Sure but let's look at some values here : 750M for 8000 jobs (3K+5K) over say 10 years (at 0%) that's a requirement of 9.3K/person/year in state taxes to recover. Just from income, that would require each person to be paid over 150K/year (with about 100K income after taxes). If we include sales taxes at about 9% and we assume that each person spends half of his/her after-tax income, we get to down to a requirement of 101K/year salary per person. There are certainly other indirect sources (you mentioned some) to consider to get to the complete picture here, but still... it seems far-fetched...
In cluster environments, the NVidia products are well ahead of anything made by AMD. And a good portion of the other core components (management, scheduler,...) are already built to support NVidia hardware (with NVML/SMI/...). Some of the Intel accelerators might get close but are also pretty pricey.
Ok, I am not well versed in economics but may be someone here can answer these questions : With roughly 250B$ market cap between the 4 first crypto-currencies, would a collapse of Bitcoin send significant ripples through the "real" economy? Do we know how much of this value was really invested in the currencies versus how much comes from the speculation?
Both Thousand Oaks Optical and Baader are really well-known in the astro community. They both have been making solar filters for a long time and I doubt they would jeopardize customers safety and their brand recognition like this.
No, in astronomy you are interested in reducing the noise for the equivalent of a sub-length. This means that if you combine say 100 images of 5 minutes, the result should be better in terms of noise (and thus DR) than a single 5 minutes exposure. Here we are interested in a totally different normalization which consists in deciding the total number of sub-frames dividing the total exposure (500 minutes with the previous analogy). For a simple stochastic sensor model, the smallest number of sub-frames (1) will *always* be the best.
To prove this, let's say that you will count an average of F electrons in a single pixel over the total exposure time and that each read-out operation follows a 0-mean normal/Gaussian distribution of variance s^2 (normalized in electrons). Then, the stochastic output of the pixel for a single read-out is given by : O ~ Poisson(F) + Normal(0,s^2). If we now decide to divide the observation interval in k sub-frame, we should observe for each : O_k ~ Poisson(F/k) + Normal(0,s^2) as the read-out noise is a constant cost. The standard deviation of the sum of the k sub-frames can be written as follow : sqrt(k*(F/k+s^2)) = sqrt(F+k*s^2). As the local dynamic range D is defined as the ratio between the full flux detected F and the previous standard deviation, we obtain F/sqrt(F+k*s^2). Thus to increase D, you want to reduce the number k of sub-frames recorded down to 1, or reduce the sensor read-noise s (RMS). And ultimately, you will hit the shot-noise limit D = F/sqrt(F).
And every time you read out a sub-frame you are penalized by the read noise... after accumulation of the variances, you end-up with an extremely noisy image. If you want to do that you don't just need a very good quantum efficiency (the probability of a incident photon to be absorbed and to release an electron) you need an almost perfect read-out circuitry (if you want to operate without cooling). Eric Fossum has proposed a "Quanta" binary sensor which would do this with a ~0.15e- RMS read-out noise which has to be compared with the 1.5+e- of the best sensors used in consumer applications today.
I think the term "hacking" for the election fits perfectly... in the context of the tech audience here. Because hacking does not have, for us, the same meaning that it acquired through the media : that of breaching electronic systems, most often for criminal gain (note the extra negative connotation). Instead, here its meaning is about finding and implementing a subversive approach to work around the limitations or rules of a system : all the news manipulation, polls, fact-checking wars are the expressions of that hack to attract the voters one way or another.
Is it really? Are all of these events caused by improper decision of the AI or is it more slight over-corrections imposed by the humans? Also, I guess Uber's vehicles operates mostly in the dense, chaotic traffic of the inner-cities rather than say speedway. 0.8 miles between take-over on speedway would be much more alarming.
You are right in the lottery sense : if your particular phone or app crashes, it is very unlikely that it is due to cosmic rays. However, it might be likely that it happens fairly often around the world. This is similar to the lottery : it is unlikely that you will win, but it is likely that someone will win.
It's all a matter of cross-section of the devices actually. If we want to compare, the IPhone 4 (an old baseline, smaller than today's generation but close to most of the low-cost devices) measures 0.007 m^2, while the top 10 largest data centers (from this random link) combined measure about 1.7 x 10^6 m^2. I am going to assume only 1% of the surface is occupied by sensitive chips (?). You would need about 2.4 millions IPhone 4 to cover the same area. Thus, it is very possible that mobile hardware is victim of more high energy burps than immobile hardware.
These P100 come with sweet HBM2 and around 500GB/s in memory bandwidth... everything based on dense linear algebra (AI, but also physics simulations) is basically flying on them.
Slashdot Mirror
Or your screen main framebuffer...
Sure but let's look at some values here : 750M for 8000 jobs (3K+5K) over say 10 years (at 0%) that's a requirement of 9.3K/person/year in state taxes to recover. Just from income, that would require each person to be paid over 150K/year (with about 100K income after taxes).
If we include sales taxes at about 9% and we assume that each person spends half of his/her after-tax income, we get to down to a requirement of 101K/year salary per person.
There are certainly other indirect sources (you mentioned some) to consider to get to the complete picture here, but still... it seems far-fetched...
In cluster environments, the NVidia products are well ahead of anything made by AMD. And a good portion of the other core components (management, scheduler, ...) are already built to support NVidia hardware (with NVML/SMI/...).
Some of the Intel accelerators might get close but are also pretty pricey.
Ok, I am not well versed in economics but may be someone here can answer these questions :
With roughly 250B$ market cap between the 4 first crypto-currencies, would a collapse of Bitcoin send significant ripples through the "real" economy?
Do we know how much of this value was really invested in the currencies versus how much comes from the speculation?
Both Thousand Oaks Optical and Baader are really well-known in the astro community. They both have been making solar filters for a long time and I doubt they would jeopardize customers safety and their brand recognition like this.
USAF has used SWIR (Short Wave Infra-Red) sensors for a long time already.
No, in astronomy you are interested in reducing the noise for the equivalent of a sub-length. This means that if you combine say 100 images of 5 minutes, the result should be better in terms of noise (and thus DR) than a single 5 minutes exposure. Here we are interested in a totally different normalization which consists in deciding the total number of sub-frames dividing the total exposure (500 minutes with the previous analogy). For a simple stochastic sensor model, the smallest number of sub-frames (1) will *always* be the best.
To prove this, let's say that you will count an average of F electrons in a single pixel over the total exposure time and that each read-out operation follows a 0-mean normal/Gaussian distribution of variance s^2 (normalized in electrons). Then, the stochastic output of the pixel for a single read-out is given by : O ~ Poisson(F) + Normal(0,s^2). If we now decide to divide the observation interval in k sub-frame, we should observe for each : O_k ~ Poisson(F/k) + Normal(0,s^2) as the read-out noise is a constant cost. The standard deviation of the sum of the k sub-frames can be written as follow : sqrt(k*(F/k+s^2)) = sqrt(F+k*s^2). As the local dynamic range D is defined as the ratio between the full flux detected F and the previous standard deviation, we obtain F/sqrt(F+k*s^2). Thus to increase D, you want to reduce the number k of sub-frames recorded down to 1, or reduce the sensor read-noise s (RMS). And ultimately, you will hit the shot-noise limit D = F/sqrt(F).
And every time you read out a sub-frame you are penalized by the read noise... after accumulation of the variances, you end-up with an extremely noisy image. If you want to do that you don't just need a very good quantum efficiency (the probability of a incident photon to be absorbed and to release an electron) you need an almost perfect read-out circuitry (if you want to operate without cooling). Eric Fossum has proposed a "Quanta" binary sensor which would do this with a ~0.15e- RMS read-out noise which has to be compared with the 1.5+e- of the best sensors used in consumer applications today.
Too may resources, not enough problems...
I think the term "hacking" for the election fits perfectly... in the context of the tech audience here. Because hacking does not have, for us, the same meaning that it acquired through the media : that of breaching electronic systems, most often for criminal gain (note the extra negative connotation).
Instead, here its meaning is about finding and implementing a subversive approach to work around the limitations or rules of a system : all the news manipulation, polls, fact-checking wars are the expressions of that hack to attract the voters one way or another.
arXiv:1703.08544, D. TRUMP: Data-mining Textual Responses to Uncover Misconception Patterns
Is it really? Are all of these events caused by improper decision of the AI or is it more slight over-corrections imposed by the humans?
Also, I guess Uber's vehicles operates mostly in the dense, chaotic traffic of the inner-cities rather than say speedway. 0.8 miles between take-over on speedway would be much more alarming.
Yeah, they are about to roll-out ICBM terrestrial delivery next week... and it's a straight road ahead after that...
And here is why...
You are right in the lottery sense : if your particular phone or app crashes, it is very unlikely that it is due to cosmic rays. However, it might be likely that it happens fairly often around the world. This is similar to the lottery : it is unlikely that you will win, but it is likely that someone will win.
It's all a matter of cross-section of the devices actually. If we want to compare, the IPhone 4 (an old baseline, smaller than today's generation but close to most of the low-cost devices) measures 0.007 m^2, while the top 10 largest data centers (from this random link) combined measure about 1.7 x 10^6 m^2. I am going to assume only 1% of the surface is occupied by sensitive chips (?). You would need about 2.4 millions IPhone 4 to cover the same area. Thus, it is very possible that mobile hardware is victim of more high energy burps than immobile hardware.
These P100 come with sweet HBM2 and around 500GB/s in memory bandwidth... everything based on dense linear algebra (AI, but also physics simulations) is basically flying on them.
When the technological revolution is here to serve the communist revolution...
Frickin' soviets...
Here's the list of fucks I give :
You'd better patent that one fast... or let Apple do it for you...
No no no... I think he meant distributing a Linux image through PXE...
Well, the guy is the first victim of automation : a machine is speaking for him...
Ah yes, the goold ol' : /dev/earth
# shred --flood
... Systemd Edition?
"Clippy AI was re-born... Just as Intelligent as before... More Artificial than ever..."
I am just terrified right now.
Wait until Google Research Scientists learn about matrices...