I obtained my PhD in France and now I am a professor in a US universities. I think you fell into the classical trick of publish or perish. US academic are under constant pressure to get material out as reviews are based on it and your expertise is estimated by the length of your publication list. So the publication rate is very high is the US compared to the rest of the world, europe for instance have much lower requirement for publication. So they do publish less articles.
I am not sure that European contribution to research is lesser than the US contribution to research (This is a real statement, I do not know how to quantify that properly). But I am pretty sure that because of publication policies, you end up reading an US authors much more often than European ones.

I know a couple professional chess players. Against somebody that do not play often, it typically does not take them ANY TIME to think, the time they take to play, is exactly the time it takes to move the piece from one square to the other.

What is more surpising is that it took only 9 moves. That means you do not know how to defend, which is not surprising since gates is probably not much of a chess player.

Solar and wind energy are not producing energy all the time. When there is no wind, wind turbine will not produce energy. When it is night, solar will not produce energy.
Storing energy is quite difficult and ineeficient. So it is not realistic to stay we will store solar energy for when it is night.
The energy consumption is not constant over time, you need to be able to deliver the proper amount of energy at any time. This is why nuclear power plant did not make coal power plant obsolete. Because starting a nuclear powerplant takes a long time, while a coal one is much faster.

I do not think we should rely on a single energy source. We need to rely on a mix of energy sources so that when one fails, other ones can pick up the pieces.

As a non US citizen, living in the US for a while, I must say that I really don't understand why SSN is used as a form of authentification. You pretty much give it to everybody. You give it to all companies that are either: banking, insurance, tax or health related. I think SSN is great as a way to identify people, but it is definitely terrible for proving who you are.

Where are you Princess Peach? Here comes the plumber!

What a jerk!

The CPU-GPU latency is already very small, smaller than a millisecond. and I assume that most of the CUDA startup latency comes configuring the kernel launch and not from the interconnect. Because PCI-express has very low lateny, infiniband card get network communication with a latency less than 2 microseconds. That's about 6000 CPU cycle (assuming 3GHz CPU).

If you want to gain by removing latency, you need the computation to be VERY small and frequent.

Moreover, we are very good at overlapping communication and computations so that latency is masked. That is what all the multiple kernels of cuda 4.0 and later was about.

The thing is you will never have similar specs because of power dissipation, CPUs already outputs north of 80W (actually all xeon E7 are over 100W), 4000 GPU-cores chips are over 100W. Dissipating 200W is going to be difficult. I think to make the architecture work, you will have to scale it down. (If only because you won't have enough memory bandwidth to feed data to all these cpu and gpu cores.)

I am a little bit skeptical about that. I am not really sure how much it will really change things. The use case actually seems very thin to me. You need a kernel which is compute intensive and where the data transfer from memory to the core is expensive. Because if there is little data to transfer, then the overhead is small. I read some benchmarks from AMD and only few kernels seemed to be in the sweet spot. On top of that becasue of the memory architecutre, I feel like raw memory to core bandwidth will be closer to what a cpu get (50GB/s) rather than what a gpu get (250GB/s).

Anyway, I'll probably buy one and give it to a student to play with.

Sure, you are right. We NEVER hear about Obamacare. When I see a TV screen, there is NEVER a debate on Obamacare. On radio, they talk about everything BUT Obamacare. And the newspaper NEVER mentionned it. We really need to hear more about it!

There are lots of advice, but I'll contribute to it. I have been teaching for a few years and I was a good student before that. So I guess my advice are meaningful.

1/ SLEEP and HEALTH! Sleep is critical. Health is critical. Plenty of people apparently already pointed it out. But your body and your brain do need the sleep. Getting 2 more hours of sleep is typically more important than getting 2 more hours of study. Also, you will sleep better if your optic nerve is not too stimulated before you fall asleep. So try to stay aways from screens in the last hour of your day.

2/ Beware of slides. Many classes are slide based nowadays. I find it idiotic in most topics. Slides are typically short bullet points, they won't help you remembering everything. Also they contribute to student not writing during the class. Most people have a visual memory which means that you need to SEE the material. Also you learn better when you are active rather than passive. Most of the time you can accomplish both by writing your class.

3/ You need to study in the calm. No TV or radio in the background do not help you study, they are likely to be distracting you. Music can be helpful to cover background noise, I'd stay away from high BPM tracks or tracks with understandable lyrics. Classical, instrumental, electronic or pop music in a language you do not speak are typically good candidates.

4/ You need to study smart. Typically I find that going over notes, correcting them or re-writing them helps a lot to get the data in. I tend to spend at least 30 minutes per hour of lecture. Making picture illustration (schemas) helps. Before a test, I usually try to make a 2 or 3 pages summary of what is important. Once again, the point is not to have it, but to make it. Doing extra exercise help.

5/ Out of class material. The way the informations are presented in class is not the only way to present the informations. I found it extremely useful in the past to pick other sources of informations. Go to your library and pick a book related to what you study and read it when you have time. It might say the same thing as your class in a different way, or just say something different. But for sure that will help you getting different perspective of your subject.

6/ Exercises. There is no secret in how to deal with exercise. You need to do exercise. If you are too slow, you need to do more. It might be important to understand WHY you are slow. Do you not understand the techniques involved? Or does it take you too much time to apply them?

Finally remember that going to college is essentially a full time job. Even if you are enrolled for only 15 credits, it should probably take about 45 hours of your time each week.

While I aggree with your first sentence. The second one is puzzling to me. I find it natural that some people understood the concept of multiplication at that time. It is not very old, it is essentially 200BC. There was plenty of commerce, armies and large government at that time which uses lots of multiplications. Pythagoras' work is about 300 years older than that and is much more complex than multiplications.

It is nice to have the artifact, but it is not very surprising IMHO.

There are parallel strategies to do some of these things. As far as I know text layouting is mostly done with dynamic programming algorihtms. These algorithms are usually very parallel.
Even if they are not, you can always use some kinds of speculative algorithms to deal with that. You assume the 3 most likely scenario for line 1 and while line 1 is being processed, you layout line 2 multiple times using different assumption on line 1. This will not give you perfect parallelism but it will give you some improvement.

As for image decompression. I am not sure how PNG works, but parallel compression and decompression of text, images, videos and sound streams as kept the parallel programming community active for a very long time. If PNG can really not be parallelized (which I doubt, but it might not be efficient) then we will move to different formats with time which supports parallel processing.

Many algorithms have been claimed as impossible or difficult to parallelize. With time, they are falling the one after the other. The one that are inherently sequential are often replace with "just as good" other algorithm which can be performed in parallel.

Have faith in the parallel computing scientific community, we are here to help! :)

API is not meant as a replacement for PCI-e. That's just the technology that links multiple processors together (and memory controllers). KNL is essentially the next generation MIC processor. The current generation is KNC which is a separate PCI-e card. I think it is in that sense that QPI replaces PCI-e.

I don't understand. Mic is your regular cache based architecture. Accessing L1 cache in mic is very fast (3 cycle latency if my memory is correct). You have similar register constraints on mic with 32 512-bit vectors per thread(core maybe). Both architectures overlap memory latency by using hardware threading.

I programmed both mic and gpu, mainly on sparse algebra and graph kernels. And quite frankly there are differences but i find much more alike than most people acknowledge. The main difference in my opinion being the programming model where gpus are used with millions of threads while mic is better used with less number of threads and more of a work pool. Atomics are really fast in gpus and not so fast in mic. But you also have much more fine thread synchronization opportunities in mic whichsomewhat remove the interest of fast atomics.

Actually the in-order execution isn't so much of a problem in my experience. The vectorization is a real problem. But you essentially have the same problem except it us hidden in the programming model. But the performance problem are here as well.

Anybody that understand gpu architecture enough to write efficient code there won;t have much problem using the mic architecture. The programming model is different but the key diffucultues are essentially the same. If you think about mic simd element as a cuxa thread, the programming different are mostly syntactical.