https://en.wikipedia.org/wiki/History_of_supercomputing#Petascale_computing_in_the_21st_century
Supercomputer from 2004 (IBM Blue) could perform at 70.720 Tflops.
Geforce RTX 4090 can perform at 82.58 tflops of processing power.
Geforce RTX 4090 was released in 2022.
In 18 years we got tech that would take an entire room or two to fit it in and miniaturized it to fit within a desktop pc.
What we call the cutting edge in computing today should be accessible to your kid as a consumer level product by the time he or she is 18 years old.
The human genome project, a project meant to get a complete sequence of the human genome, cost 2.7 billion dollars to complete. Today, human genome mapping costs are as low as 600 dollars in some cases.
https://3billion.io/blog/whole-genome-sequencing-price-and-trends-2024
https://biology.mit.edu/the-human-genome-project-turns-20-heres-how-it-altered-the-world/
In 21 years the price of genome sequencing for humans went down 4.5 million times.
Another case of the price of tech falling to consumer grade prices within one generation since the first version.
It puts into perspective exactly how fast technology is moving from a practical standpoint rather than a theoretical one. The most expensive tech out there today is going to be what the next generation would be carrying around as a consumer product, and this trend will continue, generation after generation.
The ISS ( fist segment launched in 1998 ) cost 150 billion USD to finish development. The Tiangong space station ( by China 2022 ) meanwhile cost only about 8.5 billion USD.
A 17.64X reduction in price in a 24 year timespan.
Conclusion:
The cost is going down and quality of tech products is going up 10-20X every generation. With the current scale of ongoing research into lab grown food, we might even expect a decline in the prices of food products over time, or at least an increase in their quality over time. The quality of consumer goods available to humans is increasing by a wide margin every generation. The only issue currently lies in solving continental scale and planet scale problems such as global warming, clean water access, global food supply chains, etc. With the population growth rate going down while at the same time tech advancement rates keep stable or accelerate, it is pretty much guaranteed that your children will have a higher quality of life than you as long as you are not poor or in the wrong place on the planet.
We are currently moving at rates where we should be able to solve all our problems faster than we create new problems. A continual improvement in quality of life.
The slowdown is most likely coming from two points:
1. If it ain't broke don't fix it attitude, where humans won't solve a problem until it is effecting too many people.
2. People so used to doubling or tripling of specs in the technological sphere that they are no longer focused enough on improving the efficiency of the software they build on top of the hardware.
We are two industrial revolutions away from living in a post scarcity everybody is provided for society.
Stay strong, just survive the next 40 years and it will all work out.
Jump in the discussion.
No email address required.
GPUs are not equivalent to CPUs in performance. the high flops are only achievable for certain operations. branching is slow to impossible.
Solvers for boolean satisfiability problems are very useful in a wide range of applications and also very difficult to implement on GPUs. The best one I know (not necessarily the best one there is) is https://www.win.tue.nl/~awijs/articles/parafrost_gpu.pdf --- a large part of that still has to be done on CPU, only certain parts can be done by the GPU. One of those parts is 50x faster on GPU than on CPU, but the overall algorithm is only about 2x faster than a state of the art CPU-only solver.
Even on CPUs the speedup compared to older CPUs is not the same for all problems. E.g. AVX-512 instructions perform an operation on eight 32bit numbers simultaneously, yielding an 8x speedup compared to sequential execution, and many common algorithms can take advantage of that. But for problems where that isn't possible a state-of-the-art CPU seems 8x slower than what you would expect when comparing its performance to a CPU from 2004 on e.g. linear algebra problems.
Jump in the discussion.
No email address required.
Supercomputers were always for vectorial application
Boolean satisfiability can easily be solved by gpu, just bruteforce it.
Jump in the discussion.
No email address required.
but then you don't get the x10000 (?) speedup over a 2004 consumer PC.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
None of these words are in the Bible
Jump in the discussion.
No email address required.
an omnipotent, omniscient, and benevolent god would have mentioned them. QED
Jump in the discussion.
No email address required.
More options
Context
More options
Context
While this is largely true, to my knowledge the IBM Blue Gene supercomputer also had massively wide execution units in order to achieve those TFlop numbers, so they weren't like regular desktop/server CPUs of its time. The gigabit interconnect bandwidth between nodes probably didn't do it any favors either.
A combination of a pair of 96 core Genoa-X's, 12 TiB of RAM, and a couple 4090's would probably compare quite favorably for many, many workloads given equivalent amount of time spent optimizing the code for the target hardware.
Jump in the discussion.
No email address required.
For sure.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
So what year supercomputer is the modern day computer as good as?
Jump in the discussion.
No email address required.
I don't know honestly. It depends on the type of problem. For dense linear algebra your 2004 estimate is pretty accurate.
Jump in the discussion.
No email address required.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
More options
Context