Haswell model specs leaked

[ShintaiDK]

you don't need to tell me hz are not everything. But still, it matters. A lot. And going from pii 266 mhz -> p4 2.0 ghz we did see an ipc increase too. In addition to an 8.6x cpu frequency increase.

2000/266=7.5....But since the 300Mhz was also there in 1997 its down to 6.67x.

Also power consumption went radically down from Yorkfield to Ivy. Where as from P2 266 to P4 2Ghz it went rapidly up. The 2Ghz in 2001 was also the top bin CPU. 3770K aint the top bin.

 

[Idontcare]

2000/266=7.5....

He was referring to his earlier statement on PII 233MHz, but got the numbers mixed up in the more recent post. 2000/233=8.6

 

[TuxDave]

You don't need to tell me Hz are not everything. But still, it matters. A lot. And going from PII 266 MHz -> P4 2.0 GHz we did see an IPC increase too. In addition to an 8.6x CPU frequency increase.

Going from Yorkfield->IB we've perhaps seen more IPC increase. But only 2.83 GHz (Q9550) -> 3.5 GHz (3770K) => 1.23x frequency increase. So the IPC increase going from Yorkfield->IB is nowhere near enough to compensate for the higher frequency increase we saw going from PII->P4.

To sum it up: Show me benchmarks where the relative CPU performance increase going from PII 266 MHz (1997) -> P4 2.0 GHz (2001) is lower compared to going from Q9550 (2007/2008) -> 3770K (2012), then I'll believe you.

Can I use an AES benchmark?

 

[bunnyfubbles]

Sandy Bridge improved IPC AND clocks. Ivy Bridge didn't improve much, barely at all. I read somewhere that Ivy was not fully optimized for the 22nm process but Haswell is. So I expected Intel to catch up a bit.

yeah, because an extra 100MHz after a process shrink is a major change

 

[ShintaiDK]

He was referring to his earlier statement on PII 233MHz, but got the numbers mixed up in the more recent post. 2000/233=8.6

Just wondering why he picks the lowest bin in 1997 instead of the highest. And then compares it to the top bin in 2001.

 

[Fjodor2001]

Just wondering why he picks the lowest bin in 1997 instead of the highest. And then compares it to the top bin in 2001.

Sure, you could even select the P II 300 MHz instead. However it was $1981 when released on 7 May 1997...

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_II_microprocessors#Desktop_processors

So not really a comparable mainstream CPU at the time. The 233 MHz was "only" $636 on 7 May 1997 though, so it should be considered a mainstream CPU at the time (since CPUs were priced at that level back then). Also prise-wise comparable to the P4 2.0 GHz which was $562 when released on 27 August 2001, see:

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_4_microprocessors#Willamette_.28180.C2.A0nm.29

But I'm prepared to ignore the insane price of the PII 300 MHz. So select that one instead of the PII 233/266 MHz if you'd like for the comparison. Just don't stray away from the main subject.

Now please show me the benchmarks I requested here... :whiste:

 

[TuxDave]

Now please show me the benchmarks I requested here... :whiste:

I tried just to see what the numbers ended up being. But then I couldn't find a common benchmark that stretched from P2 to Ivybridge.

 

[Lepton87]

Sure, you could even select the P II 300 MHz instead. However it was $1981 when released on 7 May 1997...

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_II_microprocessors#Desktop_processors

So not really a comparable mainstream CPU at the time. The 233 MHz was "only" $636 on 7 May 1997 though, so it should be considered a mainstream CPU at the time (since CPUs were priced at that level back then). Also prise-wise comparable to the P4 2.0 GHz which was $562 when released on 27 August 2001, see:

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_4_microprocessors#Willamette_.28180.C2.A0nm.29

Don't compare it against P4 2GHz it wasn't the top dog at the time Athlon XP: 1733 MHz (2100+) was faster. So going from P2 266MHz to Atholon XP 1733MHz WAS a HUGE performace increase, IPC increased more than 2x times and frequency increased 6.5 times. Recent advances pale in comparison when it comes to raw speed.

 

[ShintaiDK]

Sure, you could even select the P II 300 MHz instead. However it was $1981 when released on 7 May 1997...

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_II_microprocessors#Desktop_processors

So not really a comparable mainstream CPU at the time. The 233 MHz was "only" $636 on 7 May 1997 though, so it should be considered a mainstream CPU at the time (since CPUs were priced at that level back then). Also prise-wise comparable to the P4 2.0 GHz which was $562 when released on 27 August 2001, see:

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_4_microprocessors#Willamette_.28180.C2.A0nm.29

But I'm prepared to ignore the insane price of the PII 300 MHz. So select that one instead of the PII 233/266 MHz if you'd like for the comparison. Just don't stray away from the main subject.

Now please show me the benchmarks I requested here... :whiste:

Prices are due to volume mainly. And the 562$ P4 2.0Ghz was announced as a "very aggressive" pricing.

IPC wise depends in what. A Q9550 for example gives something like 30Gflops in Linpack. My own 3570K gives over 100Gflops. And Haswell will yield way past 200Gflops(most likely around 230Gflops). Or over 6.67x (7.67x) from Yorkfield to Haswell. Not to mention again, much lower power consumption. Unlike the P2->P4.

Priorities simply changed in favour of lower power consumption and IGPs.

Not to mention the option of the issue of defining timeframes. Going from a 2Ghz P4 in 2001 to a dualcore 3.6Ghz P4 in Q2/2006 isnt exactly noteworthy. Then the progress today looks pretty good.

 

[Idontcare]

I tried just to see what the numbers ended up being. But then I couldn't find a common benchmark that stretched from P2 to Ivybridge.

I have a benchmark that I've tracked since I owned a PII 450MHz. Haven't ran it on Ivy yet, but I will.

 

[Torn Mind]

You don't need to tell me Hz are not everything. But still, it matters. A lot. And going from PII 266 MHz -> P4 2.0 GHz we did see an IPC increase too. In addition to an 8.6x CPU frequency increase.

Going from Yorkfield->IB we've perhaps seen more IPC increase. But only 2.83 GHz (Q9550) -> 3.5 GHz (3770K) => 1.23x frequency increase. So the IPC increase going from Yorkfield->IB is nowhere near enough to compensate for the higher frequency increase we saw going from PII->P4.

To sum it up: Show me benchmarks where the relative CPU performance increase going from PII 266 MHz (1997) -> P4 2.0 GHz (2001) is lower compared to going from Q9550 (2007/2008) -> 3770K (2012), then I'll believe you.

There is a Q9650 that runs at 3.0 Ghz.
http://www.anandtech.com/bench/Product/49?vs=551
Q9650(3.0 Ghz) vs. i7-3770k(3.5 Ghz)
Performance is between 1.5 times to 2 times the performance of the Q9650, all while having a measly-looking clock increase of 500 Mhz.

http://www.anandtech.com/bench/Product/49?vs=701
Q9650 to i7-3570k
I use the 3570k as well because Hyperthreading can affect IPC.



Since we're considering 4-year periods, then there is the Pentium II -> Willamette Pentium 4s and the Coppermine Pentium III -> Northwood Pentium 4s.

Pentium II benchmarks are hard to find, but the following at least compares some vintage CPUs to the top-of-the-line Northwood released on November 25, 2002.

http://www.tomshardware.com/reviews/benchmark-marathon,590.html
This is the closest I can find. Too bad the only Pentium IIs they had were Celeron Mendocinos, it seems.

But if it's a Mendocino vs. the Athlon XP 2100+, then see an increase of approximately 7 times that of the Mendocino at its lowest clock.

There does seem to be a pattern of diminishing returns with regards to clockspeed increase though. After accounting for Coppermine's IPC increase, the performance gain is 4 times that of the Mendocino for only 600 Mhz for the MP3 Maker Platinum. Even having a clockspeed jump of 2.0 GHz only triples the performance compared to the Coppermine at 1.0 Ghz, even with Hyperthreading.

Comparing the Mendocino to the Pentium 4 2.0A, the performance gain is about five times that of the Mendocino 400 Mhz.

 

[boxleitnerb]

yeah, because an extra 100MHz after a process shrink is a major change

Where did I say that?
I said, they improved barely at all with Ivy (on the CPU side, that is), which is absolutely true.

 

[Torn Mind]

Where did I say that?
I said, they improved barely at all with Ivy (on the CPU side, that is), which is absolutely true.

There isn't much more room to improve. Pentium 4 and Piledriver already have basically shown where the border for stock clockspeeds are (4.0 Ghz). Any further and things start getting toasty, literally.

 

[boxleitnerb]

I don't believe that they could not have clocked Ivy 10 or even 15% higher. That is 4 GHz, which is very easy to achieve when overclocking with very moderate increase in power consumption, especially when not raising voltage (much).

 

[Torn Mind]

I don't believe that they could not have clocked Ivy 10 or even 15% higher. That is 4 GHz, which is very easy to achieve when overclocking with very moderate increase in power consumption, especially when not raising voltage (much).

It's probably due to the TDP envelope and/or stability then, i.e corporate customers or professionals cannot afford a hiccup.

The Sandy Bridge 3970X is clocked only 200 Mhz higher than the 3960X, but it has a TDP of 150W compared to the 130W of the 3960X.

 

[Fjodor2001]

Prices are due to volume mainly. And the 562$ P4 2.0Ghz was announced as a "very aggressive" pricing.

IPC wise depends in what. A Q9550 for example gives something like 30Gflops in Linpack. My own 3570K gives over 100Gflops. And Haswell will yield way past 200Gflops(most likely around 230Gflops). Or over 6.67x (7.67x) from Yorkfield to Haswell. Not to mention again, much lower power consumption. Unlike the P2->P4.

Priorities simply changed in favour of lower power consumption and IGPs.

Not to mention the option of the issue of defining timeframes. Going from a 2Ghz P4 in 2001 to a dualcore 3.6Ghz P4 in Q2/2006 isnt exactly noteworthy. Then the progress today looks pretty good.

Well, I'd at least prefer to not use benchmarks from special use cases such as AES256. Linpack/GFlops may be relevant though. But also e.g. SPECint and "SYSMARK overall" would be interesting.

Also, Q9550->Haswell is 5 years, not 4. And we do not know exactly what Haswell will bring anyway. But as I understood you Q9550->3570K gives 30->100 GFlops in Linpack (do you have exact numbers BTW, and not just "something like"), so a 3.3x increase. Did PII/300->P4/2.0 really not bring more than that? Any benchmarks to prove that?

 

[ShintaiDK]

Well, I'd at least prefer to not use benchmarks from special use cases such as AES256. Linpack/GFlops may be relevant though. But also e.g. SPECint and "SYSMARK overall" would be interesting.

Also, Q9550->Haswell is 5 years, not 4. And we do not know exactly what Haswell will bring anyway. But as I understood you Q9550->3570K gives 30->100 GFlops in Linpack (do you have exact numbers BTW, and not just "something like"), so a 3.3x increase. Did PII/300->P4/2.0 really not bring more than that? Any benchmarks to prove that?

You obviously missed my point. What about P4 2Ghz to P4 3.6Ghz. 5 years. Why dont you use that time period? Or previous periods for example back in the 80s to early 90s? You simply picked a golden period and then claim it as a baseline. Also you completely disregard the performance/watt case. If we had to follow your P2 to P4 line. CPUs today should have been 200W+. The power envelope was just brutally abused as a low hanging fruit.

The performance/watt for the p2 -> p4 in your case would most likely sit around 4x. And from 2001 to 2006 its below 2x. Its not as easy when you cant abuse the power envelope. SB to IB improved this around 30-35% alone. And from Q9550 to 3770 the performance/watt if we use Linpack (How accurate or misleading it might be.) we got an increase of 4.5x.

The Linpack numbers cept for Haswell is pretty much exact. Q9550 gets between 30 and 31. And 3570K around 100. The issue about "around" is it also depends on the memory used. But instead of using the 3770, you should atleast use the 3930 chip. And then you sit with Linpack numbers around 150Gflops. Or 5x over Q9550 in 3½ years.

And as I already said before, priorities simply changed in favour of lower power consumption and IGPs. Like it or not, but thats what the 99% crowd wants.

 

[Fjodor2001]

You obviously missed my point. What about P4 2Ghz to P4 3.6Ghz. 5 years. Why dont you use that time period? Or previous periods for example back in the 80s to early 90s? You simply picked a golden period and then claim it as a baseline. Also you completely disregard the performance/watt case. If we had to follow your P2 to P4 line. CPUs today should have been 200W+

I took it as an example. But you can pick other 4 year periods too you'd like, and compare that to the last 4 years. See:

http://www.intel.com/pressroom/kits/quickrefyr.htm
(Note: It's somewhat hard to compare if you go back to 1980 or so, since new CPU models were released much less frequently then, about every 3 years or so, but then each new model was also a bigger "jump" compared to the previous one, e.g. 286/12Mhz->386DX/33Mhz)

Still you'll find the same thing. CPU performance increase the last 4 years have not been that great, if you look at general benchmarks such as SPECint and "SYSMARK overall".

Sure, if you look back there have been some other periods as well where the CPU performance increase was not so impressive (you mentioned e.g. late 80s to early 90s). But clearly, we are in such a period now too.

And as I already said before, priorities simply changed in favour of lower power consumption and IGPs. Like it or not, but thats what the 99% crowd wants.

I've never said I disagree on that. We got lower TDP, iGPU, integrated Memory Controller, and so on which is all good.

But all I'm saying is that pure CPU performance increases have not been that great the last 4 years (5 if we include what Haswell is expected to bring). Because some people here claimed that to be false, and that the CPU performance has been as great as during e.g. 1997--2001. But so far nobody has been able to show benchmarks proving that to be true.

 

[ShintaiDK]

Pure performance increase until Core 2 was mainly due to the abuse of the power envelope. Performance/watt ratios in increase is alot better today than back then.

From 1985 to 1989 we got from a 12Mhz 386 to a 25Mhz 486. But we also started to see heatsinks. And it took another 2 years until 1991 before we had doubled that performance again. But again due to power cost. In 1993 we got the first Pentium. I dont know if you recall it, but it had overheating issues. It was before active cooling really hit. Again, performance/watt ratios was pretty bad. Active cooling then again raised the envelope on how much power you could feed the CPUs. And then extract the performance due to that.

 

[NTMBK]

Pure performance increase until Core 2 was mainly due to the abuse of the power envelope. Performance/watt ratios in increase is alot better today than back then.

From 1985 to 1989 we got from a 12Mhz 386 to a 25Mhz 486. But we also started to see heatsinks. And it took another 2 years until 1991 before we had doubled that performance again. But again due to power cost. In 1993 we got the first Pentium. I dont know if you recall it, but it had overheating issues. It was before active cooling really hit. Again, performance/watt ratios was pretty bad. Active cooling then again raised the envelope on how much power you could feed the CPUs. And then extract the performance due to that.

This is a very good point, and also reflects what we've seen in graphics cards for the last 10-15 years. It seems to be accepted as natural that graphics cards will need bigger and bigger coolers, and more and more power connectors. If we were limited to the same power envelope of a Geforce 4, we'd all be using HD7750s.

 

[ShintaiDK]

Also a reason why we hit the wall with the P4. And Bulldozer for that matter. The power envelope utterly abused all the way.

 

[Fjodor2001]

Looking at history, how's this for a golden period BTW:

Apr-72: 8008, 200 KHz
Apr-74: 8080, 2 MHz
Jun-78: 8086, 10 MHz

That's a massive 5000% frequency increase in just 6 years.

http://www.intel.com/pressroom/kits/quickreffam.htm#i486

 

[IntelUser2000]

Looking at history, how's this for a golden period BTW:

Apr-72: 8008, 200 KHz
Apr-74: 8080, 2 MHz
Jun-78: 8086, 10 MHz

That's a massive 5000% frequency increase in just 6 years.

http://www.intel.com/pressroom/kits/quickreffam.htm#i486

Back then the "CPUs were as simple as basic ICs you see in many electronics like TVs, Monitors, or whatever you name them.

It's true, various things have reduced performance gains. The first was because of the "Thermal Wall", where it became established that above 130W is a disaster for PCs. Then process scaling went down from reducing 50% power per generation, to 30%, or in some cases, less.

It's unlikely it'll change.

Also, I don't think its a huge mistake for the direction Intel is going in. They'll have to be a far smaller company if the focus stays on the Desktop road.

When you develop a process or architecture, you are bound to have a point where its really good at and some that's not so good. I wouldn't be surprised if Desktop chips are in the "not so good" portion for 22nm and beyond process.

 

[lopri]

Nehalem improved IPC also considerably if you're looking at todays applications and games. About 30% or so.

You do have to take into account platform changes for Nehalem. Native quad-cores with direct access to memory is big part of Conroe-to-Nehalem transition.

Nehalem was a fascinating chip, I agree on that. Unfortunately Intel botched it (for me and many others here) with exorbitant pricing on both the CPUs and the buggy platform. ($400 for motherboard effectively makes a $300 CPU a $500 one, no matter what reviewers say) The platform (x58) was also a furnace.

All of that led to low adaption and there weren't many places to discuss the inner workings of Nehalem (since not many people had it), so my interest subsided rather quickly.

 

[Lepton87]

You do have to take into account platform changes for Nehalem. Native quad-cores with direct access to memory is big part of Conroe-to-Nehalem transition.

Nehalem was a fascinating chip, I agree on that. Unfortunately Intel botched it (for me and many others here) with exorbitant pricing on both the CPUs and the buggy platform. ($400 for motherboard effectively makes a $300 CPU a $500 one, no matter what reviewers say) The platform (x58) was also a furnace.

All of that led to low adaption and there weren't many places to discuss the inner workings of Nehalem (since not many people had it), so my interest subsided rather quickly.

But Intel later fallowed with mainstream version in the form of lynnfield which frankly wasn't slower and was an even bigger die, but 4c/8t CPUs were still as expensive for 1156 as they were for 1366 if not more. I remember that i7 920 was cheaper then the least expensive i7 for 1156 at least in my country. I didn't care about the extra threads so I bought i5 750 which was quite affordable.