Intel Starts Production of Next-Generation Haswell Microprocessors.

[cbn]

The idea is that they're increasing IPC in a very power efficient way.

If power wasn't a problem, you could see huge IPC boosts. But in this world, we like our battery life.

I've been wondering if Intel is holding back a lot of the IPC gains till we see an improvement in I/O (specifically the whole RAM and Storage convergence thing.)

I mean, on a very high level, why should Intel boost single thread performance when the ability to feed the cores data is being bottlenecked elsewhere?

 

[zephyrprime]

Okay, just for something basic, I looked up the passmark scores for a couple of chips. (Yes, I'm sure it's not the best benchmark, but it's a quick and dirty way of comparing chips.)

Intel Core2 X6800 @ 2.93GHz -- 2,037
Intel Core i7-3770K @ 3.50GHz -- 9,631

So that's a factor of about 5 in 7 years. Some of that higher number is also because the i7 is a quad and the Core2 is a dual.

Okay, how about something from Nehalem?

Intel Core i7 965 @ 3.20GHz --5,763

This is where things look pathetic. That chip came out in 2008.

I mean, I'm running an i7 920 myself, and I don't yet see anything that's giving me the itch to build a new machine.

It can't be helped. It's just that we've reached the end of Moore's law for all practical purposes. Even though transistors will still scale down for a while longer, it will hardly help performance. The number one thing that has increased performance since Moore first announced his law was the increase in clock speeds ( > 1000x gain). IPC has increased by much less.

This is actually the beginning of the end for the IT revolution. Innovation will start to slow now. Mobile is probably the last hurrah for a long time.

 

[Charles Kozierok]

How many people would actually buy them, especially when people still just overclock 77W or 95W CPUs to achieve the same performance levels?

No way to know for sure, but my guess is more than Intel thinks. There are still people out there who do real work and do not want to start tinkering with voltage levels and clock speeds.

It's been said over and over again in this thread, Intel's main focus isn't on the desktop because their latest chips' only competition is a previous gen chip. Why dedicate more resources there?

And I agree with that. They shouldn't focus there.

But why can't they pick off a bit of low-hanging fruit in the performance department? Use their power budget for cores and hertz, sell the chip for a kilobuck, and everyone's happy. Plus, as a bonus, they continue to rub it in AMD's face.

But it seems these days that corporations and CEOs get fixated on one "strategy" and everything else falls by the wayside.

This is actually the beginning of the end for the IT revolution. Innovation will start to slow now. Mobile is probably the last hurrah for a long time.

Famous last words.

 

[Vesku]

I want to know one thing:
-- will "FSB" overclock be back, with clockgen and all the power being integrated and software manipulated in haswell?

I believe there will be some "FSB" tweaking with the Haswell Ks.

http://www.xbitlabs.com/news/cpu/di...ocessors_to_Further_Improve_Overclocking.html

 

[Idontcare]

It can't be helped. It's just that we've reached the end of Moore's law for all practical purposes. Even though transistors will still scale down for a while longer, it will hardly help performance. The number one thing that has increased performance since Moore first announced his law was the increase in clock speeds ( > 1000x gain). IPC has increased by much less.

This is actually the beginning of the end for the IT revolution. Innovation will start to slow now. Mobile is probably the last hurrah for a long time.

Moore's law is about the production cost of an IC, not about the performance of the IC.

It correlated the minimum value of the cost-per-component (xtor) curve with time and noted that over time this minimum value moved out to ICs composed of larger and larger numbers of components.

For some dang reason this observation became falsely associated with "xtors are doubling, so performance must be doubling too!"...which had nothing to do with Moore's Law, ever, and wasn't true on the face of it either (clockspeed bumps were required to drive performance).

Now it is also true that Moore's Law has held, the number of transistors per IC over time has consistently doubled every 2.1yrs since the dawn of the invention of the transistor itself.

Intel has long relied on expanding the ISA, adding new extensions, as a means of increasing performance. It doesn't necessarily improve IPC for legacy instructions all that much, but the real performance strides to be found are always with the newest ISA extensions and recompiling your applications to take advantage of them.

 

[Haserath]

http://www.realworldtech.com/cpu-perf-analysis/

Increasing performance also requires refining aspects that have already been refined.

 

[cbn]

Btw, doesn't Intel still release some CPU models without integrated GPUs?

Yep, check out LGA 2011.

 

[Turbonium]

Okay, but is that really true? Is a top-of-the-line IB CPU really "dozens of times" faster than a 2006-era Conroe on typical workloads (i.e. not cheating by using stuff that's easily multithreaded)? I don't know, but I'm highly skeptical. My guess is that it's twice as fast, maybe a bit more than that, and maybe even less than that.

Consider the above "Idontcare bait".

I was actually thinking along the lines of a PIII or P4 (Willamette or early Northwood) when I mentioned older CPUs. Surely, a top-tier SB or IB CPU is at least one dozen times faster at certain tasks? Just looking at basic benchmarks seems to suggest this (at least in my opinion).

 

[notty22]

No way to know for sure, but my guess is more than Intel thinks. There are still people out there who do real work and do not want to start tinkering with voltage levels and clock speeds.



And I agree with that. They shouldn't focus there.

But why can't they pick off a bit of low-hanging fruit in the performance department? Use their power budget for cores and hertz, sell the chip for a kilobuck, and everyone's happy. Plus, as a bonus, they continue to rub it in AMD's face.

But it seems these days that corporations and CEOs get fixated on one "strategy" and everything else falls by the wayside.



Famous last words.

\
Thought that describes this?
Intel Core i7-3970X Extreme Edition Sandy Bridge-E 3.5GHz (4.0GHz Turbo) LGA 2011 150W Six-Core Desktop

 

[zephyrprime]

Moore's law is about the production cost of an IC, not about the performance of the IC.

I knew someone was going to say that. Yes, that's what original moores law was. But it doesn't matter. Original moore's law isn't the law that matters. First off, original moore's law stopped being correct in only a few years because it originally said that density double every 12 months. This was soon changed to every 18 months. And this law was then modified to performance doubles every 18 months. This last variant is the Moore's law that matters because it was the version of moore's law that held true for 20+ years. Like I said before, the original version of moore's law failed after only a few years. Screw the original version of moore's law. Nobody in the IT industry ever talked about the original version of moore's law except as a historical curiosity until a few years ago we had to fall back to the older version of moore's law because the newer version had failed. And it will keep failing. Scaling is meeting the wall.

So Moore's law is dead.

 

[Homeles]

I knew someone was going to say that. Yes, that's what original moores law was. But it doesn't matter. Original moore's law isn't the law that matters. First off, original moore's law stopped being correct in only a few years because it originally said that density double every 12 months. This was soon changed to every 18 months. And this law was then modified to performance doubles every 18 months. This last variant is the Moore's law that matters because it was the version of moore's law that held true for 20+ years. Like I said before, the original version of moore's law failed after only a few years. Screw the original version of moore's law. Nobody in the IT industry ever talked about the original version of moore's law except as a historical curiosity until a few years ago we had to fall back to the older version of moore's law because the newer version had failed. And it will keep failing. Scaling is meeting the wall.

So Moore's law is dead.

Moore's Law is actually the doubling of the population of unicorns, every two years.

See, look. I can make up stuff too.

Moore's Law has never had anything to do with performance. You can apply it to performance, just as you can apply the laws of physics to engineering. But the laws of physics themselves have nothing to do with engineering — they're just observations of nature.

And the top two search results in Google easily demonstrate this:
http://en.wikipedia.org/wiki/Moore's_law
http://www.intel.com/content/www/us/en/silicon-innovations/moores-law-technology.html

And the original report by Moore is here:
http://download.intel.com/museum/Moores_Law/Articles-Press_Releases/Gordon_Moore_1965_Article.pdf

Stop trying to rewrite history.

 

[Idontcare]

I knew someone was going to say that. Yes, that's what original moores law was. But it doesn't matter. Original moore's law isn't the law that matters. First off, original moore's law stopped being correct in only a few years because it originally said that density double every 12 months. This was soon changed to every 18 months. And this law was then modified to performance doubles every 18 months. This last variant is the Moore's law that matters because it was the version of moore's law that held true for 20+ years. Like I said before, the original version of moore's law failed after only a few years. Screw the original version of moore's law. Nobody in the IT industry ever talked about the original version of moore's law except as a historical curiosity until a few years ago we had to fall back to the older version of moore's law because the newer version had failed. And it will keep failing. Scaling is meeting the wall.

So Moore's law is dead.

Your interest in discrediting Moore's law is blinding you to the reason it was ever a relevant observation to be made in the first place.

You've built yourself a strawman that you label "Moore's Law" because it is convenient to discredit the fictional entity you hold in your mind.

Its ok, I won't make any further effort to alter your course, you have far more conviction to make the devil out of something here than I have conviction to attempt to compel you to see otherwise.

 

[Haserath]

Moore's law was every 2 years or so, which Intel is still effectively on.

According to Wikipedia, the double of performance every 18 months started thanks to Intel. This was only due to more transistors and faster speeds, which we don't get much of either anymore.

 

[Homeles]

Moore's law was every 2 years or so, which Intel is still effectively on.

According to Wikipedia, the double of performance every 18 months started thanks to Intel. This was only due to more transistors and faster speeds, which we don't get much of either anymore.

From the article:

In multi-core CPUs, the higher transistor density does not greatly increase speed on many consumer applications that are not parallelized. There are cases where a roughly 45% increase in processor transistors have translated to roughly 10–20% increase in processing power.[72] Viewed even more broadly, the speed of a system is often limited by factors other than processor speed, such as internal bandwidth and storage speed, and one can judge a system's overall performance based on factors other than speed, like cost efficiency or electrical efficiency.

 

[Haserath]

From the article:

I know.

We don't get the increase much anymore due to diminishing returns and Intel focusing on efficiency, which is why Haswell focuses mostly on the memory system.

 

[Idontcare]

Performance scaling has always been a difficult metric to firmly nail down and track over time because of the very fact that performance is application dependent, meaning it is instruction dependent and not all instructions in the ISA are targeted equally for performance enhancements year over year.

That there was ever any reasonable ability for people to plot time versus performance and fit the data to a logarithmic function (doubling every X periods of time, where X might be 12 months, 18, 24, etc) is a miraculous coincidence in and of itself because it requires a consistent FOM (figure of merit) to be applicable to the characterization of performance across a wide span of time when everyone knows that the ideal performance to be gained from any given microarchitecture would entail using a different mix of instructions in comparison to its predecessor.

That is why new instructions are introduced with every major microarchitecture release. We are up to 2000 instructions in our most recent x86 chips. 2000!

What sort of budget do you think is being allocated towards improving the IPC of the original x86 instructions, or x87, or MMX, or SSE? That isn't where the money is going, so it shouldn't be of any surprise that the performance scaling of apps that depend on those instructions is doing no better than whatever basic clockspeed scaling is happening in the background.

 

[Homeles]

Will there ever be a point where they kill off backwards compatibility for those deprecated instructions? Do modern applications even use something like MMX?

 

[alan1476]

6 Native Intel SATA 6.0 ports, thats the best upgrade to Ivy that Haswell has.

 

[Homeles]

6 Native Intel SATA 6.0 ports, thats the best upgrade to Ivy that Haswell has.

How inane. TSX, AVX2, on package VRM, embedded DRAM on ULV parts, a stronger memory controller, et cetera, et cetera. And you choose IO?

That may be the most important thing to you, but there are revolutionary technologies that are being implemented in Haswell. You can always buy a PCI-E card with SATA 6 Gbps ports.

 

[TuxDave]

Will there ever be a point where they kill off backwards compatibility for those deprecated instructions? Do modern applications even use something like MMX?

Some math libraries still call on functions that contain MMX or SSE instructions and no one has bothered to update them. Is it dumb? Yes? Oh well.

The hardware saving you get from removing MMX is very very very tiny. We would be talking fractions of mW. You would more likely see it as a performance hit over a power hit. And if the performance hit isn't significant enough for anyone to update that library.... well... then I guess we're all stuck with it.

 

[Phynaz]

Will there ever be a point where they kill off backwards compatibility for those deprecated instructions?

Doubtful. The amount of die area they use is so small there would be almost no benefit from it. And it would quickly turn into a negative when you tell your customers you are going to break their applications. Forced upgrades are despised by corporations.

Edit: Dave beat me to it.

 

[Homeles]

Even ad infinitum? There has to be some point where they move onto something better.

 

[TuxDave]

Even ad infinitum? There has to be some point where they move onto something better.

Well, what point would that be? MMX is a decade and a half old. It's one of those things that won't die.

 

[Idontcare]

Even ad infinitum? There has to be some point where they move onto something better.

They are continually moving on to something better. That is what ISA extensions are...them moving on.

But moving on to something better need not be exclusive to keeping what you've already developed.

Why would moving on require them to break backwards compatibility and pitch ISA support for any given set of instructions?

Think about why people upgrade hardware? To make their existing software investment have a higher ROI*.

Why do people upgrade their software? To make their existing hardware investment have a higher ROI*.

Rarely is there opportunity for new hardware and new software to be purchased simultaneously that will also deliver a superior ROI. We talk all the time about how new ISA extensions will be the bomb with newly compiled versions of software but rarely is that the combination that gets deployed.

New hardware needs to make a compelling ROI story for existing software, otherwise it is a no-go simply from a numbers point of view. Breaking backwards compatibility would undermine the ROI opportunity for anyone considering to buy the new hardware.

For an example of just how difficult it is to break compatibility and convince the market to adopt your hardware you need look not further than Itanium. It was an arduous path for Itanium to gain critical mass, it finally got there but it took an Intel and an HP to bankroll it year after year until it gained the traction it needed to become self-sustaining.

Apple did similar with their transition to x86 from Power but there case is unique because they had very little market share to begin with. If your user base is 3% of the TAM then it is easy to abandon them and pick up a new set of users while still hitting your 3% TAM market share.

Intel couldn't possible do that with x86, not now. Way past the point of no return there IMO.

* higher compute performance, power efficiency, performance/W, performance/$, footprint, TCO, etc.

 

[cytg111]

Why would moving on require them to break backwards compatibility and pitch ISA support for any given set of instructions?.

- Clearly Intel thought they were going somewhere with IA64, that compilers could/should take on more than they do today. They believed in it enough to try and sink their x86 xeon lines in comparison.
Just saying that, apparantly, somone at Intel thinks/thought it is/was worth breaking backwards compatibility for new hardware/ a new paradigm.
On the other hand, i think we keep reading, here and other places, that the hardware involved with a superscalar design is really not that big of a deal/hindrance for performance onwards.
So call me semiconfused.