Hey guys! Just for background, I own a FBO & tuned 2002 RSX Type S on K-Pro v4. I drive spiritedly and like to hit VTEC. I also have a pop tune that kicks in after 3k RPM. Car has about 145k miles. I do daily my car too. Still new to K20s. I did a hard pull the other day and noticed the oil light flash for a quick second. I did do an oil change around 1k miles ago with 5 quarts of high mileage, full synthetic mobil1 5W-30. Checked my dipstick and she was reading just a tad bit past the first dot. Considering the modifications, mileage, and my driving, does this seem normal? 

Thank you!

Disclaimer: I´m not a doctor, biologist or chemist and this is just my non-professional experience as someone with HI. Everything I write is based on my individual metabolism and pre-existing conditions.
However, I wanted to make people aware of possible side-effects of supplementation with HI.
I started to take supplements, especially high doses of b-vitamins starting 2019 and an even higher dose of b-vitamins starting 2021. Mostly because of a genetically predisposed type of pyrroluria. What I didn´t realize at that time was, that all of these contained methyl-donors and had the potential of interfering with my methylation and different enzymatic and chemical pathways.
My plasma histamine was tested and it was normal, meaning, there should have been no problem with my methylation. My HNMT-enzyme function was tested in a lab and it didn´t run properly, (DAO was good) but I didn´t connect the dots. I have SIBO and MCAS, which might contribute to this situation. I did run blood tests and my homocysteine was 10.3 in 2015 and 4.1 in 2020. Because I supplement, I do plasma blood tests every year and none of these showed a level in any vitamin/mineral, that was too high or too low.
Despite all of this, my HI got worse over the years.
Doing a lot of reading all over the internet, I found three points, in which I could have messed up with supplementation:
Number one is messing up my homocysteine and hurting my SAMe, HNMT-enzyme and my glutathione. This presentation from Dr. Masterjohn shows the mechanisms, at 20:25 there is a good picture of methyl donors: https://www.youtube.com/watch?v=IRQ4-lAeYJI And this post from Dr. Lynch made me aware of the dangers of low homocysteine: https://www.drbenlynch.com/low-homocysteine/ . Basically, I have not enough homocysteine as a building block for either SAMe, which is vital to the HNMT-enyzme or glutathione, which detoxes the body. So I kept taking these B12 supplements and all for building SAMe and at the same time may have reduced the essential building block for it. I never got measured either for SAMe or glutathione, so I don´t know.
Number two is making my body build methyl-histamine from all of the methyl donors I was taking. Here is a video from Dr. Anderson, where he explains it at 04.50, but the whole video is gold: https://www.youtube.com/watch?v=_q6_2wDX1i8 . Basically, I might have made my body transform normal histamine to the much-harder-to-remove methyl histamine and overloaded my pathway of HMNT and MAO. I tried the vitamin B5 , which he mentioned in the video, but even at 250 mg (he recommends 500 mg) made me feel weird. And if I keep taking methyl donors, it makes no sense to take B5 to remove the outcome of this.
Number three is maybe taking B6 enabled my body to transform more histidine to histamine: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7527268/. I read in my book for HI, that sometimes, when you take P5P (active B 6) in higher doses, the body takes it and amps up the histamine production. Especially, if SAMe and zinc/copper is not optimal. Some people have to be more careful with B 6 supplementation: https://www.youtube.com/watch?v=uSHMG8zVZVo
I will have to run further lab testing and cut back on supplementation. Basically, it s testing and trial-and-error as most doctors know nothing about these interactions or proper supplementation. I had one doctor congratulate me on lowering my homocysteine so much.
What I wanted people to take away from this long post is that if you supplement, you don´t have control of how your body metabolizes and uses the vitamins/minerals.
The more I read about all of the pathways and chemical reactions you can mess up with supplements, the less I take and the more careful I am with supplements. I don´t demonize them and they can be useful, but I got more careful
Hope, this helps some people.

hello,my friends
capcut is a video editing software launched by Douyin. It is more suitable for users who have some editing skills or have some personalized video editing needs. During the use of Capcut, it is necessary to manually add stickers, special effects, and text at a certain position, which is suitable for users who have free time to process a small amount of high-quality materials.
Difference between GhostCut and Veed in Video Translation
GhostCut is an intelligent editing software that does not require too many user operations. As long as the video is extracted through a link or uploaded from local materials, and the corresponding processing items are checked, the system will automatically recognize and process the video, such as completing the erasure of text in the video or automatically translating english videos into other languages overseas, and aligning the picture to generate new subtitles and dubbing. It greatly frees up your hands and can meet the needs of users who are not proficient in editing or have batch and fast editing needs, quickly and efficiently producing some content.
Regardless of the type of video file you use (MP4, MOV, WEBM, AVI, etc.) or the purpose of use (editing webinars, Zoom recordings, adding broadcast video sound waves, etc.), you can work on projects together here. You can edit videos without limits (including audio editing and sound effects), and even use it on your iPhone and iOS system. Use GhostCut to create wonderful marketing videos suitable for various social media platforms, intro videos (welcome to use our intro maker), add logos, and always design with your brand as the core.

Why we use GhostCut for De-duplicate Video ?

GhostCut is an intelligent de-duplicate product that uses AI technology and video processing technology to process the md5, fingerprint information, visual information, audio information, subtitle information, and more of the video, helping users to perform secondary video creation better and faster.
Compared to using editing software like Capcut or PR templates for video De-duplication, its characteristics are: fully automated, multiple templates, and support for combining subtitles and voice, making it a more advanced secondary creation tool.
Of course, you can also use it with other tools. First, use GhostCut to complete the basic operations, and then use Capcut and PR to do more advanced editing.

If you’re the type of person who likes to wake up super early, like 4 or 5 AM, to study and feel like a boss, then this article is for you.
but sometimes you just can’t focus and end up hitting snooze, don’t worry.
This is just for you early birds who want to start the day off right.
So, there are a bunch of ways to study, and I have experimented with different types of study methods but there’s this one that only works for people who like to wake up early.
It’s pretty cool because it helps you remember stuff better and makes it easier to recall everything you learned.

Getting a sense of how much fuel you have to grind

If you’re a morning person like me, you’ll likely feel more energized as the day goes on, peaking in the afternoon.
But then your energy will drop and hit rock bottom at night.
Of course, everyone’s energy levels are different, so it’s important to pay attention to your body and adjust your study schedule as needed.
I will tell you about what to do at different times of the day and what type of subjects to study.

What is the Right time to wake up for an early bird?

I would say it depends but the most common times are 4-5 am.
so make a goal and set the alarm already. And listen, When I make a goal of waking up at say 4 am, then I ensure that I don’t wake up at exactly 4 am. because I would panic after that clock hits 4 that I am wasting every second after 4 am.
it is just that I get paranoid haha. I feel like it is my boundary and any time wasted after that would make me super guilty. but don’t be like me. I am trying to change this haha.
cutting the bs So to ensure this I set the alarm at 330 or 345 to ensure that I relax before actually sitting at my desk at 4.
It’s a good idea to start your day at least 20-30 minutes before you actually want to wake up. So if I want to wake up at 7:00 AM, I set my alarm for 6:30 or 6:45 AM. This gives me some extra time in case I want to stay in bed a bit longer.
this period also helps to make the schedule for that day and I write a list of all the tasks that I have to finish that day.
If you’re going to school, waking up at 5:00 AM is perfect because most schools start at 7:00 to 8:00 AM, which gives you two hours to get ready.

Okay, but how do I do this without feeling drowsy?

Just set your alarm for 30 minutes earlier than usual, like 8:30 instead of 9:00. This will give you some extra time to relax and remove that drowsy feeling.
Keep it up for a whole week before trying to wake up even earlier. and slowly practice this every day and you will master this. and one day you won’t even need an alarm for this.

I woke up, now what do I do? How do I start my day productively?

When we wake up, our body needs to rest for a bit, as it was in an unconscious state for 7-8 hours. Personally, I like to take advantage of this period by getting stuff done, like making a to-do list as I said before.
But if you’re someone who gets up super early, your list might need to have two parts: one for stuff that needs to happen right away, and another for things that can wait a bit.

4:45-5:00 am – warm-up time/ getting ready for the day

During this time of getting ready to start my day, I like doing my homework as a warm by seeing my subjects instead of starting directly with studying. but only those homework which is easy to finish and can be done faster.
It could be just one worksheet or a couple of pages of Math problems. Plus, I’ve started spending an hour every morning working on my Math skills. It’s especially helpful for people who find Math hard because practicing regularly can really help.
Doing this stuff every day makes me feel productive and good about myself, which is good for my overall satisfaction.
Another reason why I struggle with math at night is because I just don’t have the brain energy. Plus, it’s helpful to exercise your mind before things get really crazy.

01# Session – 8:00-11:00 am –

High concentration span/most important time

If I want to get some serious work done, I tackle it between 8 and 11:00 AM. It is the best time. and my fav too. I feel the most productive during this time.
That’s when your brain is firing on all cylinders and you can really concentrate. On weekends, it’s a good time to learn something new, and during the week, you can stay focused on your studies. Basically, this is the time of day when you can tackle anything that requires brainpower and problem-solving skills.
Are you aware of one of my articles on how to handle studying for multiple subjects? If you haven’t, no worries, read it whenever you have some free time.
Anyways, I just wanted to mention that there are three kinds of subjects:

• ones that require you to remember stuff/theoretical subjects,
• ones that require problem-solving (like physics and maths), and
• ones that require interpretation.

Personally, I like to tackle the first two types in the morning since they’re a bit tougher for me than the interpretation-based ones. you need to figure out which one is tougher for you, Don’t blindly follow my brain’s capacity. some of you are way wiser or slow learners like me. so it depends from person to person. but get the central idea right.
When I’m at my best, I like to break up my study time into two parts. I usually tackle one subject from 8 to 10 in the morning and then switch to another from 10 to noon. That way, I can stay super productive and focused.

Bonus tip 01

If you want to remember everything you learn, there are two things you should do every day.

• First, write down all the important keywords about what you studied in the morning.
• Second, watch a video that explains it, a great study hack in the evening. If you’re short on time, typing your notes is a good idea.

Session 02# – 12 pm noon

Around noon, I switch my attention to stuff that’s more about interpreting things.
It’s way less tiring than dealing with Math, Biology, or Physics. This is key because we usually start feeling more drained as the day goes on. I take advantage of this time to handle things I’ve been putting off on my to-do list.

Session 03# – 3 pm after school

My least favorite time would be this. personally speaking. It’s 3:00 PM and you’re probably either just getting home from school or you’ve been studying all day.
You’re probably feeling pretty tired by now, so it’s a good idea to take a quick power nap. Try to keep it between 10 and 20 minutes so you’ll feel refreshed when you wake up.
If you sleep longer than that, you’ll just end up feeling more tired. and the day is almost over.
​

Session 04# – 4 pm: last session of the day

When it hits 4:00 PM, it’s a go time for the final push of the day.
From 4:00 PM to 7:00 PM, I like to take a breather and either think about what I learned earlier or do easy tasks like writing down some notes.
If you went to school, this is a good chance to go over what you learned in class. Being purposeful with this time can really boost your efficiency and end the day on a high note.
If you want to remember something for a long time, just make sure you study it twice a day. It’s super important!

Bonus tip 02#

Remember when I told you to make a list of important words related to what you’re studying? grab that piece of paper.
If you wanna remember stuff better, you gotta go over everything you know about a topic and connect the dots. Mind maps are a good way to do that.
It’s important to actively recall the info to keep it in your brain, and recapping is a good way to do that. That way you’ll actually remember it for the next time you need it instead of forgetting it right away.

Bed routine

As the sun goes down, a lot of people feel tired and less motivated, which is called the trough period. To deal with it, you can try the two-minute strategy – just do quick tasks like answering emails or packing your stuff. That way, you won’t have to waste your morning time on them.
Oh, and it’s a good idea to rest up and get to bed early too. If you made it this far, let me know if you are an early bird or a night owl. if you’re a night owl then why are you here? I have already posted about you too! see you in the next post bye!
I wish you all the very best in your studies!! And don't stress too much ;)
If any problem regarding your academics, then feel free to mail me here or comment below!
MORE SECRETS HERE!

Says it all in title lmk thanks and appreciated

Hey everyone!
Welcome to the /asmr weekly Free-For-All Friday - your one stop shop for images, gifs, comics, memes, art, off-topic discussions, and general ASMR fun.
Traditionally we remove this type of content from the normal feed, but we're giving you this opportunity to have some fun for the weekend and lift that rule for this thread alone.
(Please note that we will still enforce all of our other rules, so please be respectful towards one another.)
If you'd like to chat in real time, we have a Discord server open for you to use. You can use the Discord chatrooms from here - it's a friendly place so come and say hi! :)

Overview
In electronics, AI (artificial intelligence) chips are dedicated processors used to run AI algorithms. The difference from normal chips (CPU/GPU for general purpose) lies in the emphasis on specificity and versatility. CPU, GPU, FPGA, and ASIC chips can all execute AI algorithms. And to a certain extent, GPU and FPGA can also be regarded as AI chips, but ASIC has more distinctive features, so it is also the most representative. So what they are and why they matter?
​
What is AI Chip?
The so-called AI chips generally refer to ASIC (application-specific integrated circuit) chips. Both traditional CPUs and GPUs can be used to execute AI algorithms, but their speed is slow and performance is not good enough, making them impractical for commercial use. For example, automatic driving needs to recognize road pedestrians, traffic lights and other conditions, but if the current CPU is to calculate, then the car may overturn into the river before calculation results, which is a slow performance. If you use a GPU, the speed is indeed much faster, but the power consumption is high, and the battery of the car may not be able to support normal use for a long time, and the GPU is not economical, that is, ordinary consumers cannot afford it. What’s more, GPUs often out of stock. In addition, the GPU is not an ASIC specially developed for AI, so the speed has not reached its limit, and there is still room for improvement. In fields like intelligent driving requires fast calculation; in mobile terminals, AI applications such as face recognition and voice recognition must have low power consumption. Obviously, GPUs cannot meet the requirements, so the development of ASIC has become inevitable.
​
Why It Matters?
So why do we need AI chips? For AI algorithms, CNN (convolutional neural network) is commonly used in image recognition and other fields; RNN (recurrent neural network) is mainly used in speech recognition, natural language processing and other fields, which are two different types of algorithms. However, they are essentially matrix or vector multiplication and addition, and then cooperate with division, exponent algorithms.
A mature AI algorithm, such as YOLO-V3, is a large number of calculations such as convolution, residual network, and full connection, which are essentially multiplication and addition. More over, if the specific input graphic size is determined, then the total number of multiplication and addition calculations is determined, like one trillion times, of course, the real situation is much larger than this. If you want to execute it quickly, one trillion additions and multiplications must be performed.
Let’s take another example, such as IBM’s POWER8, one of the most advanced superscalar CPUs for servers, 4GHz, SIMD, 128-bit, assuming that 16bits data is processed, that is 8 numbers, and a maximum of 8 multiplication and addition calculations can be performed in one cycle, performing up to 16 operations at a time. But this is theoretical, it is unlikely in fact.
At this time, the peak calculation times of the CPU in one second = 16X4Gops = 64Gops. In this way, one CPU calculation time can be calculated. And the execution time can also be known by switching to GPU calculation. Let’s talk about AI chips, such as Google’s TPU v1, which is about 700MHz and has a pulsation array of 256X256, as shown in the figure below. A total of 64K units, and each unit can perform one multiplication and one addition at a time, that is 128K operations.
​
Systolic Data Flow of the Matrix Multiply Unit
In addition to the systolic array, there are other modules, such as activation, etc., which also include multiplication, addition, etc. Look at the peak calculation times of TPU v1 in one second is at least = 128K X 700MHz = 89600Gops ≈ 90Tops, compared with the CPU, you will find that there is a gap of several orders of magnitude in computing capability. Of course, the above data are completely ideal values, and the actual situation may be only 5%. Because the storage on the chip is not large enough, the data will be stored in DRAM, and the data extraction is very slow, then the multiplication logic often has to wait. In addition, the AI algorithm consists of many layers of networks, which must be calculated layer by layer. However, when switching layers, the multiplication logic is free. Therefore, many factors have caused the actual chip to fail to reach the calculation peak.
Although it can be used with slow calculating speed. But at present, the size of the neural network is getting bigger and bigger, and there are more and more parameters. When encountering a large NN model, it takes a few weeks or even a month or two to train, which wastes too much time. If you encounter a sudden situations, such as power outages, have to start all over again. Moreover, if the model is modified, it will take several weeks to know whether it is right or wrong, and no one has so much patience to wait. But with TPU, then you find that you can continue to run by optimizing the parameters during a meal. That is, fast calculation speed enables rapid iteration and development of a stronger AI model.
A deeper comparison can be made clearly for the kernel structure of the GPU. To be sure, the GPU is still relatively fast, at least much faster than the CPU, so most of them currently use the GPU, but don’t forget the GPU disadvantages mentioned earlier, so it is not suitable for large-scale use in data centers. Generally speaking, CPU and GPU are not AI-specific chips. Because they have to realize other functions with other logics inside, which are completely useless for the current AI algorithm. Their performance cannot be used optimally. So it can be seen that it is reasonable for Google to spend money to develop its own AI chip TPU.
​
AI Chips Design Principles
AI chips follow a hardware design rule: Sacrifice certain versatility to improve the efficiency of specific services. A well-understood example is Bitcoin mining. Bitcoin's basic algorithm is the SHA256, which is based on the hash algorithm, so there is no reverse formula, and it can only rely on mining machines to exhaustively enumerate a large number of possible values. For each block, the miner who first tries to get the correct value will be rewarded. So the faster the calculation, the higher the expectation of winning the reward.
For mining chips, both CPU and GPU have become the mainstream ever. However, as time goes by, GPU mining has become less useful. For example, the RTX3090 graphics card can perform 126.5M SHA256 calculations per second under overclocking conditions, and the total computing power of the entire Bitcoin network has exceeded 178.6EH/s, which is equivalent to the sum of the computing power of 1.4 trillion RTX3090 chips. Therefore, the current mainstream is ASIC mining, which are designed for the SHA256 algorithm, so ASIC mining far better than GPU, and even a single mining machine can overcome supercomputing centers. However, ASIC mining sacrifice generality in hardware, making it impossible to perform other computing tasks, and even unable to mine non-SHA256 algorithm blockchain currencies, such as Ethereum (which is a technology that's home to digital money, global payments, and applications. ).
In the example of Bitcoin mining, it can be seen that a dedicated chip is the most suitable: the algorithm or operator is fixed, and it is relatively simple in the application scenario. The AI chip has the same characteristics as the mining machine, emphasizing the computing efficiency in special scenarios, but the implementation scenarios are different.The purpose of the AI chip is to execute AI algorithms efficiently, so it is also called an AI accelerator. In addition, a bunch of other names are derived, such as NPU (neural network processor), TPU (Google’s tensor processor), APU ( AMD’s accelerated processor), whose properties are very similar.
​
AI Chip Classifications
AI chips can be divided into two categories. One is for both training and inference, and GPU or CPU can also do it, FPGA (Altera's Stratix series) either, but the TPU v2 and Bitmain's sophon may have an advantage in energy consumption. Compared with GPU, it retains a considerable number of floating-point processing units (or building a lot of tensor computing units), discards some useless graphics pipelines, and improves the energy consumption. There are fewer players in this part, but they are more interesting. For example, ICLR is also thinking about training with fixed-point devices, while Xilinx hopes that XNOR-net can allow fixed-point devices to participate in training.
The other type is the inference accelerator chip, which simply means running the trained model on the chip. This piece is really full of flowers, such as the Cambrian NPU, Intel Movidius (there is also a Nervana that should be similar to XeonPhi for training), Shenjian DPU, Horizon BPU, Imagination's PowerVR 2NX, ARM's Project Trillium, etc. This type of product provides IP for all developers to integrate deep learning accelerators into SoCs.
In addition, we need to talk about Tegra X2 separately. This is equivalent to a small desktop platform. The ARM processor plus Nvidia GPU can provide complete training and inference capabilities. Of course, the power consumption is also high. For other acceleration chips, I think it is best to divide them into two categories, floating-point and fixed-point. Floating point is only FP16 half-precision, of course, it supports FP16 and INT8, such as Cambrian’s NPU and Intel ‘s Movidius. In terms of fixed-point, it is easy to think of Horizon BPU and Imagination's PowerVR 2NX.
​
Development of AI Chips
Here we will take Google TPU as an example.

1. Parallelization

In order to better accelerate the reasoning of the model, it is necessary to summarize the characteristics of the actual model, find problems, and then make targeted optimization. The algorithm of the AI chip is usually based on the convolutional neural network. Faced with its huge amount of calculation, the CPU can only perform limited calculations in a single clock cycle, and calculation bottlenecks are prone to occur. Where the GPU improves the performance of parallel computing, and converts the equivalent transformation of the convolution operation into a multiplication and addition operation of the matrix, which has become the mainstream hardware for neural network training.
The core of model acceleration is to increase the level of parallelization and perform as many operations as possible in the same clock cycle. So the TPU redesigns a dedicated hardware architecture: the systolic array.
See the figure below for the systolic array of the TPU. The calculation results are not directly output, but "flow" in the array according to the law, and saved after the calculation is completed. It not only increases the degree of parallelism, but also realizes the reuse of parameters and avoids repeated loading. And this array can perform 65536 multiplication and addition operations of 8-bit integers per clock cycle. Under the main frequency of 700MHz, an amazing computing power of 65536 * 700,000,000 * 2 [email protected] is realized.
​
TPU-Systolic-Array
​

1. High-speed On-chip Storage

Google mentioned that in the neural network it deploys, the number of parameters ranges from 5 million to 100 million, and the intermediate results also have a large amount of data. If the convolution operation unit (such as a systolic array) is frequently performed from DDR:
1) Load parameters
2) Read data
3) Write results
Similar to the barrel effect, it is the shortest piece of wood that determines the water capacity, which is the performance bottleneck. The above-mentioned intensive DDR read and write operations make the IO speed easy to become the performance bottleneck of model reasoning. Therefore, TPUv1 is equipped with 28M bytes of on-chip memory (OCM). Although the cost is relatively high, the read and write speed is far superior to that of DDR. What’s more, the TPUv2 is equipped with 16GB of HBM storage, and its IO speed reached 600GB per second.
With a cache (OCM or HBM), frequently used parameters can be loaded into the cache in advance at runtime, and then the convolution kernel can quickly read the data without repeatedly loading from memory. Not only that, the on-chip memory can also save the intermediate results of the network operation, avoiding transfer (write + read back) on the memory, thereby significantly reducing the IO pressure of the memory.
The role of the cache here is similar to the L3 in the traditional CPU. The purpose is to effectively lower the pressure on the memory, so that the powerful computing power of the convolution kernel can be fully utilized, to avoid the situation of invalid waiting for data loading.
And the paper 2 experiments and counts the running performance of several mainstream models on TPU/GPU/CPU.
The horizontal axis is the required computing power divided by the amount of data that needs to be counted from the memory, which represents the computing density corresponding to the unit memory data loading, while the vertical axis is the actual computing power. Therefore, the slash part represents that the computing performance is restricted by the memory IO bandwidth, while the horizontal part represents that the performance has been fully utilized.
​
Operational Intensity: Ops/weight byte (log scale)
As shown in the figure, TPU is significantly better than CPU and GPU at high computing density. It can be seen that under high computing density, TPU can play a greater advantage.

1. CNN Quantization Algorithm

In addition to hardware design, in order to make the AI chip run the model more efficiently, it is necessary to perform lightweight processing on the AI algorithm, which can significantly reduce the size of parameters and the amount of calculation. The current neural network training is basically based on floating-point operations. When training and reasoning on the GPU, 32-bit or 64-bit is usually used to represent a floating-point number. If a lower number of bits is used, such as using 8 bits to represent numbers, the parameter size and calculation amount will be greatly reduced.
The figure below uses an 8-bit integer Q to approximate a 32-bit floating-point number.
​
32 bit float
This operation brings a certain quantization noise, and a good neural network will have a certain anti-noise ability, so the accuracy loss of the algorithm after quantization will be limited and controllable. Another advantage of quantization is that if the input and parameters of the model are quantized into integer numbers (such as int8/uint8), there is no floating-point calculation, which can simplify the hardware design. As mentioned above, the first-generation TPU only supports integer operations, not only because integer operations are more efficient in hardware, but also because the circuit is much simpler.
At present, there are two mainstream quantification methods:
​
Mainstream quantification methods
QAT uses the quantization operator when training the model. It can avoid the loss of precision to the greatest extent, but it cannot use the existing floating-point model and needs to retrain the model from scratch.
PTQ is to quantify the existing floating-point model, although the accuracy and efficiency of the model are not as good as QAT, there is still a great demand for use because there is no need to retrain the model and the threshold for use is lower. Here is a point, how to reduce the loss of accuracy is the goal of the PTQ program, and it has also produced a variety of ideas, such as linear quantization and logarithmic quantization.

1. Efficiency and Cost

In addition to the essential 3x3 convolution, there are dozens of common neural network operators. Because of the chip cost, design difficulty, and operating efficiency, existing AI chips can only support some operators. Therefore, it is necessary to understand and consider hardware capabilities during model training, and try to use operators that can be efficiently supported by AI chips to improve efficiency after deployment. Conversely, when designing the hardware of the AI chip, it is also necessary to consider the needs of algorithm design, and try to support more commonly used operators to improve the degree of freedom in model design.
If the manufacturer has the collaborative development capability between hardware, it can achieve a very good balance between model accuracy, chip cost, and inference speed. This is why more and more terminal manufacturers choose self-developed chips, and AI chip manufacturers often package algorithms for sale.
Note: This original content is from my blog KUKE Electronics.

What are the main requirements for the role?

• You must have full professional proficiency in both written and verbal English & French
• You must be living in Belgium for the last 5 consecutive years
• You must own and use a Smartphone (Android V4.2 or higher or iPhone running iOS version 13.0 or higher) to complete tasks
• The majority of the task types will require both a desktop/laptop and a smartphone
• Gmail must be your primary email
• Active daily user of Gmail and other forms of social media
• You must have familiarity with current and historical business, media, sport, news, social media and cultural affairs in Belgium
• Access to and use of a broadband internet connection and associated computer and software to perform the work, all provided at your own expense
• Experience in use of web browsers to navigate and interact with a variety of content

Working on this project will require you to go through a standard recruitment process (including passing an open book assessment). This is a long-term project and your work will occasionally be subject to quality assurance checks.
​
APPLY HERE