when will single core performance be improved? why do they insist

No, 1024 cores

>>60788074
ayy

>>60788056
Significantly?
Never
The single core train is over.
Increasing throughput at this point requires exponential complexity. Clocks are near the limits allowed by physics.

>>60788123
also we're getting pretty close to hitting the end of moore's law

3-D gates will be the last card. until we knot the circuits as tightly as we can (think 3-D maze), there won't be much more shrinking to do.

>>60788056
Because it is very difficult

>>60788056
Just do less with your machine. Like stop using gui's. And do sequential tasks only so you won't be haunted by the benefits that multithreading would provide.

>>60788056
IBM has a 2000 core CPU.

Is there anything new on the materials science front that will allow single core to be great?

File: OHr577N-.jpg (55KB, 400x400px) Image search: [Google]

55KB, 400x400px

>>60788056
When we start using something other than silicon, which is never. But the big customers in future will be data collecting agencies handling large amount of data which can be massively parallel, so they need more cores rather than IPC. So the future will be more cores.

Oh, we'll all jump on the single-core performance train as soon as you discover a new paradigm.
Come on anon! We're waiting!

>>60788208
how come cores shrink to get faster anyway?

File: 1496559490938.png (153KB, 924x1035px) Image search: [Google]

153KB, 924x1035px

>>60788801
I could run Monte Carlo integration w/o a GUI like you said and it would still take a while in a single-core, single thread system. As it has been said, we're reaching practical limits on what conventional electronics will allow, with no future in sight for optical computing to become practical (which affords tangible speed benefits in conventional problems a quantum computer could not), and yet instead of accepting this and dealing with the more tractable but still balls-hard issue of using parallelism cleverly in more day-to-day tasks the expectations in the market are planned obsolescence and code bloat. If baby's first computationally heavy numerical method I mentioned above sets a standard, as is, for what can be expected in non-hobbyist settings, the horrible programming practices I describe later make sure all we have to look forward to is disappointment

>>60788056
single core performance is very good right now. in fact it's been great for a long time, just slowed down by things like Disk I/O. However, it's hard to make CPUs that can handle this complexity of application switching while also managing contexts of waiting on things from other hardware components. Context switching between processes is still somewhat expensive, so having multiple cores can improve speed by a factor or 4 or 8 just because it can offload the busy work somewhere else and still dedicate a lot of cpu to whatever youre directly interacting with

>>60788897
I just realized I said 'do sequential tasks only' but I meant tasks that would not benefit from parallelism.

Apparently there is a factor with how numbers are rounded that does terrible things in multithreaded operations. Found the info a year or so ago and forgot about it until now.

I'm more interested in machines without impossible logic states so if a gate or switch etc. is tripped by a small amount of cosmic energy, it will amount to no more than an error and won't hang the machine.

>>60788875
Yes, graphene https://en.wikipedia.org/wiki/Graphene#Electronic
>>60788885
See ^

>>60788123
It is possible to increase single core performance even on silicon and on the same 4-5Ghz bottleneck. It is called out of order execution. All modern CPUs are already using it. Intel can do about 7.5 operations per clock on average, Arm like 2.5 last time I checked. There is a Russian CPU which can do 21 operations per clock.

how will they power the telemetry and spyware without slowing down your computer tho

>>60788056
Dumb fucking frog poster

File: file.png (111KB, 1277x716px) Image search: [Google]

111KB, 1277x716px

>>60788208

Dennard scaling died around 2006 when we realized 5 GHz and higher clocked CPUs are not practical due to the amount of power and cooling needed to dissipate all the heat generated. Even architectural processor design IPC changes are coming to an end and although we can do all sorts of stuff like change semiconductor materials, or optical computing which might give some more increases. We as a result have to think up other way to get speeds like multicore processing. But even that has a limit, as we can only speed up a certain percent of most software as Amdahl's law says. Most of the exciting gains are now on the memory front since there are a ton of new memory technologies like Xpoint, ReRAM, MRAM (which I am vouching for), FeRAM, and NanoRAM. All of them competing to be the one universal memory instead of a memory hierarchy based upon memory speed and density which we have now.

What's this obsession with single core performance?

CPU manufacturers weren't able to improve IPC significantly, but now that amd is competing they are both improving in multicore.

Are you a lazy dev who's refusing to paralellize? (Close to all devs have no excuse)

>>60790360
Vast majority of tasks is serial. Most clever algorithms are serial, even those that aren't scale horribly. Higher-level languages introduce serial overhead.

>>60788056
When FETs get replaced.
https://www.extremetech.com/extreme/185027-the-vacuum-tube-strikes-back-nasas-tiny-460ghz-vacuum-transistor-that-could-one-day-replace-silicon-fets

>>60788056

When there is some advance in cooling, maybe integrated heat pipes within the CPU die idk im not a physicist.

>>60788056
MUH SINGLECORE MOAR GIGAHURTZ LMAO

The Playstation 4 has incredibly nice graphics for it's shitty CPU by utilizing 8 GB of GDDR5 memory with 218GB/s throughput. For comparison DDR3 memory rated at 1333MHz (see: sandy bridge controllers) has a throughput rate of 10GB/s.

Now would you look at that.

>>60790575
Wrong thread?

File: moore_is_kill.png (97KB, 1025x857px) Image search: [Google]

97KB, 1025x857px

In 20 years.

>>60790360
>Are you a lazy dev who's refusing to paralellize? (Close to all devs have no excuse)

Problem solved. Just shame devs to "paralell-ize" everything, because it's just easy, and cost effective.

>>60790624
"Hyperthreading™" is SMT

>>60791484
>CPU
>graphics

>>60788875
Yes, graphene and with graphene we can have CPU on 1 core to 500ghz

>>60789784
Do you have more info over the Russian processors?

>>60790440
Yet another "will replace silicon one day" meme.

>>60791553
>Tech sample costing thousands with a single transistor at 50ghz

>Economically viable processors with billions of transistors running at 4Ghz

Choose 1

>>60791611
IKR, Can't wait to dig up all those old stanford and harvard articles about doubling battery life.

Sadly the demon of decreasing returns tells us that battery density will double, at its best, every 15 years.

With processors it WAS every 18 months.

File: 07OLVacuumtransistors-1403115198821.jpg (100KB, 620x450px)

100KB, 620x450px

>>60791611
Materials and technology are literally same.

>>60791581
https://en.wikipedia.org/wiki/Elbrus-8S
it is this processor, but there is no info on instructions per clock in there

>>60791735
From that I see, it's basically Itanium-like VLIW, it executes 64-bytes wide VLIW instruction packs, each consisting of up to 20 instructions. I guess it doesn't work in the real world tho because data dependencies make it impossible to effectively use all the instruction lanes at the same time, just like Itanium.

>>60791735
Cont.
https://en.wikipedia.org/wiki/Elbrus_2000
Here it says 20 instructions per clock, but I've read 21 somewhere else.

>>60791735
>>60791775
VLIW means explicit instruction-level parallelism, it's like combining multiple cores in one except way less flexible and way harder to implement out-of-order execution. IPC looks inflated because of that. tl;dr it's shit.

>>60788123
>Clocks are near the limits allowed by physics.
We need to change exotic semiconductors like gallium arsinide or molybdenum disulfide

Maybe clockless CPUs will make a comeback?

>>60791499
SMT
Simultaneous Multi Threading
>>>/v/

>>60792220
So discussing fucking graphics and MULTIthreading in a thread about single-core performance isn't /v/ , huh? I say >>>/v/ yourself.

>>60788896
Less distance to travel for the currents (not the electrons, because they barely move)

>>60788056
Why is this a problem anon? Our brains are "multi core" as well.

File: 1495093750862.jpg (35KB, 640x640px) Image search: [Google]

35KB, 640x640px

>>60791499
what the fuck
frogposters should be perma-banned

>>60792110
I'd love a CPU made of thinkpadium gentooinide

The human brain is the pinnacle of computing power.

It contains millions of billions of cores consuming milliwatts of power at 100hz with millions of sensors.

>>60793306
It's very unreliable and bad at math tho.

>>60793360

The hazards of biological components.

>>60792510
Are you retarded?
What do you think is electric current?
It's the amount of electrons that move on let's say a wire per second.

>>60793360
it's cuz it's busy keeping us alive. imagine what it could do if it wasn't held back by a useless lump of meat

>>60793413
electrons barely move in an electrical current, and i mean they move on orders of nanometeres.

https://en.wikipedia.org/wiki/Drift_velocity

excerpt:
Therefore in this wire the electrons are flowing at the rate of 23 µm/s. At 60 Hz alternating current, this means that within half a cycle the electrons drift less than 0.2 μm. In other words, electrons flowing across the contact point in a switch will never actually leave the switch.

>>60793306
imagine if we could figure out a way to harvest brainpower to do computing for us

like, every smartphone would have a small brain in a jar connected to its motherboard

just like in Technobabylon

>>60793446
But it itself is just a lump, mostly of fat.
>imagine what it could do
Nothing, because it evolved over millions of years to do this one job. There's no hidden supercomputer inside because it would be extremely inefficient waste of space and energy from evolutional POV.

>>60793488
brains are too much of an unreliable computer

sure, brains are extremely energy efficient, but they are not a counting machine. the operations brains do are much closer to fuzzy logic than actually being deterministic.

imho because of this they would make horrible numerical computing hardware.