made for specific platforms have all of their assets "cooked" for that platform - so textures shipping on PS5 are not in the same format as textures shipping on the Switch, even though they were both made from the same original reference file.

One side-effect, is that the separate .mmproj file (Multi-Modal Projection encoder) is no longer needed, when using the model with llama.cpp etc.

It's not? There's an mmproj in the GGUFs released by ggml-org: https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF/tree/mai...

From the visual guide, there's still the 35M parameter embedder, then the linear projector, for vision, and the linear projector for audio, so it does have some parameters used for the multimodal input to project it into the LLM latent space: https://newsletter.maartengrootendorst.com/p/a-visual-guide-...

And the Unsloth quants, which are missing this, don't support multimodal input. (edit: actually, I may have just needed to update my llama.cpp, will check with an updated llama.cpp soon)

I'm downloading the ggml-org GGUFs now, I tried Unsloth but got some weird problems, double checking with the bf16 model to see if the issue was just the quant.

Ah, Unsloth has uploaded mmproj now as well.

But do I have the option to run it 'text only'?

+1 this. Example: Using Mistral TTS voice cloning appears to be not possible via the "providers" pass-through object in the OpenRouter API because some parameters are always forwarded which conflict with the provider's parameters.

The latest Raspberry Pi 5 has one 32-bit channel (2x 16-bit subchannels) of LPDDR4X-4267 SDRAM giving 17.1GB/s of bandwidth, 52x less than this GPU. Never mind lacking the CUDA and Tensor cores, so the FP16 performance is 102x less (307 GFLOPS vs 31.4 TFLOPS). So for £200, there's absolutely no comparison for this specific use-case.

Yeah thats what I'm saying. How is it so cheap????

V100 GPUs are e-waste.

The AMD MI250X GPUs are also interesting - 128GB of HBM2E at 3TB/s, sometimes you see them second-hand for under $1k, the catch obviously is that it needs an OAM socket. Never seen an easy way to hook them up to a regular mainboard.

An additional complication is that MI250Xes are two GPUs in one package, so you need to connect the first and last x16 SERDES groups to the host, otherwise you'll only see one GPU (or it won't work at all, idk).

Also, the cheap HPE pulls on eBay need some proprietary HPE magic to work, and I have yet to see anyone figure that out.

This person has built a converter for the OAM socket, but it is only confirmed working with NVIDIA cards at the moment (https://www.reddit.com/r/NVIDIA_SXM2PCIE/comments/1d076cn/oa...)

It fits an MI250X, and the system sees it, but the drivers don't work. They tested an HPE MI250X. There's a rumor on the thread that there are two kinds of MI250X: Ones from HPEs and everyone else's. The HPEs require a special firmware, the normal ones do not. However, the majority of the MI250Xs on the secondhand market are HPE so caveat emptor.

Ahh luckily this OAM socket will prevent me from spending money.

These are interesting, and offer beefy through put. No point in adapting to a PCI lane thought, stuck behind the slot-bus bottleneck.

> canonicality matters — for signatures, content-addressing, or any kind of “two implementations must agree on the bytes” property

If you don't do this properly, you end up with things like: - SAML XSW attack due to XML signature wrapping - ASN.1 BER/DER signature forgery - Bitcoin transaction malleability attacks

The problem is AVX-512 was disabled in later Intel Alder Lake CPUs, and later generation Intel desktop CPUs, so very few Intel desktop CPUs have AVX-512 now. Ironic that AMD has better support/performance for an ISA extension that Intel invented.