cpldcpu

11 ms·

121.
by cpldcpu 2y ago
>Thus, in this study, we take the state-of-the-art ChatGPT (the default version of GPT-3.5), the recent popular product, as the representative of LLMs for evaluation. Are they really publishing a paper based on GPT3.5 in July 2024? I am
122.
by cpldcpu 2y ago
Extrapolating from the 737max to the 787 is ridiculous. Still prefer A380 and A350 though.
123.
by cpldcpu 2y ago
There is still impressive progress in large language models themselves every week. (Just to mention some of the past month: Claude-3.5-Sonnet, Gemma2, Nemetron and many more) What is, in general, strange is that noone has really figured out
124.
by cpldcpu 2y ago
>Building an FPGA board This seems to be a bit odd? This is already a more tedious hardware project to debug, but when it is about learning the basics, building a much simpler circuit would provide more insight. It's also a bit ques
125.
by cpldcpu 2y ago
Anyways, I think ARM is on to something with their Edge AI IP, which enables inference in microcontrollers. So far, there is no Nvidia of microcontroller ML. If ARM managed to build a moat with proper software, they can go very far...
126.
by cpldcpu 2y ago
whoa, that's awesome.
127.
by cpldcpu 2y ago
Yeah, you can find them in all kinds of low-cost remote controls: https://cdn.hackaday.io/images/9838991700773145118.file-1700...
128.
by cpldcpu 2y ago
Use a neural network to detect numbers? (Well, this runs on a slightly more capable MCU, though) https://cpldcpu.wordpress.com/2024/05/02/machine-learning-mn...
129.
by cpldcpu 2y ago
>the price rises for the next two years were pre-announced at davos in january 02019 by eric luo, president of gcl, a top-ten chinese solar-panel company; this is a sort of announcement that cannot happen in a competitive market https:&
130.
Interactive Best Research Solar-Cell Efficiency Chart(nrel.gov)
1 points by cpldcpu 2y ago|0 comments
131.
34.6% efficient perovskite-silicon tandem solar cell demonstrated(pv-magazine.com)
2 points by cpldcpu 2y ago|1 comments
132.
by cpldcpu 2y ago
>now they cost 8¢ a watt, down by half since last year's prices, prices which had been held steady for several years only by a price-fixing cartel. Sorry, this is most likely not true and an unbased claim. >this is precisely what
133.
by cpldcpu 2y ago
Most likely the fab sells it at a loss, given the current situation in the industry: https://www.reuters.com/business/energy/china-solar-industry...
134.
by cpldcpu 2y ago
10e6 cycles is nothing. A CPU at 10MHz writing at the same memory location would create that stress within 100ms. Note sure if this is a misinformed article or some information is missing.
135.
by cpldcpu 2y ago
that was 20 years ago
136.
by cpldcpu 2y ago
It's just continued pretraining to "heal" the damage caused by switching the activation functions and enforcing sparsity. Apparently they managed to recover original performance on standardized tests after continuing pretrain
137.
by cpldcpu 2y ago
>This reaches demoscene levels of crazy/impressive! The exp/log trick to multiply with addition does indeed look very familiar. I know that a number of demos used it in the 90ies to simplify matrix multiplications for 3d graphi
138.
by cpldcpu 2y ago
Good point, I went right to the bitnet code. I will correct my original post.
139.
by cpldcpu 2y ago
The quantization approach is basically identical to the 1.58bit LLM paper: https://arxiv.org/abs/2402.17764 The main addition of the new paper seems to be the implementation of optimized and fused kernels using triton,
140.
Neural Network inference on a "3-cent" 8 bit Microcontroller(cpldcpu.com)
4 points by cpldcpu 2y ago|0 comments
141.
by cpldcpu 2y ago
>Several people stopped to help corral the animals, including a rodeo clown and horse trainers. ok
142.
by cpldcpu 2y ago
Maybe something simpler, like a haar wavelet, would also work? Or DFT using Görtzel?
143.
by cpldcpu 2y ago
On the CH32V003, a load should be two cycles if the code is executed from SRAM, there are additional wait states for load from flash. The V2A does only cache a single 32 bit instruction word, so there is basically no cache. This publication
144.
by cpldcpu 2y ago
https://www.statista.com/statistics/677096/vr-headsets-world... This seems to be inconsistent with these numbers
145.
by cpldcpu 2y ago
>It was 23 billion dollar market in 2023. How is that possible? Assuming an ASP of $333 USD (which is too high), this would be >70mio sold headsets?
146.
by cpldcpu 2y ago
The latter one. The network is trained in full precision (this is required for the gradient calculation), but the weights are nudged towards the quantized values.
147.
by cpldcpu 2y ago
The different is in using quantization aware training, where the quantization of the weights is already simulated during the training. This helps to restructure the network in a way where it can optimally store information in the allotted n
148.
by cpldcpu 2y ago
Indeed using a mouse sensor for data input would be quite interesting. Mayber another option would just be a row of phototransistors.
149.
by cpldcpu 2y ago
Great project! I used MNIST because it is easy to work with as a dataset. Audio classification would be quite interesting as a follow up, but I assume one would need some kind of transform to deal with the data in an easier way.
150.
by cpldcpu 5y ago
nah, it will easily fit. not sure about routing ressources though. === CPU8BIT2 === Number of wires: 319 Number of wire bits: 3029 Number of public wires: 38 Number of public wire bits: 20

More ›