CPU improvements year to year

[tynopik]

Let's also assume that the light pulses are generated in the centers of the cores, and have to propagate to the farthest corners, about 4mm (4*10^-3 m) away.

But one signal doesn't have to propagate across the entire chip

you can take the clock signal and route it to multiple entrance points on the chip

as long as you keep all the paths from the clock the same length, they will stay exactly in sync

 

[Gikaseixas]

Thanks for all the inputs.
I still think that Intel could further improve IPC if pushed by a competitor but I could be wrong. Higher mhz would have an impact in power efficiency I think, but what about higher IPC? Does it affects the power profile of a chip?

 

[jj109]

Thanks for all the inputs.
I still think that Intel could further improve IPC if pushed by a competitor but I could be wrong. Higher mhz would have an impact in power efficiency I think, but what about higher IPC? Does it affects the power profile of a chip?

Depends on how that IPC is achieved...

For example if the branch predictor's accuracy is improved without adding a ton of transistors, it CPU could have higher IPC and use less energy since less cycles are wasted chasing the wrong branch.

Otherwise, if the core is made wider with everything else held constant, IPC would go up for workloads with high instruction-level parallelism and power will go up as well.

 

[AtenRa]

Yes. http://www8.hp.com/us/en/products/proliant-servers/product-detail.html?oid=5375897#!tab=features

I'm under NDA to discuss what's coming next, but I can tell you that their power limit is 100w per server.

You can have a big core with lower power, one doesnt invalidate the other. And yes we already know that both AMD and Intel are already using low power small core SKUs for server use like Jaguar and ATOM. That doesnt mean they will switch to small cores for the entire enterprise segment.

 

[BrightCandle]

Most of the improvements we saw before came from clock speed improvements. The hey days going from DX4's through to the Pentium 4 was mostly about clock speed improvements combined with instruction level improvements. Now we are only getting the instruction improvements its not a lot to go on. You can spend enormous amounts of design and research time to get pretty marginal improvements here, the only answer to which is going wider on the numbers of cores and having the software to use it (not an easy task).

 

[Phynaz]

You can have a big core with lower power, one doesnt invalidate the other. And yes we already know that both AMD and Intel are already using low power small core SKUs for server use like Jaguar and ATOM. That doesnt mean they will switch to small cores for the entire enterprise segment.

Did I say they would?

I see you can still buy some very nice buggy whips too.

 

[TuxDave]

Researchers have developed transistors with speeds close to 1 THz, so 200 times faster than the transistors in Intels CPU's. Of course a lot of work needed to reach these speeds in a CPU, but that is why we need competition.

Hopefully this analogy helps:

News Statement:
"Scientists develop rubber for a tire that can operate at ground speeds of 450mph!"

How people are taking it:
"If we put those tires on our car, we too can go 450mph!"

That's basically how I'm seeing the general reaction to IBM's news on the unity gain frequency of a transistor.

Note: The spec IS a real spec, just like tire ground speed is a REAL spec and has some REAL applications, but people are just taking words that they are familiar with and using the numbers in places that it doesn't belong

 

[ShintaiDK]

It has to be said tho, that IBM is masters of spinning R&D results. Elsewhere "irrelevant" R&D results gets frontpage news. And people get the illusion that IBM is this forerunner of technology. While in reality....

 

[Homeles]

Intel, AMD, and IBM also have had 2+ THz transistors, which they announced back in the early 2000's. Back then, Intel was claiming that they'd actually be using them in the not too far future, but obviously frequencies haven't scaled nearly as well as they had hoped.

In fact, Intel's "TeraHertz" transistor design they were pumping back then is pretty damn close to what they were using at 32 and 45nm, except they were fabbed on SOI.

The 3 major features were use of a high-K dielectric (zirconium dioxide -- Intel and practically everyone else ended up opting for the similar hafnium dioxide), a raised source and drain (Intel's 32nm featured this, and FinFETs are, in part, an extreme version of a raised source and drain), and SOI, which Intel has not adopted, citing cost concerns. Even with SOI, we'd not be terribly better off than we are now.

 

[iamgenius]

Interesting thread.....I'm following it!


So, we nearly reached the limit when it comes to Clock speed. I remember from university days back in 1999/2000 my instructor telling me that Intel & AMD will hit a wall after reaching 1000 MHz (1 GHz).

 

[Lepton87]

Interesting thread.....I'm following it!


So, we nearly reached the limit when it comes to Clock speed. I remember from university days back in 1999/2000 my instructor telling me that Intel & AMD will hit a wall after reaching 1000 MHz (1 GHz).

He must have been a rather stupid instructor. Because in March 6, 2000 AMD released Athlon clocked at 1GHz. How can you have any trust in the education system if such idiots teach.

 

[coolpurplefan]

This may sound like an odd question, but if people say GPUs are progressing faster than CPUs, could Microsoft make an OS for a GPU?

 

[Homeles]

This may sound like an odd question, but if people say GPUs are progressing faster than CPUs, could Microsoft make an OS for a GPU?

Theoretically... you probably could. It'd run beyond terrible, though, and that's something that will likely never change.

 

[witeken]

Would Xeon Phis be any different?

 

[NTMBK]

Would Xeon Phis be any different?

The Xeon Phi already runs a Linux OS internally. The PCIe card is basically a computer within a computer, in terms of how the software works. And socketed, standalone Xeon Phi is coming with Knight's Landing.

 

[lamedude]

So, we nearly reached the limit when it comes to Clock speed. I remember from university days back in 1999/2000 my instructor telling me that Intel & AMD will hit a wall after reaching 1000 MHz (1 GHz).

I got a kick out of this. The following predications were made in 89.

At the Microprossor Forum, several conference speakers predict that processor clock speeds will not exceed about 50 MHz.
John Crawford, chief architect of the Intel 80486 processor, predicts that in 1999, Intel will release a processor incorporating 50 million transistors, and be 386 compatible.

One group couldn't be more wrong while the other was close (CuMine was 21m transistors).

 

[teejee]

I got a kick out of this. The following predications were made in 89.
One group couldn't be more wrong while the other was close (CuMine was 21m transistors).

yes, I'm certain we will see much higher frequencies than 5Ghz (by order of magnitude). Of course there are obstacles to overcome, but I cannot see why it should be impossible.

 

[Revolution 11]

yes, I'm certain we will see much higher frequencies than 5Ghz (by order of magnitude). Of course there are obstacles to overcome, but I cannot see why it should be impossible.

You can just look at historical trends to see how this will be hard (higher than 5Ghz by order of magnitude). When they predicted the 50Mhz limit, the last few years still saw the traditional Moore's law scaling.

Compared to today, where Dennard scaling failed sometime between 2000-2003 and clockspeed increases shortly thereafter, you can assume with some certainty that production CPUs will not see 50Ghz in a time scale less than 10 years.

 

[teejee]

You can just look at historical trends to see how this will be hard (higher than 5Ghz by order of magnitude). When they predicted the 50Mhz limit, the last few years still saw the traditional Moore's law scaling.

Compared to today, where Dennard scaling failed sometime between 2000-2003 and clockspeed increases shortly thereafter, you can assume with some certainty that production CPUs will not see 50Ghz in a time scale less than 10 years.

50Ghz in normal PC CPU's will definitely take more than 10 years. But I think we'll get there. But more like 20 years (wild guess).
And remember that Moores law is not a law by any means, just an simple observation. It won't predict any disruptive technologies.

 

[Homeles]

50Ghz in normal PC CPU's will definitely take more than 10 years. But I think we'll get there. But more like 20 years (wild guess).

Nothing is set in stone. There are far too many factors that affect advancement. Political factors and economic factors are hugely important things that you can't account for.

 

[Revolution 11]

50Ghz in normal PC CPU's will definitely take more than 10 years. But I think we'll get there. But more like 20 years (wild guess).
And remember that Moores law is not a law by any means, just an simple observation. It won't predict any disruptive technologies.

First, I was being conservative with my statement. One thing I have learned is that the impossible is actually just hard to envision, not impossible.

Second, I don't think we will get anywhere close to 50Ghz ever using architectures similar to what Intel and AMD and ARM are offering. Remember, if you could scale Sandy/Ivy/Haswell to 50Ghz, that is a 10x improvement. Note that almost all of our performance improvements since basically 2003 have come from more cores and more IPC, not higher clocks. Clocks today are still the same level as the Pentium 4, so my hopes are not high.

 

[witeken]

No [duh] Sherlock, that's why we will move on from silicon and CMOS. Silicon can't even go to 50GHz.

No profanity in the technical forums, please
-ViRGE

 

[Ajay]

Well, if we do ever get to a 50 GHz CPU, bless the poor engineers who will be working the clock trees (or whatever will be needed at that speed).

 

[JTsyo]

It's not just the lack of competition but lack of application. How many process are that CPU is the bottleneck?

 

[sm625]

I think real competition is going to come from the integrated FPGA. Just wait until someone writes a custom graphics engine that runs so efficiently that it actually outputs better graphics than a 150W GPU.