Hacker News Reader

Comment by JesperRavn | Hacker News Reader

It's fascinating that this is so close to reality, at least in terms of preserving the information that encodes our memories and identities.

Of course there is the issue of reviving the person, either biologically or digitally. As computing power grows, I don't see any barrier to simulating an entire human brain. Sure it will take time: according to this[0] and applying Moore's law, it will take about 60 years before we can simulate a brain in real time with a single machine.

This whole field makes me very uncomfortable. I'm not saying it's wrong, but the potential for harm (particularly to those being simulated) is very great. There's a Greg Egan story where reanimated humans live as "AI" characters a game simulation, waiting for when society will wake up and grant them rights.

[0] http://www.extremetech.com/extreme/163051-simulating-1-secon...

View on HN

Which Greg Egan story is that?

I can't remember (possibly not Greg Egan). In case someone else knows, the main plot is that the characters have to act as inhabitants of a primitive village, in a computer game/simulation.

Probably neither of these (depending on your definition of primitive), but they're interesting anyway:

The Cookie Monster[1] by Vernor Vinge

Glasshouse[2] by Charles Stross

[1] https://en.wikipedia.org/wiki/The_Cookie_Monster_%28novella%... (WP link, contains spoilers)

[2] https://en.wikipedia.org/wiki/Glasshouse_%28novel%29

Possibly he refers to Permutation City. Absolutely necessary read.

Almost all of them.

>and applying Moore's law, it will take about 60 years before we can simulate a brain in real time with a single machine.

What Moore's law? This has already stopped.

Re Moore's law, the best answer I could get was from Intel via Wikipedia[0] "our cadence today is closer to two and a half years than two." So I wouldn't call that "stopped" but slowed. On the other hand, there are other kinds of advancement that don't really fit into any kind of "law" but might make up for this slowdown, resulting in long term exponential growth, e.g. the transition from CPU to GPU computing.

[0] https://en.wikipedia.org/wiki/Moore%27s_law

Intel was slow (delayed) with the Skylake, and poised to get even more delayed in delivering the next series, that jumps to 10nm.

So this "2.5 years" is just the first slowdown -- it gets worse. It's not like we just change 2 to 2.5 and continue for the next 10-20 years.

Probably not even 2 or 3. After 10nm, which is already challenging, you're basically screwed with any kind of process known/used today. And the costs for a fab are into the billions (and get higher with lower sizes), so it's not like small players can even compete in that field.

And that's for multicore -- so unless we also get a way to write more programs to be massively parallel (which lots of algorithms cannot be made) -- we don't get the "free boost" from Moore's law with regards to speed that we used to get.

Not only it's dropped to 2.5 years, but it's also more useless, just giving us more transistors in not readily usable cores.

That's why Intel devotes large part of the CPU to the embedded GPU -- it doesn't have much else to do with it.

Multiple cores are useful for virtualization or any collection of programs which each use one core. If the future is one of many cores, we can invest into software architectures that make it easy for parallel processes/VMs to cooperate securely, e.g. message bus.

That's why Intel devotes large part of the CPU to the embedded GPU -- it doesn't have much else to do with it.

I don't quite understand your point here, but if you were intending to refute my point regarding GPU computing, I think you are not understanding the nature of the CPU to GPU shift. GPGPU computing is a way to increase effective computing power without advancements in semiconductor technology. GPUs have a fundamentally different architecture that allows them to get much more processing power from the same silicon. They work differently to CPUs (amortizing memory access latency across many hardware threads instead of trying to minimize it, sharing control flow across many cores to minimize that cost, and requiring manual control of memory locality) but on many domains including training neural networks, they are much more powerful.

So GPUs are certainly not a diversion CPU makers get into when they can't make progress on CPUs.

>I think you are not understanding the nature of the CPU to GPU shift.

I do. It's just not a general way to bypass the Moore law slowdown because there are domains that are inherently non parallelizable, and there GPUs don't do much.

Besides, I'll already covered GPUs in a sense when I talked about the need to make software to work well in many cores. Part of the work for that translates to GPU computing too.

But there's no general "shift" from CPU to GPU computing, as in "here's a new paradigm that will solve the Moore slowdown". It's, and will remain, more of a special purpose thing, for graphics, rendering, number crunching and yes, neural networks.

>So GPUs are certainly not a diversion CPU makers get into when they can't make progress on CPUs.

GPUs in general might not be, but for integrated CPUs and especially those from Intel, they are.

Reply on news.ycombinator.com