Is CPU technology topping out?

[ShintaiDK]

Yes, but how much? Perhaps 15-20% performance increase? Then another Tick next CPU generation that improves performance by 5%? Should I get excited? :\

15-20% in average with current software.

Taking in account SB+IB+HW improvements during 3 years. They lowered TDP by 20%+ and increased CPU performance around 65%. Not taking about iGPU or other cost savings. Plus CPU prices if around half of what it was previously.

In the same timeframe you went from a singlecore 2.4Ghz P4 to a 3.2Ghz dualcore with a 135% higher TDP.

That turns us back to the point again with you wanted more cores without cost. Looking back then you should pay 500$ for a dualcore CPU.

500$ roughly translates into a 3930K. Meaning you get exactly the same as back then.

You are just not "excited" enough to pay the same price.

AMD didnt have better prices either:
http://techreport.com/articles.x/8295

537-1001$ for a dualcore. Are you willing to pay that for an octocore?

 

[Fjodor2001]

Taking in account SB+IB+HW improvements during 3 years. They lowered TDP by 20%+ and increased CPU performance around 65%. Not taking about iGPU or other cost savings. Plus CPU prices if around half of what it was previously.

We don't know what Haswell will bring. The latest Tick/Tock cycle (SB+IB) brought around 30% CPU performance increase in about 2.5 years.

In the same timeframe you went from a singlecore 2.4Ghz P4 to a 3.2Ghz dualcore with a 135% higher TDP.

Depends on what timeframe you compare with.

We also went from:

Pentium II 233 Mhz (May 7, 1997)
Pentium II 450 Mhz (August 24, 1998)
Pentium III 800 MHz (December 20, 1999)
Pentium 4 2.0 Ghz (August 27, 2001)

I.e. the CPU performance increased by almost 1000% in about 4 years!

And later we also went from P4 -> C2D in one CPU generation, reducing power consumption by around 30-50%, and more than doubling CPU performance... :biggrin:

That turns us back to the point again with you wanted more cores without cost.

Which again is something I've never said. Can you please stop destroying all threads with personal insults and lies?

This has nothing to do with what me or any other specific individual wants to pay for a specific CPU. We're talking general principles here.

Looking back then you should pay 500$ for a dualcore CPU.

500$ roughly translates into a 3930K. Meaning you get exactly the same as back then.

You are just not "excited" enough to pay the same price.

AMD didnt have better prices either:
http://techreport.com/articles.x/8295

537-1001$ for a dualcore. Are you willing to pay that for an octocore?

See this post by AtenRa - and that post discusses current 22 nm technology. One node shrink from now (i.e. Broadwell), an 8 core CPU at $300 should clearly be possible. If Intel wants to release such a CPU that is.

 

[pelov]

AMD didnt have better prices either:
http://techreport.com/articles.x/8295

537-1001$ for a dualcore. Are you willing to pay that for an octocore?

Back then the benefits were at least tangible. There was a noticeable difference between an Athlon64 and a Pentium II or 1400+ Thunderbird. Now your average user won't notice a difference between a C2Duo and a 3960X.

 

[piasabird]

Well to put this into contrast, It is a kind of consensus that the present Ivy Bridge Processors run 10-20 degrees hotter than they should be running with the proper kind of cpu package. However, it looks like Intel got cheap and cut some corners to save money.

However, in the past they thought Intel had hit an obstacle with P4 processors as they were running hot. Computers can be sped up to run faster and more efficiently in different ways. One way is to design better hard drive or storage technology. Reads and writes can slow down a processor. The bus on the motherboard maybe could be improved also to speed up a processor. They say we have a 64 bit processor but is it really running at 64 bit speed on a 32 bit bus?

I think there are still some ways to make a processor run faster, through better CPU design. I think it might also be possible to design processors to be better for specific tasks, like a processor designed to enhance graphics processing or encoding and decoding or some other task like better arithmatic algorithms.

CPU's are not the only thing that could be changed or altered. We could get rid of all DVD and Blu-Rays and just use some kind of flash drives to replace that. I think we need to dump all optical drives. They only reason to keep using them is due to archaic movie and music rights legal issues. As a country we need to come up with some better solutions and demand some improvements in storage size capabilities. If we had smaller (physical size) storage devices we could drastically shrink the size of computers. W dont really need a big box computer. So the only other thing holding us back is the size of the video cards.

 

[piasabird]

So you think it is possible to speed things up with a 128 bit processor?

 

[piasabird]

What about a more object oriented operating system?

 

[moonbogg]

It may be topping out for Intel, but aMd has lots of room to improve still.

 

[TuxDave]

Possibly not a shortage of ideas. But the OP wanted to know if the raw CPU performance has topped out (i.e. not taking lower TDP or added IGP into account). And to the answer to that question must be Yes I think. Currently we're only seeing 5-25% performance increases each year. Not much to cheer about...

I, too, was talking about pure CPU performance ideas. However it's true that I don't know how much % improvement these ideas count for. When you study software traces and most of them have all the hardware going full blown, you either add more hardware and reduce the number of instructions required (the 2nd one is usually the more interesting one). Just like if we never had a multiplier unit but "ah ha!" turns out people like to multiply numbers. We no longer have to do a series of additions and shift operations. That's at least how one part of the chip gets performance. Then there are still some solutions not so straight forward (and perhaps limited to x86 processors) that would improve other areas as well.

So for me, I'm definitely biased since I'm exposed to this stuff daily but for me there's definitely a long laundry list of things I want but a shorter list of things that I can get done in time. But I'll concede the point that I leave the % to someone else to figure out.

 

[TuxDave]

So you think it is possible to speed things up with a 128 bit processor?

I always hesitate with these questions since I'm always wondering if what I think as a 128-bit processor is what you think of for a 128-bit processor. To answer your question, do we need more address space? And how many high performance traces require large integer operations (there are some!).

 

[ShintaiDK]

So you think it is possible to speed things up with a 128 bit processor?

All the parts that yield the main benefits are already done at 128bit(SSE) and 256bit(AVX(Haswell will extend this)).

 

[ShintaiDK]

Depends on what timeframe you compare with.

We also went from:

Pentium II 233 Mhz (May 7, 1997)
Pentium II 450 Mhz (August 24, 1998)
Pentium III 800 MHz (December 20, 1999)
Pentium 4 2.0 Ghz (August 27, 2001)

I.e. the CPU performance increased by almost 1000% in about 4 years!

Are you claiming a 2Ghz P4 is 10 times faster than a 233Mhz P2? Maybe you can show me the benches then.

See this post by AtenRa - and that post discusses current 22 nm technology. One node shrink from now (i.e. Broadwell), an 8 core CPU at $300 should clearly be possible. If Intel wants to release such a CPU that is.

Nobody said it wasnt possible. You can already buy octocores today. You just aint willing to pay the price.

It cost Intel around 10$ to manufactor an i7 3770K and maybe 20$ for an octocore Xeon. But thats not the reason they cost what they do. And no, its not because Intel is greedy either and not giving poor you cheap octocores.

 

[Fjodor2001]

Are you claiming a 2Ghz P4 is 10 times faster than a 233Mhz P2? Maybe you can show me the benches then.

I don't know what the exact performance difference is, and it likely also varies between different scenarios since the P4 has additional performance improvements (e.g. new instruction sets) in addition to the frequency difference alone. However, the point is that the increases in CPU performance was extremely much greater back then.

Nobody said it wasnt possible. You can already buy octocores today. You just aint willing to pay the price.

It cost Intel around 10$ to manufactor an i7 3770K and maybe 20$ for an octocore Xeon. But thats not the reason they cost what they do. And no, its not because Intel is greedy either and not giving poor you cheap octocores.

Does it matter what I personally would consider buying various CPU versions for? What if I'd pay $1 or $100.000? Does it matter? Once again - we're discussing general principles here for gods sake! Can you please stop stalking me!

The point is that with the next node shrink (Broadwell), Intel can fit 8 CPU cores on the same die area as current 4 core IB. Based on current pricing, that ought to mean Intel should be able to sell it for around $300 and still keep their margins (they have to cover for development & marketing costs etc, so the cost of producing them which you mentioned is irrelevant). Whether they want to do that or not is a different story.

 

[Hugo Drax]

We are Software constrained. All that CPU power sitting idle in 95% of the average users desktop.

 

[moonbogg]

We are Software constrained. All that CPU power sitting idle in 95% of the average users desktop.

Yeah, exactly. So, we are reaching a practical limitation, first due to hardware, then due to software. Single threaded performance is leveling off and software can't take advantage of multiple cores very well. We need nanotube transistors or something to fix it.

 

[Lonbjerg]

640k ought to be enough for anybody.

[citation needed]

Let the myth die...please.

 

[moonbogg]

[citation needed]

Let the myth die...please.

What about the following quotes and sentiments from Nvidia's VP shortly after nvidia's CEO openly mocked Intel? They both claim the CPU is dead and you "Don't need a fast one anymore" Please, do read. Oh, and lol?

Link will follow.

Nvidia now claims CPU is dead

Previous Nvidia CEO criticized Intel’s graphic offerings, calling them a joke. Now it’s Roy Taylor’s turn. Roy Taylor, as we all don’t know, is Nvidia’s vice president of content relations (What?).

From: Roy Taylor [mailto:RTaylor@nvidia.com]
Sent: 10 April 2008 23:36
Subject: The best job in the world.
Guys I have the best job in the world. Official. I cant tell you how much fun we’re having here right now.
I don’t know how much this will mean to you all but for those that don’t know a war has just started that will likely be written about for years and which will affect everyone who owns a PC. Everyone.
Basically the CPU is dead. Yes, that processor you see advertised everywhere from Intel. Its run out of steam. The fact is that it no longer makes anything run faster. You don’t need a fast one anymore. This is why AMD is in trouble and its why Intel are panicking. They are panicking so much that they have started attacking us. This is because you do still [need] one chip to get faster and faster – the GPU. That GeForce chip. Yes honestly. No I am not making this up. You are my friends and so I am not selling you. This s***t is just interesting as hell.

http://i.justrealized.com/tag/nvidia/
​

 

[cytg111]

A quad-core Sandy Bridge at 130W is 50% faster than a quad-core Bloomfield at 130W. Moore's law itself only states that the number of integrated circuits on components doubles every year.

An i7-3960X has 2.27 billion transistors. An i7-990X has 1.17 billion. That's a 94% increase - pretty close to what Moore's Law says.

That "law" is more of a self-fulfilling prophecy than anything else. It is generally used as a benchmark for what should be produced.

Moore's law says nothing about lowering power consumption while increasing performance or transistor count by 100%.

- Yes, correct, number transistors, wich again is expected to translate directly into performance ... I dont want to go into the semantics, cause noone wins there. Point being, transistors not being used (as in not-used-cores) is worth zip nada nothing. And you wont see(i predict) more than 4 cores mainstream untill we get that construct that lets our algorithms scale with core count.

I dont get the "used as benchmark". I dont understand it.. It must be like "The industry produces" and then, after the fact, we will hold moores law to it.. Not the other way around.

Simplified.
Moores law says this : for every 18 months your FPS in game X doubles.
Enterpatation? Yes.
On the Money? Indeed.

 

[cytg111]

Parallelization. When you don't know how to increase performance per core, you increase cores. ...

Yes.
If it was plausible to create a 2 core monster that performed at 2x singlethread performance of a 4 core-whatever x86 within the same power envelope... AMD would (hopefully) have been all over that.
It has been in the 'news' many times now, from both amd & intel stating that there is no more low hanging fruit in regards to single threaded performance.

 

[Ferzerp]

AMD would (hopefully) have been all over that

What gives you that idea? AMD is all about moar "cores" regardless of "core" performance (and whether or not a "core" is actually a core)

 

[cytg111]

What gives you that idea? AMD is all about moar "cores" regardless of "core" performance (and whether or not a "core" is actually a core)

- Excatly, they took the only route they could.
I dont understand why they didnt have a compiler-plugin-whatever ready for something like bulldozer.. it makes zero sense atm in the consumer segment.

 

[EltonL]

No cpu technology hasn't run out of steam. Currently we are software constrained when it comes to threading. In terms of IPC we have to also look at power consumption reduced. On a core to core basis, the current generation is still being improved while simultaneously also consuming less power. That's progress.

If anything, its time to get scaling in order and have it so theres good scaling across the board rather than diminishing returns per core. Note that IBM's Power7 can hit on average 4.5Ghz stock.

 

[Lonbjerg]

What about the following quotes and sentiments from Nvidia's VP shortly after nvidia's CEO openly mocked Intel? They both claim the CPU is dead and you "Don't need a fast one anymore" Please, do read. Oh, and lol?

Link will follow.

Nvidia now claims CPU is dead

Previous Nvidia CEO criticized Intel’s graphic offerings, calling them a joke. Now it’s Roy Taylor’s turn. Roy Taylor, as we all don’t know, is Nvidia’s vice president of content relations (What?).

From: Roy Taylor [mailto:RTaylor@nvidia.com]
Sent: 10 April 2008 23:36
Subject: The best job in the world.
Guys I have the best job in the world. Official. I cant tell you how much fun we’re having here right now.
I don’t know how much this will mean to you all but for those that don’t know a war has just started that will likely be written about for years and which will affect everyone who owns a PC. Everyone.
Basically the CPU is dead. Yes, that processor you see advertised everywhere from Intel. Its run out of steam. The fact is that it no longer makes anything run faster. You don’t need a fast one anymore. This is why AMD is in trouble and its why Intel are panicking. They are panicking so much that they have started attacking us. This is because you do still [need] one chip to get faster and faster – the GPU. That GeForce chip. Yes honestly. No I am not making this up. You are my friends and so I am not selling you. This s***t is just interesting as hell.

http://i.justrealized.com/tag/nvidia/
​

I was talking about the myth about Bill Gates and the quote he never made about 640K...

 

[IntelUser2000]

We also went from:

Pentium II 233 Mhz (May 7, 1997)
Pentium II 450 Mhz (August 24, 1998)
Pentium III 800 MHz (December 20, 1999)
Pentium 4 2.0 Ghz (August 27, 2001)

I.e. the CPU performance increased by almost 1000% in about 4 years!

And the TDP went from 34.8W to 71.8W. The fact is last few decades or so of massive CPU gains came from power usage increases that equal the performance gain. That happened with GPUs too. Which is why Nvidia's claim of GPUs being on a 3x Moore's Law cadence deteriorated pretty quickly.

And later we also went from P4 -> C2D in one CPU generation, reducing power consumption by around 30-50%, and more than doubling CPU performance... :biggrin:

We also had far more applications that really demand CPU power. Even browsing and word processing had a point where it was a "power user" application, meaning those hit TDP too. Also, you are wrong about doubling CPU performance. Performance/clock doubled yes, absolute top chip to top chip doubled no.

See this post by AtenRa - and that post discusses current 22 nm technology. One node shrink from now (i.e. Broadwell), an 8 core CPU at $300 should clearly be possible.

Whether they can do it or not is not important. The thing is that people who need that CPU is in practice non existent in the enthusiast crowd. The fact that such systems moved from a need to a want means they don't need to make such cheap chips.

It cost Intel around 10$ to manufactor an i7 3770K and maybe 20$ for an octocore Xeon.

That's for the die alone. You still need to add packaging costs. The die is probably around $15. But production costs are never the sole reason for selling costs.

 

[Cerb]

Pentium II 233 Mhz (May 7, 1997)
Pentium II 450 Mhz (August 24, 1998)
Pentium III 800 MHz (December 20, 1999)
Pentium 4 2.0 Ghz (August 27, 2001)

I.e. the CPU performance increased by almost 1000% in about 4 years!

No, clock speed went up 850% in four years. Performance did increase faster than today, but that P4 2GHz would generally be about equivalent to a PIII 1.5 or 1.6, and even that needed RDRAM to work out.

And later we also went from P4 -> C2D in one CPU generation, reducing power consumption by around 30-50%, and more than doubling CPU performance... :biggrin:

Intel did not go from the P4 to the Core 2. They went from the P4 back to an updated P3 to become the Pentium M, which then became the Core Solo/Core Duo, which then became the Core 2 Duo [and MCM Core 2 Quad]. That's two in-between generations, three if you count the never-released R&D on the P3.

 

[IntelUser2000]

Intel did not go from the P4 to the Core 2. They went from the P4 back to an updated P3 to become the Pentium M, which then became the Core Solo/Core Duo, which then became the Core 2 Duo [and MCM Core 2 Quad]. That's two in-between generations, three if you count the never-released R&D on the P3.

I think we should compare top end to top end. The top CPU before Core 2 wasn't the Pentium 4 or the Pentium D, it was the Athlon X2. The "what-ifs" could have gone either way. The successor to Prescott could have been Tejas and Intel might be 1/3rd of the size they are now, with AMD dominating.

Athlon X2 to Core 2 resulted in 30% performance gain with ~10% power use reduction.

It's true that CPU gains from new generations are lower. But gains weren't free. 486 CPUs used low enough power to be used in cellphones and Pentium chips had lesser cooling than some PCH chips today! We can't afford to have that magnitude of increase in power usage anymore.