Haswell 4c8t @2.6 Fritz Chess Benchmark <Update:2.8G Haswell>

[inf64]

It's goin to be ever more difficult to milk more IPC from modern day x86 cores. We are near the point of diminishing returns since adding more HW at the problem only helps so much. What we will have though are 3 things: moar cores(sic) , new ISA and better software support for HW functions( ie. FMA,AVX2,HSA etc.) and clock increases. The IPC will surely go up but not by leaps and bound as we saw in the past. Especially if the current performance mark is set up very high like in the case of newest IB core. It would be more easier if the bar was lower,like in case of PD so you can have some decent gains with removing obvious bottlenecks from your design. IB has no obvious bottlenecks so it's just a matter of tweaking the design further and trying to milk as as much as possible with relatively moderate HW investments ( Haswell x86 cores won't be much bigger than IB ,nor much more complex- intel invested heavily in power delivery and on die VRMs though).

 

[meloz]

It's goin to be ever more difficult to milk more IPC from modern day x86 cores. We are near the point of diminishing returns since adding more HW at the problem only helps so much. What we will have though are 3 things: moar cores(sic) , new ISA and better software support for HW functions( ie. FMA,AVX2,HSA etc.) and clock increases. The IPC will surely go up but not by leaps and bound as we saw in the past. Especially if the current performance mark is set up very high like in the case of newest IB core. It would be more easier if the bar was lower,like in case of PD so you can have some decent gains with removing obvious bottlenecks from your design. IB has no obvious bottlenecks so it's just a matter of tweaking the design further and trying to milk as as much as possible with relatively moderate HW investments ( Haswell x86 cores won't be much bigger than IB ,nor much more complex- intel invested heavily in power delivery and on die VRMs though).

We have clearly evolved from focus on pure performance to performance / watt. Intel just reach there before everyone else in the industry because they lead the manufacturing process (and also because x86-64 is so refined and mature), and thus Intel also face all the difficulties first.

As a consumer, I am mentally prepared for the fact that in near future we might not get more than 5% performance improvement (on CPU side) from one generation to next, yet gain 15% or more improvement in system performance / watt terms. I see no reason to pretend sky is falling.

Common consumers are yet to latch onto the performance/watt concept, the datacenter segment does so intuitively since energy consumption (for powering equipment as well as cooling it) is their major cost.

 

[Idontcare]

Sadly AMD always gets raped in the end

That was unnecessary.

I agree, SB - IB was a side grade with lower OC potential and higher temps, I hope haswell is more than a lowly 10% faster.

Its not all that different than what happened with 45nm penryn over 65nm conroe. Anyone who had a G0 Q6600 was better off keeping it versus attempting to "upgrade" to a Q9xxx because the OC potential was not really any better.

The bottom line in my mind is that people's expectations are unjustifiably inflated when it comes to year-on-year improvements that can come from refining a product that has been persistently refined already for some 30+ yrs.

 

[ShintaiDK]

People got unrealistic expectations to IPC increases. The only real way to increase IPC today is new instructions like AVX2.

 

[HutchinsonJC]

Weren't part of the expectations of Haswell toying with the idea of more room to OC and not just IPC improvements?

 

[pantsaregood]

10% improvement in the case of this benchmark is nice considering SB-IB only yielded less than a 1% improvement. Also of note is that Haswell, as a result of Intel's tick-tock approach, is being produced on a mature process node - unless there were some major architectural changes (doesn't look like it), it should overclock at least as well as Ivy Bridge.

 

[Remobz]

For people still with an i3 proc Haswell is looking mighty good

 

[RussianSensation]

Nehalem to SNB was roughly 10% if I recall correctly. SNB to IVB is another 5%.

Nehalem to SB is more than 10%.

Final Performance Rating in 22 benchmarks (top chart)

i5 760 2.8ghz => 100%
i5 2500K 2.8ghz => 114% (+14%)

i7 930 2.8ghz => 100%
i7 2600K 2.8ghz => 114% (+14%)

i7 930 2.8ghz with HT => 100%
i7 2600K 2.8ghz with HT => 113% (+13%)

The average from Nehalem/Lynnfield to SB seems to be 13-14% increase in IPC in apps/games.

iXbtlabs put i5 760 against i5 2500K both clocked at 2.8ghz as well and SB won by 14% on average:
http://ixbtlabs.com/articles3/cpu/sandybridge-core-vs-lynnfield-p2.html

Haswell's 10% IPC increase over SB would be the worst since any previous Intel generations starting with Nehalem. Since Intel hinted at a 10% IPC increase, I guess they didn't set the expectations that high. 10% and hopefully superior overclocking to IVB is still great in 2 years time given the rate of progress in the CPU space has been slowing down. I would preferred a 6-core Haswell though at $325 instead.

 

[AtenRa]

The ~10% in this benchmark(Fritz Chess) is not IPC but Performance over Ivy. The bech was run with 8 threads not a single thread.

IPC could be lower or higher than 10%. If Thread Scaling is higher than Ivy then IPC may be lower than 10%.

 

[mikk]

What IPC increase in Fritzchess from Nehalem to Sandy Bridge have we got?

 

[Idontcare]

<- slaps forehead and exclaims "doh!"

The ~10% in this benchmark(Fritz Chess) is not IPC but Performance over Ivy. The bech was run with 8 threads not a single thread.

IPC could be lower or higher than 10%. If Thread Scaling is higher than Ivy then IPC may be lower than 10%.

This is so damn spot on, I feel silly, can't believe I totally overlooked this :headinshame:

You are absolutely right. The 10% performance improvement is convoluted, we cannot rightly claim to attribute it to be an IPC improvement without first determining how much of the performance increase comes from increased multi-threading efficiency that comes into play in the perfomance scaling of all multi-threaded apps (the one part that is hardware-dependent, the Almasi/Gottlieb portion).

In order to know the single-threaded IPC improvement of Haswell over IB we need the Haswell bench to be ran with one thread (easy to do from the application side of things, but we don't have access to Haswell to run the test )

 

[parvadomus]

The ~10% in this benchmark(Fritz Chess) is not IPC but Performance over Ivy. The bech was run with 8 threads not a single thread.

IPC could be lower or higher than 10%. If Thread Scaling is higher than Ivy then IPC may be lower than 10%.

That benchmark is pointless to know actual IPC. It's mainly a branch prediction test (CPUs will mispredict a lot in this test), that can be influenced by two things:
- Branch prediction capability
- Pipeline length
So haswell might have a shorter pipeline, or just better branch prediction (I think this, it should have larger structures to learn loop behaviors).
We need more benchmarks with real workloads to really know how much haswell has improved over SB/IB.

 

[AtenRa]

That benchmark is pointless to know actual IPC. It's mainly a branch prediction test (CPUs will mispredict a lot in this test), that can be influenced by two things:
- Branch prediction capability
- Pipeline length
So haswell might have a shorter pipeline, or just better branch prediction (I think this, it should have larger structures to learn loop behaviors).
We need more benchmarks with real workloads to really know how much haswell has improved over SB/IB.

IPC is application depended and not constant. IPC may be lower in Fritz Chess than another application but the comparison is still valid between processors.

I do agree that we need more apps to really have an idea on the performance gains over Ivy.

 

[Acanthus]

Unless maximum OCs rise I will sit out yet another CPU generation.

It's not like AVX2 apps are going to be coming out any time soon.

 

[Zargon]

more and more thinking I'll be doing a jump to haswell-e and HD 8000 at the same time, in a bout 18 months or so

 

[Edrick]

People got unrealistic expectations to IPC increases. The only real way to increase IPC today is new instructions like AVX2.

Winner Winner!

 

[Hulk]

Most of my applications run on a single core of my 2500k just fine. I'm talking about browsing the web, word processing, spreadsheets, e-mail, financial software, etc.. probably what makes up 90% of what I do. That is NOT to say that IPC and clockspeed, and more compute in general, aren't always welcome. But I'd like an 8 core/16 thread consumer oriented processor that can bring those cores on line when I do some video editing, audio editing, photo editing or other apps which can make use of the additional cores.

It seems like 14nm would make this possible from both a cost and thermal standpoint. The issue, as always, is Intel doesn't want to savage it's high end parts. We just have to pray for 12/24 core and 16/32 core high end parts, which would allow them to bring 8/16 into play in the consumer marketplace.

 

[NTMBK]

Winner Winner!

Yes, they'll see a really great IPC boost on the applications of 2017!

 

[TuxDave]

People got unrealistic expectations to IPC increases. The only real way to increase IPC today is new instructions like AVX2.

Put me under "I disagree". Just how AVX2 only benefits certain types of computation benchmarks, there are other architectural features that can skyrocket other types of traces. And it can be done without additional instructions or additional ALUs. On top of that, remember instruction level is just one level of abstraction. Consider the uop flows for complex instructions that can be improved. Same instruction... just faster.

I have never run SPECfp but I do know it composes of many FP intensive traces and I guess the final score is a big average. Has anyone attempted to see the speedup of the individual traces? You might be surprised.

Edit But in case anyone wanted to nitpick. Yes, SOME traces need new instructions to improve. When a certain trace is well understood and therefore very optimized for the CPU to minimize misprediction, stalls, etc.... then you start hitting the point that IPC improvements is going to require some creativity.

 

[FlanK3r]

People got unrealistic expectations to IPC increases. The only real way to increase IPC today is new instructions like AVX2.

Im agree . Or cut some bugs in first relase of architecture. The future could be in combination CPU+iGPU power (for some aplications more cores/threads). But Trinity is only little baby in diapers. Maybe after 5-10 years....

 

[videogames101]

anyone with more data than me want to analyze those voltage values?

higher/lower than IB at the same clock?

 

[Idontcare]

anyone with more data than me want to analyze those voltage values?

higher/lower than IB at the same clock?

Which values? The idle voltages in the screencaps of the OP?

 

[RussianSensation]

Which values? The idle voltages in the screencaps of the OP?

IDC noob question. Why do Intel's CPU need 0.889 voltage in 2D mode to run at just 800mhz? Do they really need that much voltage just to sustain 800mhz clocks? Also, hwy can't they let the CPU drop to just 200mhz at 0.20V and turn off the remaining 3 cores fully. I mean in 2D, who needs 4 cores running at 800mhz and 0.9V?

 

[Idontcare]

IDC noob question. Why do Intel's CPU need 0.889 voltage in 2D mode to run at just 800mhz? Do they really need that much voltage just to sustain 800mhz clocks? Also, hwy can't they let the CPU drop to just 200mhz at 0.20V and turn off the remaining 3 cores fully. I mean in 2D, who needs 4 cores running at 800mhz and 0.9V?

They don't need that much, not by a long shot. My 3770k is LinX stable at 1.6GHz with a mere 0.636V.

So cut that clockspeed in half and put the cores to idle and you should be looking at a heck of a lot lower Vcore than 0.636V.

Why 0.889V? I can think of some reasons - all having to do with ease of binning (economics) and engineering margin - but first my suspicion falls to whether or not that CPUz value is correct to begin with.

 

[Tsavo]

It's not like AVX2 apps are going to be coming out any time soon.

I'm still waiting for excellent AVX ones. :|