f_devd

13 ms·

211.
by f_devd 3y ago
<4x would be quite optimistic, at ~11x the tokens the amount of compute/memory required would be n^10 (even with the lower starting point of flash attention) so unless they are already have excessive margins it wouldn't make mu
212.
by f_devd 3y ago
I can actually answer this (as it is how I stumbled on to the repo), it's through a signal from the motherboard called Pwrbrk (Power Brake), Pin 30 on PCIe. It tells the PCIe device to maintain a low-power mode, in the case of Nvidia G
213.
Reverse engineering Dell iDRAC to get rid of GPU throttling(github.com)
249 points by f_devd 3y ago|90 comments
214.
by f_devd 3y ago
Normally that's correct except that the USB-C pinout is well designed to not put GND and VBUS near each other so it's pretty difficult to actually short anything. Still it's better to use something non-conductive.
215.
by f_devd 3y ago
You can clean USB-C the same way (I have before) but you need a needle or a very thin sim ejector
216.
by f_devd 3y ago
To be fair ML is (used to be?) pretty broad, so unless someone is actively keeping up with the sota in the high-data sequence modeling area it's quite possible to miss. I know ML teams which were entirely made up of OSML practicioners,
217.
by f_devd 3y ago
For the framework 13th gen it only required running a fairly recent kernel (>=6.1) for it to work, but that's definitely hindsight
218.
by f_devd 3y ago
I believe so, I remember an article on hibernation+luks being more difficult to set up
219.
by f_devd 3y ago
Seems to be just WiFi M.2 cards developed by MediaTek in collaboration of AMD[0], I guess the benefit would be additional verification/optimization. [0]: https://www.amd.com/en/press-releases/2021-11-18-amd-an
220.
by f_devd 3y ago
It might be a bit of tinkering but you can enable suspend-then-hibernate with logind/systemd which gives you a similar setup to Macbooks but slightly slower to wake from deep sleep as it needs to go through post/grub again. With t
221.
OpenROAD: Open IC Design Sythesis from Verilog(github.com)
3 points by f_devd 3y ago|0 comments
222.
by f_devd 3y ago
I had to reread it a few times but I think it's just poor phrasing, in the first sentence they are referring to the limited number of recharge cycles of most rechargeable batteries (which is generally much less applicable capacitors).
223.
by f_devd 3y ago
> Just let the runtime do it and throw an exception. Afaik there are no exceptions in rust only panics & `Result`s (and `Option` I suppose)
224.
by f_devd 3y ago
Well if you include internal you'd have a blob of neurons which on it's own wouldn't do anything. The point is not that it can "know time", it's the ability to handle time in a continuous matter with respect to
225.
An experimental deep learning framework for Nim using array programming(github.com)
5 points by f_devd 3y ago|0 comments
226.
by f_devd 3y ago
That's not quite the same as I meant, sense of time passing is not 'X seconds passed" as these are you are describing when time passed into the text modality (which means it's treated as text), a sense of time passing me
227.
by f_devd 3y ago
If you include a sense of time passing (i.e. continuous operation) and memory with respect to that time passing (persistence of events), maybe.
228.
by f_devd 3y ago
> deterministic lossy compression function This would only be true if it didn't mix or 'forget' any of the training data for each prompt, and only then for the exact same inputs of the training data. That is a lot of quali
229.
by f_devd 3y ago
Seems to be Linux 5.15, so it has a chance but SBC manufacturers don't have great track records on this so wouldn't hold my breath
230.
by f_devd 3y ago
Neither have I but it actually lines up pretty well depending on your read of GP. Large religions often use a proving-a-negative with a pascals wager to achieve an inarguable story regardless of content. And cult religions generally do this
231.
by f_devd 3y ago
For the people in favor of this proposal, please consider that the packaging story in python is already quite complicated without the option for local package directories. They admit the current system is not optimal (they recommend against
232.
by f_devd 3y ago
I did consider iCE40 (mainly due to the icestorm/yosys project) for our project at the time but they weren't available I believe, I will have to look into it again. Unfortunately a microcontroller was not really an option as the o
233.
by f_devd 3y ago
I wish it was possible to do this without the crowd sourcing, I have a low-volume HDL project of just ~40 gates and 32 shift registers (+4*32 gates) which discrete is quite bulky, FPGAs are overkill and CPLD works but are often too small or
234.
by f_devd 3y ago
LoRA (Low-Rank Adapter) is way to customize/finetune the LLM to a new datasets without needing to retrain the entire network (which makes it better (and in theory easier to do). It doesn't not change the speed significantly afaik
235.
by f_devd 4y ago
> If you trained GPT specifically on arithmetic tasks Sure but you'd have a lot of overlapping tokens with BPE, which doesn't help with convergence. GP is claiming it's specifically worse at arithmetic because of BPE which
236.
by f_devd 4y ago
Important here: it's breaking it through side-channel attacks, which means the algorithm itself might not be vulnerable but the implementation is (given sufficient information on things like timings)
237.
by f_devd 4y ago
To add to this, it's likely to 'originate' from (and is supported by) Thorn[0], which was discussed the previous post of mullvad's campaign. [0]: https://www.thorn.org/about-our-fight-against-sexual-explo
238.
by f_devd 4y ago
32EUR from JLCPCB for priority 7d and 26EUR for standard 7d (not sure what the difference is to be honest). The claimed price from AISLER is 57EUR. Although I have ordered from JLC before I don't have any receipts (and usually can affo
239.
by f_devd 4y ago
3x 2 layer 10cm*10cm: * JLCPCB Green 5 €1.86 total €0.37 each (Minimum order qty 5) * AISLER Green 3 €37.20 total €12.40 each Even with shipping & customs (20EUR extra only for JLC), AISLER is still 76% (1.76x) more expensive. Now I do
240.
by f_devd 4y ago
> They're also pretty cheap Maybe I'm making different types of PCBs but they are consistently the most expensive compared to all other fabricators [0] even with shipping cost savings. [0]: https://pcbshopper.com&#x2

More ›