nrp

9 ms·

91.
by nrp 2y ago
You’ll be able to run distills of DeepSeek R1. For the full model, you can network together multiple and use exo or llama.cpp RPC. You can do image generation as well! AMD has some demos of this, and they are maturing PyTorch support to b
92.
by nrp 2y ago
Definitely! ROCm is getting really solid for inference. LM Studio (and therefore the underlying llama.cpp) work out of the box already, and we see AMD pushing forward on PyTorch and other areas rapidly.
93.
by nrp 2y ago
Amazing to hear that!
94.
by nrp 2y ago
Thanks, glad to hear that!
95.
by nrp 2y ago
We've been moving to fTPM (AMD's Pluton TPU) in our last couple of generations, including on Framework Desktop. This is something where if we see strong demand to go back to dTPM that we can look at for future generations.
96.
by nrp 2y ago
I’m happy to answer questions folks have on the Framework Desktop (though probably not until later today).
97.
by nrp 2y ago
Just sent you an email.
98.
by nrp 2y ago
Super impressive, and awesome to see that you were able to use Framework Laptop hinges. Let me know if you need more. We have a ton of remaining 3.3kg ones!
99.
by nrp 2y ago
Framework Laptop 16 without a dGPU is 2.1kg.
100.
by nrp 2y ago
We’ve taken to recommending Rufus in our setup guides rather than the Windows installation media tool due to the lack of recent Wi-Fi drivers in vanilla Windows.
101.
by nrp 2y ago
We do for a wide variety of products and IP: https://www.trade.gov/us-export-controls See also the US sanctions on SMIC.
102.
by nrp 2y ago
In this case it’s because Dylan Patel of Semianalysis interviews Lisa Su regularly and presumably has a direct line to her, and because Lisa and the rest of AMD leadership are absolutely reading the article. It’s unclear if Pat would have
103.
by nrp 2y ago
> Give AMD Engineers more compute and engineering resources to fix and improve the AMD ecosystem, they have very few internal gpu boxes relative to what Nvidia provides to their engineers. This is real. We’ve found ourselves having to g
104.
by nrp 2y ago
This is designed to fit the Expansion Bay Shell, but not the Graphics Module. To fit the dGPU, we had to make the mounting scheme different within the Graphics Module.
105.
by nrp 2y ago
We expect most Framework Laptop 16 users won't need more than the two M.2 storage slots that are on the Mainboard itself, so we don't want to end up with a wasted module.
106.
by nrp 2y ago
Our engineering team did the engineering!
107.
by nrp 2y ago
We have that on the most recent generation of Framework Laptop. When the hardware privacy switch is engaged, the image sensor is electrically powered off and the camera controller feeds a dummy frame with an illustration of the switch.
108.
by nrp 2y ago
Note that this is a business-focused pre-release from DeepComputing, not the pricing for general developers at release of the Mainboard.
109.
by nrp 2y ago
We have gotten inbound interest from other Mainboard makers too, at least one of which is pretty far along on a design.
110.
by nrp 2y ago
Cameras use isochronous mode, which maxes out at less than half the theoretical max USB 2.0 bandwidth, and they use MJPEG compression on each frame rather than something like H.265 that can compress much more by leveraging P/B-frames.
111.
by nrp 2y ago
USB cameras are interfaced over USB 2.0 because USB 3.0 is extremely difficult to make work from an EMI perspective running through a laptop’s hinges and lid. USB 2.0 means heavy compression to get 1080p60, and on pretty mediocre ISPs in t
112.
by nrp 2y ago
That’s one of the most interesting parts of this document. Many people will read it and think “I would never work at a place like that,” and many others would think “that’s exactly the environment I want to work in!” More startups should b
113.
by nrp 2y ago
This viewpoint resonates for me. OpenSCAD is a useful tool for hacking together 3D parts quickly, but an excellent tool for exercising the shape rotating parts of your brain.
114.
by nrp 2y ago
How are you finding 2b/3b quantized llama 405B? Is it behaving better than 8b or 16b llama 70B?
115.
by nrp 2y ago
As noted in the article, Architecture was indeed previously an Olympic category: https://en.m.wikipedia.org/wiki/List_of_Olympic_medalists_in...
116.
by nrp 2y ago
We ship a large quantity of RP2040’s in real products, and agreed that the RP2350 looks great too! Part of the reason we went with RP2040 was the design philosophy, but a lot of it was just easy availability coming out of the chip crunch.
117.
by nrp 2y ago
The Telo Truck is more directly a modern take on getting the most truck into the smallest footprint: https://telotrucks.com/
118.
by nrp 2y ago
For context on the cooling system, it is absolutely overkill for the SoC, but to keep the project simple, DeepComputing is re-using the same heatsink and fan from our 11th Gen Intel Core Framework Laptop 13. This means that it should be ab
119.
by nrp 2y ago
We have a set of documentation on GitHub: https://github.com/FrameworkComputer/Framework-Laptop-13 We're using this program with DeepComputing to help develop our Mainboard documentation further as well.
120.
by nrp 2y ago
The Mainboard itself is going to be a lot more accessible than you (apparently) think! We'll have more to share on pricing as we progress with DeepComputing on the program.

More ›