Daniel Lemire's blog
Science and Technology links (April 13 2024)
1.
1.
1. Our computer hardware exchange data using a standard called PCI Express. Your disk, your network and your GPU are limited by what PCI Express can do. Currently, it means that you are limited to a few gigabytes per second of bandwidth. PCI Express is about to receive a big performance boost in 2025 when PCI Express 7 comes out: it will support bandwidth of up to 512 GB/s. That is really, really fast. It does not follow that your disks and graphics are going to improve very soon, but it provides the foundation for future breakthroughs.
2. Sperm counts are down everywhere and the trend is showing no sign of slowing down. There are indications that it could be related to the rise in obesity.
3. A research paper by Burke et al. used a model to predict that climate change could reduce world GPD (the size of the economy) by 23%. For reference, world GDP grows at a rate of about 3% a year (+/- 1%) so that a cost of 23% is about equivalent to 7 to 8 years without growth. It is much higher than prior predictions. Barket (2024) questions these results:
4. The firm McKinsey reports finding statistically significant positive relations between the industry-adjusted earnings and the racial/ethnic diversity of their executives. Green and Hand (2024) fail to reproduce these results. They conclude: despite the imprimatur given to McKinsey’s studies, their results should not be relied on to support the view that US publicly traded firms can expect to deliver improved financial performance if they increase the racial/ethnic diversity of their executives.
5. Corinth and Larrimore (2024) find that after adjusting for hours worked, Generation X and Millennials experienced a greater intergenerational increase in real market income than baby boomers.
source
Science and Technology links (April 13 2024)
1.
1.
1. Our computer hardware exchange data using a standard called PCI Express. Your disk, your network and your GPU are limited by what PCI Express can do. Currently, it means that you are limited to a few gigabytes per second of bandwidth. PCI Express is about to receive a big performance boost in 2025 when PCI Express 7 comes out: it will support bandwidth of up to 512 GB/s. That is really, really fast. It does not follow that your disks and graphics are going to improve very soon, but it provides the foundation for future breakthroughs.
2. Sperm counts are down everywhere and the trend is showing no sign of slowing down. There are indications that it could be related to the rise in obesity.
3. A research paper by Burke et al. used a model to predict that climate change could reduce world GPD (the size of the economy) by 23%. For reference, world GDP grows at a rate of about 3% a year (+/- 1%) so that a cost of 23% is about equivalent to 7 to 8 years without growth. It is much higher than prior predictions. Barket (2024) questions these results:
It is a complicated paper that makes strong claims. The authors use thousands of lines of code to run regressions containing over 500 variables to test a nonlinear model of temperature and growth for 166 countries and forecast economic growth out to the year 2100. Careful analysis of their work shows that they bury inconvenient results, use misleading charts to confuse readers, and fail to report obvious robustness checks. Simulations suggest that the statistical significance of their results is inflated. Continued economic growth at levels similar to what the world has experienced in recent years would increase the level of future economic activity by far more than Nordhaus’ (2018) estimate of the effect of warming on future world GDP. If warming does not affect the rate of economic growth, then the world is likely to be much richer in the future, with or without warming temperatures.
4. The firm McKinsey reports finding statistically significant positive relations between the industry-adjusted earnings and the racial/ethnic diversity of their executives. Green and Hand (2024) fail to reproduce these results. They conclude: despite the imprimatur given to McKinsey’s studies, their results should not be relied on to support the view that US publicly traded firms can expect to deliver improved financial performance if they increase the racial/ethnic diversity of their executives.
5. Corinth and Larrimore (2024) find that after adjusting for hours worked, Generation X and Millennials experienced a greater intergenerational increase in real market income than baby boomers.
source
2018年听过他的课，当时思路还很清晰，至少比我清晰
Daniel Lemire's blog
A simple algorithm to compute the square root of an integer, byte by byte
A reader asked me for some help in computing (1 – sqrt(0.5)) to an arbitrary precision, from scratch. A simpler but equivalent problem is to compute the square root of an integer (e.g., 2). There are many sophisticated algorithms for such problems, but we want something relatively simple. We’d like to compute the square root bit by bit…
For example, the square root of two is…
1. 5 / 4
2. 11 / 8
3. 22 / 16
4. 45 / 32
5. 90 / 64
6. 181 / 128
7. …
More practically, 8-bit by 8-bit, we may want to compute it byte by byte…
1. 362 / 256
2. 92681 / 65536
3. 23726566 / 16777216
4. …
How can we do so?
Intuitively, you could compute the integer part of the answer by starting with 0 and incrementing a counter like so:
Indeed, the square of the integer part cannot be larger than the desired power.
You can repeat the same idea with the fractional part… writing the answer as x1+x2/B+... smaller terms.
It will work, but it involves squaring ever larger numbers. That is inefficient.
We don’t actually need to compute powers when iterating. If you need to compute x**2, (x+1)**2, (x+2)**2, etc. You can instead use a recursion: if you have computed (x+n)**2 and you need the next power, you just need to add 2(x+n) + 1
because that’s the value of (x+n+1)**2 – (x+n)**2.
Finally, we get the following routine (written in Python). I left the asserts in place to make the code easier to understand:
The algorithm could be further optimized if you needed more efficiency. Importantly, it is assumed that the basis is not too large otherwise another type of algorithm would be preferable. Using 256 is fine, however.
Obviously, one can design a faster algorithm, but this one has the advantage of being nearly trivial.
Further reading: A Spigot-Algorithm for Square-Roots: Explained and Extended by Mayer Goldberg
Credit: Thanks to David Smith for inspiring this blog post.
source
A simple algorithm to compute the square root of an integer, byte by byte
A reader asked me for some help in computing (1 – sqrt(0.5)) to an arbitrary precision, from scratch. A simpler but equivalent problem is to compute the square root of an integer (e.g., 2). There are many sophisticated algorithms for such problems, but we want something relatively simple. We’d like to compute the square root bit by bit…
For example, the square root of two is…
1. 5 / 4
2. 11 / 8
3. 22 / 16
4. 45 / 32
5. 90 / 64
6. 181 / 128
7. …
More practically, 8-bit by 8-bit, we may want to compute it byte by byte…
1. 362 / 256
2. 92681 / 65536
3. 23726566 / 16777216
4. …
How can we do so?
Intuitively, you could compute the integer part of the answer by starting with 0 and incrementing a counter like so:
x1 = 0
while (x1+1)**2 <= M:
x1 += 1
Indeed, the square of the integer part cannot be larger than the desired power.
You can repeat the same idea with the fractional part… writing the answer as x1+x2/B+... smaller terms.
x2 = 0
while (x1*B + x2 + 1)**2 <= M*B**2:
x2 += 1
It will work, but it involves squaring ever larger numbers. That is inefficient.
We don’t actually need to compute powers when iterating. If you need to compute x**2, (x+1)**2, (x+2)**2, etc. You can instead use a recursion: if you have computed (x+n)**2 and you need the next power, you just need to add 2(x+n) + 1
because that’s the value of (x+n+1)**2 – (x+n)**2.
Finally, we get the following routine (written in Python). I left the asserts in place to make the code easier to understand:
B = 2**8 # or any other basis like 2 or 10
x = 0
power = 0
limit = M
for i in range(10): # 10 is the number of digits you want
limit *= B**2
power *= B**2
x*=B
while power + 2*x + 1 <= limit:
power += 2*x + 1
x += 1
assert(x**2 == power)
assert(x**2 <= limit)
# x/B**10 is the desired root The algorithm could be further optimized if you needed more efficiency. Importantly, it is assumed that the basis is not too large otherwise another type of algorithm would be preferable. Using 256 is fine, however.
Obviously, one can design a faster algorithm, but this one has the advantage of being nearly trivial.
Further reading: A Spigot-Algorithm for Square-Roots: Explained and Extended by Mayer Goldberg
Credit: Thanks to David Smith for inspiring this blog post.
source
省流：Neoverse V2
Google Axion Processors – Arm-based CPUs designed for the data center https://cloud.google.com/blog/products/compute/introducing-googles-new-arm-based-cpu/
Google Cloud Blog
Introducing Google’s new Arm-based CPU | Google Cloud Blog
The custom Axion Arm-based processor is designed for general-purpose workloads like web and app servers, databases, analytics, CPU-based AI, and more.
Chips and Cheese
Intel’s Ambitious Meteor Lake iGPU
#ChipAndCheese
Telegraph | source
(author: clamchowder)
Intel’s Ambitious Meteor Lake iGPU
#ChipAndCheese
Telegraph | source
(author: clamchowder)
Arch Linux: Recent news updates
Increasing the default vm.max_map_count value
The vm.max_map_count paramater will be increased from the default
This change should help address performance, crash or start-up issues for a number of memory intensive applications, particularly for (but not limited to) some Windows games played through Wine/Steam Proton. Overall, end users should have a smoother experience out of the box with no expressed concerns about potential downsides in the related proposal on arch-dev-public mailing list.
This
Before upgrading, in case you are already setting your own value for that parameter in a
source
(author: Robin Candau)
Increasing the default vm.max_map_count value
The vm.max_map_count paramater will be increased from the default
65530 value to 1048576.This change should help address performance, crash or start-up issues for a number of memory intensive applications, particularly for (but not limited to) some Windows games played through Wine/Steam Proton. Overall, end users should have a smoother experience out of the box with no expressed concerns about potential downsides in the related proposal on arch-dev-public mailing list.
This
vm.max_map_count increase is introduced in the 2024.04.07-1 release of the filesystem package and will be effective right after the upgrade.Before upgrading, in case you are already setting your own value for that parameter in a
sysctl.d configuration file, either remove it (to switch to the new default value) or make sure your configuration file will be read with a higher priority than the /usr/lib/sysctl.d/10-arch.conf file (to supersede the new default value).source
(author: Robin Candau)
杰哥的{运维，编程，调板子}小笔记
开发一个链接器（3）
前言
这个系列的前两篇博客实现了一个简单的静态链接器，它可以输入若干个 ELF .o 文件，输出 ELF 可执行文件。接下来，我们进一步支持动态库：输入若干个 ELF .o 文件，输出 ELF 动态库。
source
开发一个链接器（3）
前言
这个系列的前两篇博客实现了一个简单的静态链接器，它可以输入若干个 ELF .o 文件，输出 ELF 可执行文件。接下来，我们进一步支持动态库：输入若干个 ELF .o 文件，输出 ELF 动态库。
source
唉原来王羽佳都已经37岁了，怪不得看起来老了。想当年第一次看到王羽佳的演奏还是在这个古代视频
https://b23.tv/o6E1gtH
https://b23.tv/o6E1gtH
Chips and Cheese
Inside Control Data Corporation’s CDC 6600
#ChipAndCheese
Telegraph | source
(author: clamchowder)
Inside Control Data Corporation’s CDC 6600
#ChipAndCheese
Telegraph | source
(author: clamchowder)
本频道温馨提醒：若遇到疑似诈骗可拨打反诈咨询电话 96110 咨询。若发现上当受骗，请注意及时保存好相关证据，并第一时间拨打 110 报警。
杰哥的{运维，编程，调板子}小笔记
开发一个链接器（2）
前言
这个系列的第一篇博客实现了一个最简单的静态链接器，它可以输入单个 ELF .o 文件，输出 ELF 可执行文件。接下来，我们需要把它升级到支持输入两个或者更多的 ELF .o 文件。
source
开发一个链接器（2）
前言
这个系列的第一篇博客实现了一个最简单的静态链接器，它可以输入单个 ELF .o 文件，输出 ELF 可执行文件。接下来，我们需要把它升级到支持输入两个或者更多的 ELF .o 文件。
source
