No Mo' Moore's - Time for new Architectures and Software Languages

[rcmaehl]

Source:
Institute of Electrical and Electronics Engineers (IEEE)

 

TL;DR:
We're WAY behind on Moore's law, by a factor of 15 if it was still viable. It's time to focus on optimizing software instead of hardware.

 

Media:

Quotes/Excerpts:

Quote

“We are now a factor of 15 behind where we should be if Moore’s Law were still operative. We are in the post–Moore’s Law era.”  Patterson told engineers attending the 2018 @Scale Conference. “we’re at the end of the performance scaling that we are used to. When performance doubled every 18 months, people would throw out their desktop computers that were working fine because a friend’s new computer was so much faster.” But last year, he said, “single program performance only grew 3 percent, so it’s doubling every 20 years. If you are just sitting there waiting for chips to get faster, you are going to have to wait a long time.” This is actually good news. It’s also good news for innovative software engineers. “Revolutionary new hardware architectures and new software languages, tailored to dealing with specific kinds of computing problems, are just waiting to be developed,” Rewriting Python into C gets you a 50x speedup in performance. Add in various optimization techniques and the speedup increases dramatically. The biggest area of opportunity right now for applying such new architectures and languages is machine learning. There’s a vigorous debate surrounding which type of computer architecture is best for machine learning, with many companies placing their bets.

 

My Thoughts:

We going to start seeing more and more ASIC's as time progresses. As mentioned by Linus in the video below, application specific processors and languages will increasing in popularity as clock-speeds and performance stagnates. Machine learning cores will be just one of the ASICs to come out of this stagnation.

 

[The Benjamins]

All this proves is that single core is capped out, we have to focus on multi core performance.

[Bcat00]

Not surprising.

the easy path is now used up so now the harder path must be taken.

[rcmaehl]

4 minutes ago, The Benjamins said:

All this proves is that single core is capped out, we have to focus on multi core performance.

Multicore can only be improved so much, such as we can improve how the cores talk to each other and how they share resources but not by a lot.

[vanished]

My understanding was that transistor count was what doubled.  That appears to be on track.  I know it's often quoted as performance doubling but I'm not sure it was ever intended like that.

[The Benjamins]

1 minute ago, rcmaehl said:

Multicore can only be improved so much, such as we can improve how the cores talk to each other and how they share resources but not by a lot.

Ya but the core race has only just started gaining momentum, and both Intel and AMD are working on solutions for that. 

[Linus_torvalds]

9 minutes ago, rcmaehl said:

We're WAY behind on Moore's law, by a factor of 15 if it was still viable. It's time to focus on optimizing software instead of hardware.

Tell that to MS who release windows 10. On fresh install it has 150 services, 8.1 had about 50 and 7 had about 45 in my experience and i have reinstalled those about 50 times each! On my and friends and relatives PCs. Lately i just install linux on everyones PCs.

Also Nvidia who does not make good drivers for linux and does not make them open source.

Also many other companies. For example my thinkpad t440s had dolby digital home theater software that improved sound quality very much and it's not available for linux and they will not open source it.

 

Proprietary software is very bad! Look at linux it's open source and it became so good that almos whole internet runs on linux, 90% of smartphones run on linux, smart watches run linux, supercomputers run linux and 2.5% of desktop runs on linux too.

If game developer companies were not so lazy and create more games for linux Desktop segment would run on linux too and that would be big win for humanity. Windows 10 is full of spyware and it needs fast hardware. It's using way more resources then an OS should.

 

Look at games. Those are so unoptimized that they use 6 GB VRAM, 10 GB RAM and so on. Not so long ago i played crysis 3 on ultra setting on a 2 GB GPU +2 GB RAM + i7 3770. Now many games will crash on those specs. Hell windows 10 takes almost 1.5 GB RAM just to start. My arch linux uses 170 MB RAM on fresh start and i have more usefull services running then windows 10 has.

 

So yes open source software  = win for all humanity!

[Spotty]

1 minute ago, The Benjamins said:

Ya but the core race has only just started gaining momentum, and both Intel and AMD are working on solutions for that. 

AMDs solution for increasing core count on multi core CPUs is the same now as it was 10 years ago.
 

Spoiler

[Crunchy Dragon]

4 minutes ago, Ryan_Vickers said:

My understanding was that transistor count was what doubled.  That appears to be on track.  I know it's often quoted as performance doubling but I'm not sure it was ever intended like that.

We're well past the days of performance improving exponentially between architectures, as far as I'm aware, that hasn't happened since maybe Sandy Bridge at latest.

 

The transistor count was what doubled, and that is on track according to the graph.

[Rune]

amdcorescores.jpg

[rcmaehl]

5 minutes ago, Ryan_Vickers said:

My understanding was that transistor count was what doubled.  That appears to be on track.  I know it's often quoted as performance doubling but I'm not sure it was ever intended like that.

It originally was performance doubling with the transistors but that died around late 2000s and people have been hoping we'd recover, we haven't.

[The Benjamins]

6 minutes ago, Spotty said:

AMDs solution for increasing core count on multi core CPUs is the same now as it was 10 years ago.
 

  Reveal hidden contents

and AMD sells the fastest CPU currently because of it.

[Sniperfox47]

My perspecitve is that machine learning is going to get a hell of a lot more important in the coming years. There are many problems it can't reduce the complexity of, but if it can be "good enough" at tasks to have reasonable accuracy while being less serially intensive then we could see pretty substantial performance boosts. But for some kinds of tasks we're just going to have to accept that for now performance isn't going to get much higher.

 

I love the reference to Python because that's pretty misleading. Switching python to C++ isn't a guaranteed 50x performance improvement. Optimized C++ is better than optimized python but if you're precompiling the Python stuff and not using crazy inefficient functions the difference should be a fraction of that in most cases. Rust and C++ may be great from a performance perspective but their complexity is also *MUCH* higher than Python and they're simply not a solution for a lot of devs.

 

Any time you add abstraction to decrease difficulty you also sacrifice performance in the process. Making a whole new language doesn't solve that issue. You can't have the super easy for the compiler to optimize code like Rust/C++ and the high level of simplified abstraction like Python/Java/Dart.

 

Specialized hardware is great, and paired with machine learning it can be even better but that only solves so much of the issue. Not every problem can be accelerated with specialized hardware, and you can't have an infinite amount of super specialized hardware in a generalized device.

 

9 minutes ago, The Benjamins said:

Ya but the core race has only just started gaining momentum, and both Intel and AMD are working on solutions for that. 

The point is, from a technical perspective, you can't just keep strapping more cores on. And each additional core has diminishing returns. You can only make most problems so serial before they cap out. There's a *lot* of things you can't parallelize very well. And a lot of the things you can parallelize can't be done massively so.

[jagdtigger]

29 minutes ago, Spotty said:

AMDs solution for increasing core count on multi core CPUs is the same now as it was 10 years ago.
 

  Reveal hidden contents

And it is so bad that intel didnt panic at all when ryzen came out.... /s

Edited September 21, 2018 by jagdtigger

[HunterAP]

I agree with The Benjamins, this seems to point more to an end of single core improvements and a need for an increase in core count.
Also the fact that power usage has stayed relatively the same in the past 20 years speaks a lot: we're using slightly more power for much higher performance, even though the frequencies are not really changing all that much.

 

I do agree we need an improvement in software development (looking at any coding language that's fully interpreted or partially compiled and then interpreted). We mostly need improved Assembly code and such, which x86 and x86_64 have both been THE standard for a long time. Even ARM is hitting a physical hardware limit similar to x86 but their software is quickly catching up to x86 performance.

The other problem is that having people use ASIC's for each of their different devices means that modularity of different machines will take a hit: a CPU that fits in a specific socket means that the ASIC version would basically be limiting the number of options for a system to have. This would be HELL for developers since now a company has to develop multiple architectures in terms of hardware, build separate microcode for each one, and make it monetarily feasible to do all of this, and even then they have to compete with the likes of x86 and ARM architectures.

Even after all this, software languages have to adapt to the new architectures. It's be hard to convince the likes of Java, Python, C#, or even C and C++ to make their own compilers/interpreters for new architectures. And now you say "we need new software languages anyways": you're right, but that also takes an immense amount of effort, which is why you don't just see new languages come out of nowhere, you do get some of them but they never hit a huge market share. For most languages, or at least any that are based on C, are SOL until C itself is written to work on the new architecture, then have to wait for the C-based languages to follow suit, which could take so much time that I'd be in my 40's (I'm in my early 20's now for reference).

[Curufinwe_wins]

Amdahl's law everyone...

[Deus Voltage]

Linux presents to the casual and enthusiast pc communities a paradox that I don't see anyone discussing in great depth.

 

One of the reasons why Linux is so great is because you have far less risk of having a virus than running a windows machine. 

If you want to feel extra secure, just download clam (or clam tk for graphical interface).

 

Linux destros also tend to be far lighter, another big plus. 

 

The problem (the paradox), as I alluded in the beginning of my comment, is that if Linux becomes sufficiently popular, and enough people migrate to it, then more viruses will be made that tailor to Linux destros. 

 

So on the one hand, Linux is great, we love its open source/ light/ and almost virus free nature. But at the same time, we risk losing some Linux's benefits if it is no longer as exclusive. 

 

Maybe I'm rambling, I don't know.

 

 

[yian88]

So going by the article/topic  we should start with throwing out windows since its an overbloated piece of shit? yet no one is doing that, lets improve software and programming languages? good luck getting people to use linux and get rid of C++, traditional silicon chips and architectures will  be long dead and you still wont be able to get people to quit windows and C++ garbage.

 

 

[TechyBen]

1 hour ago, The Benjamins said:

All this proves is that single core is capped out, we have to focus on multi core performance.

Ballistic processing is my bet...  https://aip.scitation.org/doi/10.1063/1.4943000

 

Though optical, single electron (https://aip.scitation.org/doi/full/10.1063/1.2888164) or other electron tunneling techniques could out pace it's development. https://en.wikipedia.org/wiki/Nanoscale_vacuum-channel_transistor

 

These will at least lower the power cost, and in theory could then increase the density... no idea if they would increase the speed, or if they would cost a LOT more.

[jagdtigger]

17 minutes ago, Deus Voltage said:

The problem (the paradox), as I alluded in the beginning of my comment, is that if Linux becomes sufficiently popular, and enough people migrate to it, then more viruses will be made that tailor to Linux destros. 

 

So on the one hand, Linux is great, we love its open source/ light/ and almost virus free nature. But at the same time, we risk losing some Linux's benefits if it is no longer as exclusive. 

Not necessarily, its a possibility but its far harder than on windows. For one the prompt for elevated privileges isnt a simple yes/no question, there is a built in safe software source, not to mention handy stuff like jails, or simply running the browser as nobody .

[dtaflorida]

2 hours ago, Crunchy Dragon said:

The transistor count was what doubled, and that is on track according to the graph.

I agree. I'd say the law is still in effect. The densities have doubled, the actual transistor count that is on CPUs hasn't, or is dedicated to other things like the igpu and what not.

 

I assume they just add on as much performance as they want in additional transistors(of course its more complicated that that) rather than fill all that extra space they've gained on a wafer when they shrink the process. Hence we have very controlled gain in performance. Just enough to be an upgrade, but not so much that they can't entice you to upgrade again in a year or two.

 

See die size shrink at each new node then increase to add more processing power prior to the next process iteration.

https://www.anandtech.com/show/7003/the-haswell-review-intel-core-i74770k-i54560k-tested/5

 

When I hear/see someone say Moore's law is dead, I assume they're trying to get attention. Shock/provoke, then pitch your spiel. That's not to say Patterson isn't right about time for new architectures and languages. 

[BuckGup]

I'll say it again. Devs have gotten sloppy and lazy because the hardware is powerful enough that they can

[pas008]

Moving away from silicon?

Many have been making graphene break throughs but nothing concrete yet

[leadeater]

4 hours ago, HunterAP said:

Also the fact that power usage has stayed relatively the same in the past 20 years speaks a lot: we're using slightly more power for much higher performance, even though the frequencies are not really changing all that much

That power usage is intentional though, it isn't so much a feat of engineering and technology that we have been able to stick to the same power usage but an actual design and technology limitation that they are unable or unwilling to exceed, mostly both. There's only so much heat, therefore power, that materials can take before they break down.

[leadeater]

4 hours ago, HunterAP said:

The other problem is that having people use ASIC's for each of their different devices means that modularity of different machines will take a hit: a CPU that fits in a specific socket means that the ASIC version would basically be limiting the number of options for a system to have. This would be HELL for developers since now a company has to develop multiple architectures in terms of hardware, build separate microcode for each one, and make it monetarily feasible to do all of this, and even then they have to compete with the likes of x86 and ARM architectures.

The other problem I see with more ASICs is memory and data flows, the big advantage to a big powerful CPU with a lot of instruction sets is everything stays in one place all with access to the same memory controlled by the same memory controller. Once you start moving things off the CPU to ASICs you have to do protocol translations to get data, memory in reality, to the ASIC which has a big overhead along with hitting other limitations like bus bandwidths.

 

That's why NVLink was developed, it addresses one part of the problem which is bus bandwidth but it still doesn't address the translation issue.

 

For there to be a big uptake in ASICs there will need to be a big shift in how hardware and memory is handled within computers, Gen-Z is one of those things I have been following with a lot of interest that seeks to solve this.

 

The focus has always been processor centric, what the source is talking about and people generally have conversations about when we need to start thinking more memory centric and start trying to remove the closed off silos of resources all talking their own language (hardware/memory wise).

https://www.nextplatform.com/2017/11/21/breaking-memory-free-compute-gen-z/

https://www.nextplatform.com/2018/05/17/forging-composable-infrastructure-for-memory-centric-systems/

https://www.nextplatform.com/2018/02/15/gen-z-interconnect-ready-restore-compute-memory-balance/