What seems to be happening is that economic Moore's Law is continuing: computation is becoming cheaper, which means that the same amount of money can buy more of it. (This may be on hold post-2020, due to the shortage crisis.) On the other hand, clock speeds aren't being pushed past 4 GHz because (a) thermal issues become severe, and (b) there is no real industrial need--single-threaded programs can already do millions of things in time we consider imperceptible, so slow computation is usually a software or network problem in the first place.
In theory, this could be a good thing, because parallel and concurrent computing are hard and something has to flush out the mediocre programmers (the ones who put with Agile Scrum and Jira) who've flooded the market. In practice, though, it seems to be pushing society in the wrong direction. Just as regular people don't benefit from low or negative interest rates (since they'll still pay high interest on personal loans), regular people aren't going to be able to set up a cluster on AWS--or even know why they might want to. The result, then, of the move toward cloud computing seems to be worsening, not improvement, of the power imbalance between capital and labor.
In the long long run, we're still seeing technical advances, which are cause for hope. In the short term, though, the picture is ugly because the technical gains are mostly going to the few who have the capital necessary to profit from them.
3D engines are still extremely limited by single thread performance, at least if you want fast refresh rates.
That's the case for VR were Valve as shown with their headset that 120Hz, and even 144Hz, is a huge comfort and presence improvement over the 90fps standard (or even 72 on Oculus Quest 2) from the rest of the industry.
Hitting the 144fps mark, means a CPU fast enough to consistently clear the main loop, while leaving enough time for rendering (!), in a shared <7ms time budget.
Creation is cool in VR, I've had a lot of fun in tilt brush. Apparently there are ways to be productive building in VR from 3D engines too. Some even try coding even though the resolution is still extremely insufficient.
Meta promises that most work, solitary or collaborative will eventually be better done in VR/AR.
However, while I love VR, I sure won't be spending much time suffering at low FPS.
We're far from not needing faster CPUs anytime soon, there are not enough John Carmack and Michael Abrash able and willing to optimise the shit out of every routine in assembly.
This makes sense, but it might be a mercy that progress on VR is slow, given that it's Facebook ("Meta") that is trying to push it.
It'll be best for all of us if technology slows, except perhaps for those innovations that reduce energy usage and GHG emissions, until corporate capitalism is overthrown (which may be 50 years or more away). It's actually a good thing, under our current economic system, that technological capabilities are so low compared to what they could be.
>(b) there is no real industrial need--single-threaded programs can already do millions of things in time we consider imperceptible, so slow computation is usually a software or network problem in the first place.
Or a problem of the difference in retrieval speed between CPU cache and RAM -- let alone the HDD.
Even though 2017 is on the date, this appears to cover up to 2012, which is a bummer, since it is an interesting topic, but I'd rather see an analysis that included more recent data. I feel that the time period covered represents only the transition from single to multi core, and I wonder what the impact looks like as we have developed tools and techniques for dealing with parallel software. Still, was an interesting read!
What seems to be happening is that economic Moore's Law is continuing: computation is becoming cheaper, which means that the same amount of money can buy more of it. (This may be on hold post-2020, due to the shortage crisis.) On the other hand, clock speeds aren't being pushed past 4 GHz because (a) thermal issues become severe, and (b) there is no real industrial need--single-threaded programs can already do millions of things in time we consider imperceptible, so slow computation is usually a software or network problem in the first place.
In theory, this could be a good thing, because parallel and concurrent computing are hard and something has to flush out the mediocre programmers (the ones who put with Agile Scrum and Jira) who've flooded the market. In practice, though, it seems to be pushing society in the wrong direction. Just as regular people don't benefit from low or negative interest rates (since they'll still pay high interest on personal loans), regular people aren't going to be able to set up a cluster on AWS--or even know why they might want to. The result, then, of the move toward cloud computing seems to be worsening, not improvement, of the power imbalance between capital and labor.
In the long long run, we're still seeing technical advances, which are cause for hope. In the short term, though, the picture is ugly because the technical gains are mostly going to the few who have the capital necessary to profit from them.
3D engines are still extremely limited by single thread performance, at least if you want fast refresh rates.
That's the case for VR were Valve as shown with their headset that 120Hz, and even 144Hz, is a huge comfort and presence improvement over the 90fps standard (or even 72 on Oculus Quest 2) from the rest of the industry.
Hitting the 144fps mark, means a CPU fast enough to consistently clear the main loop, while leaving enough time for rendering (!), in a shared <7ms time budget.
Creation is cool in VR, I've had a lot of fun in tilt brush. Apparently there are ways to be productive building in VR from 3D engines too. Some even try coding even though the resolution is still extremely insufficient. Meta promises that most work, solitary or collaborative will eventually be better done in VR/AR.
However, while I love VR, I sure won't be spending much time suffering at low FPS.
We're far from not needing faster CPUs anytime soon, there are not enough John Carmack and Michael Abrash able and willing to optimise the shit out of every routine in assembly.
The end of Dennard Scaling is a tragedy.
This makes sense, but it might be a mercy that progress on VR is slow, given that it's Facebook ("Meta") that is trying to push it.
It'll be best for all of us if technology slows, except perhaps for those innovations that reduce energy usage and GHG emissions, until corporate capitalism is overthrown (which may be 50 years or more away). It's actually a good thing, under our current economic system, that technological capabilities are so low compared to what they could be.
>(b) there is no real industrial need--single-threaded programs can already do millions of things in time we consider imperceptible, so slow computation is usually a software or network problem in the first place.
Or a problem of the difference in retrieval speed between CPU cache and RAM -- let alone the HDD.
That falls under slow software, since optimized programs will use prefetching so latency isn't an issue.
Even though 2017 is on the date, this appears to cover up to 2012, which is a bummer, since it is an interesting topic, but I'd rather see an analysis that included more recent data. I feel that the time period covered represents only the transition from single to multi core, and I wonder what the impact looks like as we have developed tools and techniques for dealing with parallel software. Still, was an interesting read!
He’s looking at that inflection, which is part of the difference-in-differences trend in econometrics
https://en.wikipedia.org/wiki/Difference_in_differences
The idea is to look for events that provide you with a natural experiment, like the switch to multicore.
Anyway, the cloud has really accelerated parallel loads, so you are right: I’d love to hear about now
https://twitter.com/lookinggdlouis/status/154003428764216934...