IPC comparison between Sandy Bridge and old CPUs?

[Dark_Archonis]

I've been looking deep into the technical info and benchmarks, but have had trouble figuring this out:

How does the IPC of Sandy Bridge compare, clock for clock, core for core, to some much older Intel CPUs such as Dothan, or Coppermine or Conroe?

I'm curious just how much of an IPC progression Intel has made in the past 10-12 years.

 

[BallaTheFeared]

http://www.tomshardware.com/charts/x86-core-performance-comparison/benchmarks,128.html

 

[Dark_Archonis]

http://www.tomshardware.com/charts/x86-core-performance-comparison/benchmarks,128.html

I've already seen that data before, and it's not what I'm looking for. Most of the old CPUs I mentioned are not listed there.

 

[BallaTheFeared]

OIC, sorry I couldn't help.

 

[Dark_Archonis]

Yeah, thanks anyways. I'm just curious percentage-wise, how much better clock for clock is Sandy Bridge versus the older Intel CPUs I mentioned?

I know there are some CPU gurus in here, hopefully some of them can shed light on this?

 

[BallaTheFeared]

It goes back to 2005 with the Prescott, Conroe is on there.

Anything before netburst wasn't hitting high enough clock speeds to really compare, since SB runs it's CPUNB and L3$ off the processor speed, downclocking it to 3GHz is actually lower IPC more than a straight clock speed drop would.

Pent3 would require an 800-1000MHz modern processor, and is pretty much a futile effort imo. IPC is only one factor, clock speed is just as important. Conroe to Bloomfield saw about a 17% increase in IPC, Bloomfield to SB saw another ~20% increase.

Later C2D/Q chips clocked just as well as most Bloomfields, some even better because C0 wasn't that great of a clocker. SB offers both the IPC increase we saw back then, as well as a healthy avg clock speed increase of about 500-700MHz.

 

[Dark_Archonis]

Well, since very few replies here, I decided to do the hard work myself. After some thorough analysis and calculations, here is what I have found:

At the very worst, clock for clock, thread per thread, Sandy Bridge outperforms Dothan by around 45%. At the very best, clock for clock and thread per thread, Sandy Bridge outperforms Dothan by around 90% or more. So roughly speaking, on average, clock for clock and thread per thread Sandy Bridge outperforms Dothan by around 67% (or more). On average means a mixture of regular business and productivity apps, multitasking, browsing, media performance, and gaming.

Now this is completely excluding clock speed differences, the fact that Sandy Bridge has Hyper-Threading, the fact that Sandy Bridge has a dedicated media processor for encoding/decoding media (in most configurations), and the fact that Dothan is single core while Sandy Bridge is mainly quad core.

In total performance, Sandy Bridge obviously outperforms Dothan by a factor of several hundred percent.

My numbers are not exact, but I think a fairly good representation.

To me the numbers are impressive overall. Even more impressive is the fact that each Sandy Bridge core is incredibly more power efficient compared to a single Dothan core, clock for clock. Much of that is due to a more advanced process, but much of that is also due to architectural advancements.

I didn't bother yet making the comparison to Coppermine, but the clock for clock performance difference between Coppermine and Dothan is not very big.

 

[fixbsod]

Maybe this is more up the poster's alley --

edit 4/9/12 -- I retested Cinebench on the P4 and actually got a whopping 0.27 !!! As 0.23 --> 0.27 is fairly significant I revised this post:

I have a P4 2.0GHz 'A' Northwood w/1GB RAM (2x512 MB PC400) which is single core/single thread and scores a 0.27 in Cinbench 11.5 (WXP Pro SP2). This was near top of the desktop line at the time -- came out 1/2002 and purchased 8/2002 when the fastest intel desktop CPU was a P4 2.53 GHz Northwood.

My i7-2700k running at 3.9 GHz (even using all cores will turbo this high) w/16GB RAM (4x4 GB 1600) which is 4cores/8threads will get apx 7.6 or so overall with all cores/threads and 1.66/core in Cinebench 11.5 (W7 Ult SP1). This is and was top of the desktop line at the time of purchase 10/2011, which was also its release date. Aside -- I've been tops for 6 months !!

1 core adjusted only for MHz results in a 0.85 score. On a 1 core to 1 core basis and adjusting for MHz results in a raw speed increase of 3.15x which I want to say is... subpar, adjusting again for non CPU improvements like OS and larger RAM would probably bring it down to 2.5 times as fast???

 

[IntelUser2000]

I didn't bother yet making the comparison to Coppermine, but the clock for clock performance difference between Coppermine and Dothan is not very big.

Where'd you get that? Turion 64(mobile chip based on Athlon 64) based on the Athlon 64 core running 20% higher frequencies than Dothan were only in average 10% faster: http://techreport.com/articles.x/9377/8

Athlon 64 itself was 30% faster than Athlon XP which in turn was faster than the Pentium III. In overall, even the original banias were 30-40% faster than Pentium III chips.

On a 1 core to 1 core basis and adjusting for MHz results in a raw speed increase of 3.7x which I want to say is... subpar, adjusting again for non CPU improvements like OS and larger RAM would probably bring it down to 3.3 - 3.5x ?

If you say so. But the 2700K has instruction set advantages along with having 4x the cores.

The original Core 2 Duo was about 1.9x the speed per clock of the Pentium 4. Penryn isn't significantly faster than Conroe, and Nehalem was a small advancement over Penryn in single threads. Sandy Bridge adds 10-15% in everything over Nehalem.

In addition, CPUs lose IPC as clock speed increase. Some advancements made by modern processors are merely to counter that. Cinebench score scales 85% by purely increasing frequency. If that Pentium 4 straight up doubled in clocks to 4GHz, it would get 0.43, which is 4x away from the 2700K at 3.9GHz.

Single thread throughput was once the sole determining factor for CPU performance, but nowadays is merely one factor out of many. Even the low threaded applications today can run 2 threads. If Nehalem came out back in 2002, it would have been less than 10% faster than its predecessor simply due to programs not being able to take advantage of 4 cores + Hyperthreading.

 

[peonyu]

I dont have IPC comparisons but would like to see some. In general though its safe to say that on a straight mhz basis it goes like this-

SandyBrige > Core 2 series > AMD Phenom series > Athlon 64 > Athlon XP > Pentium 3 > Pentium 4


Where Bulldozer fits in Idk, but I would guess between the Phenom and A64...If its IPC is less than that then I would be surprised.

 

[Bman123]

This is what I gathered online and my own testing. Celeron g530 is faster then a e6600, about equal to a e7300. Pentium g620 is equal too e8400-8500. Sorry but that all the info I can give from my personal experience as I've owned all those chips. Just off a guess I would say the i3 2100 I have has to be close to a q6600 too.

 

[OVerLoRDI]

This might help. Not sure how much information about older cpus is in this thread. But the data from this thread combined with older data that you might be able to find could answer your questions.

http://forums.anandtech.com/showthread.php?t=2057154

 

[Dufus]

You can get some idea of IPC improvents here section 4, but perhaps your more interested in gains through newer instructions in which case that will be much more software specific. i.e. running Linpack SSE vs AVX can show a ~90% increase in performance.

 

[taltamir]

I'm curious just how much of an IPC progression Intel has made in the past 10-12 years.

IPC does not mean performance per clock (MHz). It means Instructions per Clock (MHz).
What you want to know is Performance/Watt and Performance/Clock. Not Instructions/Clock.

 

[Magic Carpet]

SandyBrige > Core 2 series > AMD Phenom series > Athlon 64 > Athlon XP > Pentium 3 > Pentium 4

Tualatin [especially Celly] gave Athlon XP a good run for its money. They traded places, depending on the benchmark used. Clearly, at 1.8+ Tualatin w/ 512kb would beat any offering of its day. Too bad, Intel had scraped those plans and never released anything past 1.4 Ghz.







The above LinX benchmark should give you an idea... how things have changed in the past decade. Lack of instruction sets and... low memory bandwidth play the major role in such a low score.

End-of-Life date Last order date for embedded processors is July 30, 2012
Last shipment date for embedded processors is Jan 28, 2013

Surprisingly, according to cpu-world.com, Intel still ships them in embedded solutions.

 

[IntelUser2000]

Clearly, at 1.8+ Tualatin w/ 512kb would beat any offering of its day. Too bad, Intel had scraped those plans and never released anything past 1.4 Ghz.

1.4GHz might have been a real limit. On 0.18 micron, they pushed really hard to get the 1GHz part out, and at 1.13GHz it basically blew up. 80% frequency gain from process even back then was pushing it, and that's without considering 1GHz was already at the limits. Banias reached 1.6GHz on 0.13 micron but only with pipeline modifications.

Tualatin [especially Celly] gave Athlon XP a good run for its money.

Athlon XP is a modification of Thunderbird Athlon, the one which was compared. It was a pretty good advancement considering it was on the same 0.18 micron process and was 7-10% faster.

 

[Vesku]

Yeah, my dual Pentium III 1GHz machine lasted a loooong time. No surprise from me that they went back to the pre-P4 architectures for inspiration.

Intel has managed to power through on the IPC front so far, at least a generation longer of 10 to 15% gains than I thought we'd see. If only they'd let us OC these SB Celerons and Pentiums. Wonder if Haswell will keep the IPC improvements coming... have a feeling we will see more transistors dedicated to specific accelerations as overall IPC improvements get tougher and tougher (as we've seen happen with Video and Encryption CPU engine blocks).

 

[IntelUser2000]

Intel has managed to power through on the IPC front so far, at least a generation longer of 10 to 15% gains than I thought we'd see.

Pentium III Coppermine review
http://www.anandtech.com/show/399/7

The gains don't seem better than they do with recent architectures. I've also looked at Pentium II benchmarks, the gains are similar as Tock today.

The difference between then and now is that back then in-between gains happened with clock speed and FSB increases. Now its with more cores, ISA enhancements, and accelerators.

 

[Cogman]

You're not going to get solid IPC numbers because IPC is pretty much meaningless.

The problem with IPC is that CPU architectures have been constantly evolving. To the point that you can't really say that a multiplication instruction will take X amount of clocks and a add will take Y clocks. Everything depends on how what instructions were ran beforehand, how the CPU decides to do the ordering, and if some portion of data falls in the cache or not (and what level of cache does it fall into?)

Add to this mess the fact that the program doing this measurement can be interrupted at any time and you've got a real mess on your hand.

Better measurements of CPU speed increases can be found by benchmarking the same application on the old and then the new CPU.. Even that has its problems, if the application isn't programmed to use new CPU instructions it could easily give a lower performance than is available.

The only thing we can absolutely say about the IPC is that it most likely has gone up (Even though it did go down during the P3->P4 era).

 

[Tempered81]

Some gaming benches might help u