The spy slop is fed to the masses to make the otherwise mundane export control issue more intriguing to them. If you watch cable news, it's non stop slop about how it's sensitive IP or his closeness to Trump, AI strategy, or other labs that is jeopardized. GPU architecture doesn't make for good pillow talk.
Democrats have gained power and are about to engage in a protracted colonoscopy whilst asking questions that amount to "Are you a commie?". For a refresher, review the TikTok CEO's hearing.
Yes, my point being that Democrats have gained the ability to hold the hearings as they won in midterms. If they hadn't, Republicans were not going to initiate this. All of this is 1-2 years old; Republicans could have done this at any time.
It doesn't have much to do with this CVE, but if you're trying to sell to the US government, it will often be a requirement, which means that a company making money from those sales will not easily be able to move to a distro without that certification.
It's about using a certified cryptography implementation. Presumably RHEL has such a certification, while Debian has a compliant but uncertified implementation.
This isn't relevant to the CVE, but counters the claim that Debian is always a suitable choice of distribution.
Yes, there is no point in grading homework or essays. This has been largely true in India and China anyway, and the US will have to move to similar model of standardized tests and in person exams. Universities are anyway proxies for IQ filtering. It is perhaps impolite to say so explicitly in the west, but people will have to confront that.
Maybe for "soft" subjects like History and Philosophy, where almost nobody who does those courses actually becomes a historian or a philosopher... But there are plenty of subjects where you actually do learn things that are useful beyond university. Engineering, law, medicine, chemistry, etc.
It's describing a different type of parallelism - pipeline parallelism. The network is laid out spatially on the wafer in CB's case. So weights are stationary and activations flow. A GPU is going to process the network one layer at a time. GPUs also use other types of parallelism (tensor and expert mainly in inference). Pipeline parallelism doesn't make much sense on a GPU for inference.
I think the point is they need several wafers to fit the model. Data parallelism between requires too much bandwidth (they only have so much edge real estate for io) so you have to do pipeline parallelism with say one layer per wafer. No one does it on gpu unless you can’t fit the model in one fast network domain (which is 72 GPUs nowadays for Nvidia). There is a Dwarkesh episode from like two months ago that explains this reasonably accurately.
The reason 2D grids work well is because you can split tensor axes along whatever dim you want across PEs (batch, sequence, hidden, etc.). And your per PE code is relatively simple. It's of the type : get data from neighbor, compute my subtensor, send to neighbor. And you can implement broadcasts and reductions along any axis.
My guess is it is more like the opposite. They distribute spares evenly along every array in every dimension, so that there are always the same number of selected and unselected cores in each core-set.
With flexible addressing and bus-linking, along any set-line. So configuration happens the same way, with the same efficiency, regardless of whether any of the unselected elements are failed or not. As long as failures don't exceed the unselected-elements per element-array-set number.
And that number of unselected cores per element-array-set, is chosen to put wafer yields where they want them.
Trading a generous fixed wafer under-utilization for high yields, uniform configuration and processing efficiency, despite variable element failures.
No, I think this is a valid threat model and a good rebuttal. The evil maid steals the original laptop, replaces it with a second laptop with a look-alike PIN entry screen without an attestation. You enter your PIN, it is transmitted to the evil maid with the first laptop, and your data is unlocked. You would notice, but your data would be exposed.
Now, to be clear, I don't think this negates TPM-sealed disk encryption or TPM-and-PIN sealed disk encryption; they're still meaningfully better than the alternatives (with a passphrase alone, your data is already compromised regardless in this invented scenario, and all scenarios are equally variable to the swiss-cheesiness of UEFI firmware unless it is replaced). But it is a valid model.
What does it matter if it doesn't unlock the disk? It's a different device with a logger. The purpose of the maid is to be evil and steal your pin. The maid has your real device. The only way to prevent this is by you authenticating the device before you enter anything. Look at the heads project (https://github.com/linuxboot/heads).
What problem is this solving though? The post says they're doing this to avoid typing passphrases. Why would you want the private key hidden in the tpm chip of that device? How will you decrypt if the laptop goes bust or if you move the drive elsewhere?
The TPM generates a strong, non-exportable key in hardware that is far less susceptible to brute-force attacks.
As opposed to accidentally creating your own key of "4", for instance.
A PIN can be used to protect the TPM from unlocking. Even if you set a PIN of "1234" — it would still be far more secure if the disk and laptop were ever to become separated, because the TPM will have generated a strong key anyway.
The TPM is basically a smart card soldered to the motherboard.
> How will you decrypt if the laptop goes bust or if you move the drive elsewhere
Windows and Apple both generate a recovery key as part of the enrollment process which you can store offline.
When you format LUKS you can add multiple decryption keys in case the TPM is hosed for some reason like the mobo failing. Can even use a Yubikey instead of a TPM. Or both. Or none. It's really up to you.
That's a minor convenience to not type passphrases and a fairly high risk of you losing the backup key. Not to mention, the entirely separate problem of storing the backup key securely. It doesn't pass the xkcd "correct horse battery staple" test for me.
Fedora uses LUKS where your backup key can just be a regular passphrase. A key derivation function is used internally to help with mitigating brute force attacks.
No, yubikey in the fido/webauthn form protects against phishing/TOTP snooping and replay over the web. TPMs for luks don't. See the other thread about evil maid.
Why is everybody (in the government) in the EU so illiterate? Do they not understand that their economies are dead and football streaming isn't going to save them.
reply