I wonder if it is time for Intel to rethink Tick-Tock

[Hugo Drax]

It seems every new CPU now only offers a 5-10% improvement over the last. And having to design/release a new chip and all the associated work involved does not seem to make sense anymore.

Business and consumers do not upgrade every year.

So wont it make sense from a cost/profit perspective and for simplicity to just release a new CPU every 4 years?

 

[SiliconWars]

It seems every new CPU now only offers a 5-10% improvement over the last. And having to design/release a new chip and all the associated work involved does not seem to make sense anymore.

Business and consumers do not upgrade every year.

So wont it make sense from a cost/profit perspective and for simplicity to just release a new CPU every 4 years?

The problem is, the industry changes so fast that 4 years is like a lifetime. It is now almost pointless trying to design a CPU for where you think the industry will be in 4 years time.

The lack of improvements are down to a lack of demand. If more consumers wanted faster CPU's and were willing to pay for it, Intel would provide them with that. As it is, they aren't. Sure there are plenty of us here who want it, but in the grand scheme of Intel's cash cow we're nothing to them.

 

[cytg111]

Heh, no, tick-tock is, if im guesstimating correctly, intels biggets achievement to date, they are effectively - not theoretically - two generations ahead of competitors .. that goes hand in hand with the chips they design. Sorry to say, there is, atm, no equal.

 

[Cerb]

Yup, no reason to ditch it, and plenty of reason to actually put similar R&D money into Atom, the same way they've been doing tick tock with the fat CPUs.

Fundamentally, tick-tock is an ongoing R&D policy. Rather than starting a big design, and then finishing it, they decide where to go next, and freeze some design goals at some point, and get that done, while continuing to improve it for gen n+1, n+2, n+3, n+4, and so on, internally. Then, next year, repeat the same thing. While it takes a few years to finish a CPU, that way they won't be too far out of touch with the market, like they were with the P4s. It may take 3 years to adapt, but as long they're internally working on a few different branches, they can choose the most compelling features to finalize, then refine.

Also, if they only released new CPUs every several years, business would only want to upgrade shortly after the new CPUs came out, which would be devastating to the rest of the PC market*, which would become insanely conservative at that point. It would also drive Intel to act insane, like RAM companies, pumping out massive quanitities, then rationing them while they update their fabs.

* PC = runs standard Windows or x86 Linux, and isn't branded Apple

 

[jones377]

In fact, Intel is moving to Tick-Tock into the Atom space, where they really need it right now.

 

[Zodiark1593]

The Tick-Tock cycle also serves to separate node shrinks and Arch changes. I believe at one point, it was said that introducing Arch changes and a node shrink at the same time is more likely to introduce issues than doing one and then the other in succession.

 

[Hugo Drax]

Interestingly enough I heard Intel might stretch Haswell to 2014 and introduce a new desktop chip in 2015.

And broadwell being more of a low power cpu.

 

[ShintaiDK]

Look at the performance/watt. Tick-Tock is a huge success.

You are not gonna get 50% boosts. Unless its 4 years cycles or something.

And Broadwell comes to Xeons.

 

[OCGuy]

10%? The fight is in power consumption. As far as performance, they have no rival to speak of.

Efficiency is where the money is. Having the performance crown is just frosting on the cake.

 

[beginner99]

It seems every new CPU now only offers a 5-10% improvement over the last. And having to design/release a new chip and all the associated work involved does not seem to make sense anymore.

Business and consumers do not upgrade every year.

So wont it make sense from a cost/profit perspective and for simplicity to just release a new CPU every 4 years?

That is if you just look at CPU performance. If we go back 4 years there was the i7-920 and soon to be released i5-750. Comparing the i5 to the newest haswell i5-4670k:

http://anandtech.com/bench/Product/109?vs=837

We see that the i5-4670k is up to 100% faster in quiet a few benchmarks and a short look tells me it's never under 50%. Of course the higher clocks help here but we are comparing stock to stock.

But this is not the point. In the same 4 years we went from no iGPU to one that can play 3D games (albeit at low res and quality) while at the same time consuming about half the power. Yeah, TDP is not much different but at idle like when writing a forum post here, the i5-750 will just use a lot more power.

So one can hardly says there was no progress. The progress just was in a direction CPU-enthusiast don't really care about: iGPU and low power.

 

[CakeMonster]

Tick tock still makes perfect sense. The design changes are no less significant than they've been historically, even though one might get that impression from the constant whining over clockspeed and IPC.

 

[BrightCandle]

Some new architectures are going to be more radical than others. There is every reason to suspect that at some point Intel will move a research project into the mainstream processors and show a more substantial improvement in performance. But for now Intel is playing with the laws of physics and loosing. They can't get anymore clock speed out of the CPU, there are rapidly diminishing returns from cache so there is no basic way to throw transistors at the CPU and have it go faster.

Its amazing Intel gets 10% at all, it costs them billions a year to do so!

 

[cytg111]

10%? The fight is in power consumption. As far as performance, they have no rival to speak of.

Efficiency is where the money is. Having the performance crown is just frosting on the cake.

- Also, guesstimating 10 years ahead, threading software is emerging, we figured out how to parallize our workloads, more than 4 cores are in demand, 6, 10, 16 cores for the average consumer.
16 Cores running what? 300 watts? I think the fight for perf/watt is a righteous one atm, we are gonna need that tech later.

 

[ShintaiDK]

New instructions is essentially the only way to get higher IPC. Conroe and Haswell is major points there due to the faster execution of the new instructions.

 

[Idontcare]

Some new architectures are going to be more radical than others. There is every reason to suspect that at some point Intel will move a research project into the mainstream processors and show a more substantial improvement in performance. But for now Intel is playing with the laws of physics and loosing. They can't get anymore clock speed out of the CPU, there are rapidly diminishing returns from cache so there is no basic way to throw transistors at the CPU and have it go faster.

Its amazing Intel gets 10% at all, it costs them billions a year to do so!

Intel still gets more clockspeed, they just choose not to bin it and sell it.

Instead they sell lower power processors at the same price-point, increasing their parametric yields in the process (all the more profit).

 

[shady28]

OP has a good point, being taken out to an extreme by many responders though.

4 years is clearly too long. Going from 12 months to 18 months, that may already be happening. That by itself could reduce expenditures for R&D and fab investments by as much as 1/3. Given lack of a compelling case for upgrades, this makes sense for the desktop.

Mobile like bay trail, different story.

 

[jpiniero]

Still think they are (or will be forced to) since I don't think 10 nm is happening in a timely fashion.

 

[R0H1T]

Still think they are (or will be forced to) since I don't think 10 nm is happening in a timely fashion.

Yup, 14nm is probably gonna be a quarter or two behind schedule & 10nm by roughly twice as much & anyone who says that Intel can continue putting an increased amount of $$ for R&D going forward, especially with shrinking revenues & profits, is smoking some serious [stuff] & I hope that this age old adage of Intel being two nodes ahead of the rest of the pack gets corrected in due time because it certainly isn't true anymore !

Please do not use profanity in the tech forums
-ViRGE

 

[Exophase]

New instructions is essentially the only way to get higher IPC. Conroe and Haswell is major points there due to the faster execution of the new instructions.

New instructions rarely increase IPC, in fact they often decrease it (even while still increasing overall performance). I think the term you're looking for is perf/MHz, not IPC.

 

[ShintaiDK]

New instructions rarely increase IPC, in fact they often decrease it (even while still increasing overall performance). I think the term you're looking for is perf/MHz, not IPC.

New instructions+new code tends to increase IPC.

 

[Arachnotronic]

New instructions+new code tends to increase IPC.

I think the idea is that new instructions increase work per instruction, which leads to a performance/MHz improvement. It's not necessarily an instructions per clock increase.

 

[pablo87]

In value consumer, 4 years would be good enough. To wit, the quad core i5 750K launched 4 years ago. The cost to build these if the 45nm fab on which it was built still existed, would certainly allow Intel to meet or exceed their gross margin % target, and the contribution would be better than the Celerons 900's built on 22nm...

Based on a sub $100 selling price. Does Intel sell any quad cores for under $100 today?

 

[Exophase]

New instructions+new code tends to increase IPC.

No, they really don't. Think a little more carefully about what you're saying. New instructions are meant to perform more complex operations that would have previously taken several instructions to execute. So they decrease the number of instructions needed for a task. But that doesn't increase the number of instructions executed per cycle. Since new instructions often use special execution units of which there are fewer than the simpler primitive units that the previous code would utilize this means that often IPC goes down, not up.

 

[2is]

The runtime improvements in haswell laptops is reason enough not to ditch their cycle. It's not always about overclocking or huge increase in IPC. Once I can get my hands on a Retina MacBook Pro with Haswell, i'm buying.

 

[TuxDave]

The Tick-Tock cycle also serves to separate node shrinks and Arch changes. I believe at one point, it was said that introducing Arch changes and a node shrink at the same time is more likely to introduce issues than doing one and then the other in succession.

That's probably a good enough view for the average person.

Truth is, ticks and tocks both have arch changes except that the tick has much less time to finish it (therefore less scope) because ticks need to plan in additional time to spec out the new process and analyze how the design changes. At the same time tock also has the opportunity to introduce small process adjustments even if it's on the same "process generation" as the previous tick. There's not a clear division of "arch only change" and "process only change".