Ryzen' Shine ppl :Đ

i got interested into what our technical limits are atm. through this debate. so went a bit researching. I kept it within current design philosophy, so no quantum computing stuffz.

summarized my findings: physically, it seems, we won’t get much above 8 GHz  ![:)](<fileStore.core_Emoticons>/emoticons/001j.png “:)”) with electrons, but mainly atm. silicon heat treshholds in cmos limits us to ~3.6 GHz;

interestingly, lately the travel time of data on the mobo starts to affect limits aswell. doesnt matter if the cpu is fast as it gets, if the dataflow isnt - we really have to find a way to speed up light.

thankfully, thats only clock speeds - there is lots of stuff, which influences our progress in efficiency more than that. e.g. its also a matter of time until our applications make use of multicore designs more properly, more preoptimizations, more precognition stuff. and in gaming, besides sheer graphics intensitity, for me personally, networking techniques get more important as a limiting factor aswell.

why this fluff? because it kinda shows, that our current hardware is quite mature. innovation started to evolve around other things than clock speed by now. unfortunately, we can’t expect the THz cpu appearing miraculously aswell. in terms of miniaturization we will reach a treshhold soon aswell, moores law ends in 2025.

but in the year 2525, you gotta survive.

a little curiosity: in cryptography, they make use of bremermanns limit - assume a giant computer the mass of 1 earth with maximum absolutely theoretically possible computing power per mass unit - to test whether a cryptographic algorithm can be broken in time with brute force. quite funky.

Yes!

…the Ryzen is chance that multi-core sensible games software design will show up even more and solid in games to come… because Ryzen is consumer oriented reasonable price for 8 core cpu

20 hours ago, g4borg said:

i got interested into what our technical limits are atm. through this debate. so went a bit researching. I kept it within current design philosophy, so no quantum computing stuffz.

summarized my findings: physically, it seems, we won’t get much above 8 GHz  ![:)](<fileStore.core_Emoticons>/emoticons/001j.png “:)”) with electrons, but mainly atm. silicon heat treshholds in cmos limits us to ~3.6 GHz;

interestingly, lately the travel time of data on the mobo starts to affect limits aswell. doesnt matter if the cpu is fast as it gets, if the dataflow isnt - we really have to find a way to speed up light.

thankfully, thats only clock speeds - there is lots of stuff, which influences our progress in efficiency more than that. e.g. its also a matter of time until our applications make use of multicore designs more properly, more preoptimizations, more precognition stuff. and in gaming, besides sheer graphics intensitity, for me personally, networking techniques get more important as a limiting factor aswell.

why this fluff? because it kinda shows, that our current hardware is quite mature. innovation started to evolve around other things than clock speed by now. unfortunately, we can’t expect the THz cpu appearing miraculously aswell. in terms of miniaturization we will reach a treshhold soon aswell, moores law ends in 2025.

but in the year 2525, you gotta survive.

a little curiosity: in cryptography, they make use of bremermanns limit - assume a giant computer the mass of 1 earth with maximum absolutely theoretically possible computing power per mass unit - to test whether a cryptographic algorithm can be broken in time with brute force. quite funky.

It’s not only a problem about conductor’s heating/power limits. More megahartz need longer pipelines in the chip, with all the problem a long unoptimized pipeline can get (redundancy, miscalculation, failed soft requests) all stuffs that makes them doing sometimes two times the job a smaller pipeline/mhz chip can do.

In this regard we had an hystorical example: P4 willamete and athlon xp (barton).

Where the first iteration of P4 was that bad, for the reason I describe above, that the low freq, barton had an higher IPC with a lower clock speed (that’s why they basically moked Intel by calling their cpus athlon 2500+ when instead their clock speed was around 1800mhz, since they had the same IPC at that freq.), all I wrote here is history, and not: me bashing Intel. Two minute research can be done to verify this.

That’s the reason Intel left the first iteration of P4 to die fast and started to move (somehow) to the amd approach with the “core” arch. Well ok, during years branch prediction improved a lot, that’s why we started to see again high freq. but it’s always a work around to something that don’t really have a real solution if you want to achieve the higher clock speed, even if you “solve” the thermal/power limits of the material involved.

That’s why we have multiple cores arch. Sadly the software optimization always struggled since the mass market is oriented toward 2/4 cores cpus, so basically: who cares. I blame AMD for that, for too long they left that faildozzer on the market, without releasing competitives multiple cores on mass market. Let’s see what happens now.

Btw on Ryzen topic, as far as I saw, the mainboard are really struggling to achieve decent reliability and performance optimization. It’s really a bleeding edge issue we are seeing right now. Expecially Asus is acting really weird on their top brand motherboards. They already released like 3 bios update in a month, and it’s not even enough to say that they are at a good “start”.

Even like this, we are seeing something totally whorth attention (finally). 

It certainly wont be easy to match up to the new cpu with mobos but it shouldn’t take too much time either.

4 hours ago, Spongejohn said:

It’s not only a problem about conductor’s heating/power limits. More megahartz need longer pipelines in the chip, with all the problem a long unoptimized pipeline can get

I was more talking about theoretical limits - production limits after all are reducing over time, depend a lot on innovation in the production technologies anyway. the physical limit of 8 GHz for electrons is e.g. not directly related to heating or power, but quantum-physical boundaries. the 3.6 GHz on silicon e.g. is a heat limiting factor - not saying there isn’t even already alternatives with non-silicon materials, or phototronics. So those are limits under theoretical perfect conditions, where we would not have to think about “how to produce it”, just “if we could make it perfect”, and it’s just impressive, that our current cpus kinda operate at those speeds. Independent of brand ![:p](<fileStore.core_Emoticons>/emoticons/004.png “:p”)

do have to admit, have no idea, why pipelines would have to get longer for higher clock speeds, might already touch the relativistic signal speed issue? Still, I was under the impression, that is only an issue in the dimension of decimeters, still, at GHz, so might be something design related.

3 hours ago, g4borg said:

…the physical limit of 8 GHz for electrons is e.g. not directly related to heating or power, but quantum-physical boundaries. the 3.6 GHz on silicon e.g. is a heat limiting factor - …

if the material get too thin, quantum effects (‘quantum tunneling’) will counter effect memory state states or heavily influence currents…

Yes, its design related one needs more logic to feed longer pipelines at higher speeds (branch predictions…etc.pp) this causes heat and makes dies enlarge lengthen signal paths and overall resulting in higher prices because of higher rate of spoiled wavers… also

On 5/3/2017 at 0:34 PM, g4borg said:

do have to admit, have no idea, why pipelines would have to get longer for higher clock speeds, might already touch the relativistic signal speed issue? Still, I was under the impression, that is only an issue in the dimension of decimeters, still, at GHz, so might be something design related.

In a chip the performance is bounded to the slowest stage*time execution. If you have a stage that take longer time then others, the other hardware unit have to litteraly: wait.

That’s why you have multiple overcutted stages and parallel pipelines. Instructions are sent trough those like a flow to let all the hardware unit take all the time they need to execute stages, without waiting times. Since other instructions can be sent to others hardware units in a constant flow.

Plus: those stages can be “cutted” into multiple executions flows to override the “waaaaait i didn’t finished” situation that would block everything and cause a “flush” if anything goes wrong.

Only with overcutted stages*instructions you can boost the clock from hardware units and get advantages from it. Otherwise it is useless. 

Also the problem is, during an execution of a stage, if this is cutted in substages on multiple hardware units, it’s really easy that an error can occours: so you have to flush all the pipelines, not only the one that is executing the instructions. Causing a dramatic loss in troughput.

There are no real solution to this apart the overcomplicated branch predictions algorithms (or you know… Magic) they developed in order to fullfil hardware units in the most reliable and fastest way  to prevent critical “flush” errors and a consequential loss of troughput.

 

Sorry for the oversymplified explanation that I’m sure some tech guy will destroy 'cause is pretty dumbed down. Tech english is not easy for me, and I don’t have a degree on that ![:p](<fileStore.core_Emoticons>/emoticons/004.png “:p”)

 

Edit: let me try again, you will have speed improvements from higher clocks only if you let the instructions flow trough multiple stages instead of less stage for pipeline. Otherwise you will be bounded (and wait for) to the slowest hardware unit that have to process the “bigger” stages. Litterally, making an higher clock for hardware unit is useless, since they have to wait for the slowest one to process his stage.

The first remarkable crisis came as early as the introduction of the very high speed intel Pentium 4 with up to 4 GHz, in them the pipelines were significantly longer, but were designed in length due to loss of throughput, if brach predicting algorithms gave wrong prediction and the total pipeline had to be emptied and filled with new stuff (costing cycles to dumb pipeline content)

With all the major CPU producers, even very big like the z13  5 GHz octa-core processor from IBM, AMD and intel, 4-5GHz seems to be the sweet spot for modern cpu architectures with roomtemperature operation and mostly air cooling…

3 hours ago, avarshina said:

The first remarkable crisis came as early as the introduction of the very high speed intel Pentium 4 with up to 4 GHz, in them the pipelines were significantly longer, but were designed in length due to loss of throughput, if brach predicting algorithms gave wrong prediction and the total pipeline had to be emptied and filled with new stuff (costing cycles to dumb pipeline content)

With all the major CPU producers, even very big like the z13  5 GHz octa-core processor from IBM, AMD and intel, 4-5GHz seems to be the sweet spot for modern cpu architectures with roomtemperature operation and mostly air cooling…

Yep, you always need to balance mHz, pipeline’s stages and tdp according to the nm process.

Prediction algorithms are only a “patch” to this, that can only works to a given extent. 

That’s why we need to go multicore to see real perfomance inmprovements on CPUs. 'Cause, aside from “semi conductors” limits, the actual Cpu arch have some limits that is not easy at all to deal with.

So we have a sweet spots for the arch around 4-5ghz (i’ll tell you, going around 4ghz for both amd and intel on 8 core is not an easy task at all), but the software is stil behind. For example, Windows scheduler use cores like a baby eats  candies… the more the merrier, but both Intel and AMD arch suffer from microsoft scheduler weirdness when you have more than 4 cores with HT or SMT.

This doesn’t means we all should became AMD fan 'cause they pushed long time ago for multicore arch (and also pushed some others improvements to the mass market in order “to sell”), since Intel also introduced some interesting stuffs like avx2 and others dedicated instructions that can boost some tasks a lot. But in the end, those are marginal improvements, to be honest. We are basically bounded, speaking about CPU ark and software development, to a model developed 20 years ago… And it’s damn sad that tech improvements are slowed down just 'cause of “market reasons”. 

“You know, all you need is a dual 5ghz cpu and you will run everything nowadays”… a lot of people thinks like that even amongst software developers… that’s just damn sad. 

Like if in Star trek they stopped developing warp’s engine, since, you know… who needs more than warp 7 nowadays…