4th Generation Intel Core, Haswell summarized

[Idontcare]

It's pretty big if you believe this: http://assets.vr-zone.net/15272/dieimage.jpg

The amount of effort they are using to push Ultrabooks, there must be more reason than just cost for having only 30% gains despite having GT3.

I beleive it but that must mean they planned to use legacy process node 1Gb parts (60nm or older) for the 1Gb die to be that large.

Even on an older legacy node like 48nm, a 1Gb (128MB) die weighs in at 29mm^2.

Shrunk and produced on a more modern 30 nm or 2x nm process and that 128MB L4 dram cache should be <10mm^2 and silly cheap (<$1).

The reasons not to do it should come down to risk to timeline concerns (package level complexity and interposer validation) and not a question of there being enough space under the IHS. The socket itself would have been designed long long ago to be able to accomodate the interposer footprint if Intel really had plans to bring such a thing to the market.

 

[Revolution 11]

For general use? Probably not. If you have a specific program that will benefit from Haswell's optimizations? Probably.

I'm betting that emulators compiled to take advantage of AVX2 will really fly on Haswell, for instance. Dolphin already makes good use of AVX.

What? I read last year that Dolphin would not benefit from AVX substantially. At best, the reaction is mixed on supposed benefits.

http://forums.dolphin-emulator.com/showthread.php?tid=14439&page=2

Now Dolphin may have progressed beyond that build in terms of benefiting from AVX, but that upperbits penalty would still be in place. (not that I know anything about this) Perhaps BenchPress can elaborate on this.

 

[Sweepr]

Increased IPC: http://www.xbitlabs.com/images/cpu/haswell-uarch-idf/SF12_ARCS001_10005.png

 

[ShintaiDK]

Increased IPC: http://www.xbitlabs.com/images/cpu/haswell-uarch-idf/SF12_ARCS001_10005.png

:thumbsup:

 

[NTMBK]

Increased IPC: http://www.xbitlabs.com/images/cpu/haswell-uarch-idf/SF12_ARCS001_10005.png

No scale graphs are useless. For all we know that could be a log scale.

 

[Idontcare]

No scale graphs are useless. For all we know that could be a log scale.

It is pretty obvious what the graph relates to though. Blue are ticks, red are tocks.

The blue for 2013 is for Haswell. The prior blue before that is for SB, the intervening red is for IB. The blue bar third from the left is for Nehalem, and the first blue bar is for Conroe.

We know what the IPC improvements were for each successive tick going from Conroe to Nehalem to SB - it was ~7-10%.

The blue bar that we are to interpret as being Haswell for 2013 is roughly the same increase (percentage-wise) as the prior blue bars in the progression in comparison to the prior red bar (IB) but the gap to the prior tock (SB) is surprisingly higher than the gap between prior ticks.

Also we see in this graph part of the reason why Intel referred to IB as a "Tock+" as the red versus blue increase for IB over SB is notably higher than the red vs blue gap for the prior tocks over their ticks.

Assuming that Nehalem averaged a 10% increase in IPC over Conroe (Merom) and that SB averaged a 7% increase in IPC over Nehalem yields a value of roughly "305" arbitrary units (average of 299 and 311) of scale for the Y-axis offset.

Plugging back in the arbitrary offset yields an expected Haswell IPC increase of 12% over SB. Inline with expectation.

 

[inf64]

^^ And that's what intel states also, ~10% over SB. But lets not forget that we will see huge gains if/when the workload supports AVX2/FMA.

 

[Sweepr]

I'm personally expecting ~10% IPC jump over IB and equal or slightly higher clocks vs IB (100-200MHz).

That puts Haswell in line with Anandtech statement about ~10% increase in performance (average) over IB, with possible gains in the low double digits.

 

[mikk]

7% is wrong, Sandy improved IPC over Nehalem by 10-15%.

 

[Riek]

I'm personally expecting ~10% IPC jump over IB and equal or slightly higher clocks vs IB (100-200MHz).

That puts Haswell in line with Anandtech statement about ~10% increase in performance (average) over IB, with possible gains in the low double digits.

I expect lower frequency for hasswel. (possibly a higher max turboboost for one core though).
And a higher ipc gain than 10%. Making the overall result ~10% due to the lower clockspeeds.

 

[exar333]

7% is wrong, Sandy improved IPC over Nehalem by 10-15%.

That's a little high. Check out Tom's clock for clock comparisons. 7-10% is probably pretty accurate for most gains. Some more, some less.

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

 

[mikk]

That's a little high. Check out Tom's clock for clock comparisons. 7-10% is probably pretty accurate for most gains. Some more, some less.

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

http://ixbtlabs.com/articles3/cpu/sandybridge-core-vs-lynnfield-p1.html
http://www.computerbase.de/artikel/prozessoren/2011/test-intel-sandy-bridge/47/
http://www.hardware.fr/articles/815-15/lynnfield-vs-sandy-bridge-2-8-ghz.html


= +10-15% in comparison with a Nehalem Quadcore (Bloomfield or Lynnfield) compared to Sandy Bridge at the same frequency.

 

[BallaTheFeared]

That's a little high. Check out Tom's clock for clock comparisons. 7-10% is probably pretty accurate for most gains. Some more, some less.

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

Key word is games, Intel is stating over a mixed workload, over a mixed workload SB has a better instruction capability that enhanced it's ICP gain. I was more akin to 13-17% myself over mixed load types.

13% is pretty "meh", SB was awesome because the IPC increase was met with a huge clock rate increase over what 45nm chips were capable of doing along side a drastic increase in perf/watt. Allowing both high clocks and great performance out of cheap coolers like the 212+ which is now one of the highest recommended coolers and we have to conclude it is so because of how well SB performed.

Unless Intel can get the clock speeds up Haswell at least on forums like this is going to end up getting the "Ivy Bridge" treatment.

 

[AtenRa]

HW is a nice APU, it is time for the $300 6-core socket 2011 in 2013.

 

[Nemesis 1]

HW is a nice APU, it is time for the $300 6-core socket 2011 in 2013.

Don't even start this crap . HW has no APU and never will have. That AMDs name for a cpu with graphics . AMD will not be naming intel devices not now not ever. APU is = anything AMD/ATi have named meaningless . ATI called GPUS VPUs . Ya we all know wghat a VPU is . and thats wonderful but we use the GPU term the market lleader NV gave us. AMD is no market leader never has been never will be . So calling an Intel CPU an APU is incorrect and willl never stand . Debate it with intel

 

[BallaTheFeared]

Don't even start this crap . HW has no APU and never will have. That AMDs name for a cpu with graphics . AMD will not be naming intel devices not now not ever. APU is = anything AMD/ATi have named meaningless . ATI called GPUS VPUs . Ya we all know wghat a VPU is . and thats wonderful but we use the GPU term the market lleader NV gave us. AMD is no market leader never has been never will be . So calling an Intel CPU an APU is incorrect and willl never stand . Debate it with intel

lol

 

[Idontcare]

7% is wrong, Sandy improved IPC over Nehalem by 10-15%.

Give me a specific number (not a range) to use for SB over Nehalem, as well as a specific number for Nehalem over Conroe. I will gladly recompute the value for Haswell based on Intel's published results.

 

[Nemesis 1]

Key word is games, Intel is stating over a mixed workload, over a mixed workload SB has a better instruction capability that enhanced it's ICP gain. I was more akin to 13-17% myself over mixed load types.

13% is pretty "meh", SB was awesome because the IPC increase was met with a huge clock rate increase over what 45nm chips were capable of doing along side a drastic increase in perf/watt. Allowing both high clocks and great performance out of cheap coolers like the 212+ which is now one of the highest recommended coolers and we have to conclude it is so because of how well SB performed.

Unless Intel can get the clock speeds up Haswell at least on forums like this is going to end up getting the "Ivy Bridge" treatment.

Intel isn't competeing with AMD anymore . Intel is competing with ARm cpus and China cpus . Yes China has cpus and they are growing in usage. AMD has nothing for intel. I would very much like to see the link were you seen Intels clocks for haswell desktop. Your insane iff you believe IB is getting the cold shoulder from anyone . SB is about the same and most that are upgrading did so already with SB . But if you build now a system choosing SB over Ib be a retard move.

Haswell brings so much more than SB/IB its not even funny until one reads post like yours. You do understand a 15% increase of a HW processor would require about 22% increase in IPC on the next AMD release to stay in step with intel aDVANCES. iTS ACTUALLY HIGHER BUT DON'T WANT THE BABIES CRING.

 

[BenchPress]

Plugging back in the arbitrary offset yields an expected Haswell IPC increase of 12% over SB. Inline with expectation.

Who's expectation?

Haswell increases the number of execution ports by 33%, and on top of that it also grows the buffers used by out-of-order execution by over 10%. Even taking into account that extracting more ILP gets increasingly harder, the expectation would be to gain more than 10% in IPC.

So there appears to be something else which keeps the IPC gain modest. Also, increasing the number of execution ports typically sacrifices clock frequency, due to requiring extra time to pass the results around. But with an IPC gain of only 10%, that doesn't seem like a smart gamble when you're on the winning team. So I suspect these two to be connected...

To keep clock frequency high while increasing the number of ALUs, all the while keeping power consumption low, they may have increased the latency of certain instructions. The 64-bit arithmetic instructions seem like a good candidate, because even 64-bit software mostly consists of 32-bit calculations, and the 64-bit address calculations are handled by the AGUs, not the generic ALUs.

This compromise would explain why the IPC gain is lower than expected, and gives overclockers something to look forward to.

 

[jones377]

Give me a specific number (not a range) to use for SB over Nehalem, as well as a specific number for Nehalem over Conroe. I will gladly recompute the value for Haswell based on Intel's published results.

You know better than this IDC. But I suppose you could take all applications ever made, weigh them by the number of users for each and then average the IPC improvement. Anything else and you'll get a different number depending on who you ask, when you know the real number is 42.

 

[Nemesis 1]

Give me a specific number (not a range) to use for SB over Nehalem, as well as a specific number for Nehalem over Conroe. I will gladly recompute the value for Haswell based on Intel's published results.

The biggest problem is gaming benchmarks . You don't add in a bunch of game benchmarks to get an average . you have to average all the games and that average is the IPC increase in gaming . than you add that average to other benchmarks you average . Uaing gaming benchmarks is kinda phony anyway as the gpu is the biggest metric in those benchmarks unless your willling to use low res. results to remove the gpu as the bottle nick . Lets average game benchmarks IB against BD you use what ever bench mark you want I will use low res see what kind of a differance we come up with . If your going to choke IB with todays GPus to get your % that a cheat we all know its a cheat.But its the only thing AMD fanbois will except in their fantasy world.When the present agreement between NV/Intel runs out I believe 3 more years Intel will know longer offer a PCi-E slots / If you want NV you will have to buy an nv cpu . For PHi intel will bring that ondie or use another socket . You will not see NV accelrators on Intel server products .

 

[Bat123Man]

Good thing the GF wanted a desktop so I can give her the IB while I get Haswell next year



Dumped discrete HTPC cards when I got the i5-661. Never missed it since.

Hi ShintaiDK,

What are the specs of your HTPC with the i5-661 if you don't mind me asking? And what do you use it for? Netflix, BluRay, Matroska ? Any casual gaming?

Have been using an old original XBOX as my HTPC, but it can't do BR or HD above 720p. Also have a PS3, but it doesn't handle all formats. Time to build an HTPC, but only if I can build a fanless one, lower power the better. If that i5-661 can handle the load without active cooling or a discrete GPU to help it, that sounds ideal.

Tks,
BM.

 

[inf64]

lol

Balla you forget the rule number one man, do not feed the trolls! And you tossed a decent meat chunk to godzilla of that kind .

 

[sm625]

Knowing intel, they will try to charge $478 apiece for the minimum SKU, thus keeping tablets priced well over $1000 and thus continuing to bleed massive profits to apple and google.

 

[VirtualLarry]

When the present agreement between NV/Intel runs out I believe 3 more years Intel will know longer offer a PCi-E slots

Oh really. I suggest you look up the history of the MCA bus.