Ivy Bridge vs Haswell

[cbn]

The mainstream CPUs have been "stuck" at quad-cores for too long: Q6600 => Q9400 => i7-860 => i7-2500. Its time to move on to eight cores! Too bad its not until Haswell.

Trouble is mobile seems to be the standard for consumer computing now and ARM doesn't seem interested in fanning out with more cpu cores for smartphones.

Rather they seem more interested in pursuing single threaded performance with lower core counts. (with measures like "big.LITTLE" throw in to support that effort)

EDIT: Therefore It is hard for me to imagine most mainstream software (browsers, Office, etc) taking advantage of octo-core and beyond for a long time. Why program for hardware that very few people are likely to own in the near future?

 

[OVerLoRDI]

EDIT: Therefore It is hard for me to imagine most mainstream software (browsers, Office, etc) taking advantage of octo-core and beyond for a long time. Why program for hardware that very few people are likely to own in the near future?

Adding more cores generally is a case of diminishing returns, except for in cases where loads can be distributed easily across many cores.
1 -> 2 cores: huge increase in "snappiness" of day to day tasks and real world performance
2 -> 4 cores: moderate increase in "snappiness" of day to day tasks and real world performance
4 -> 8 cores: basically zero increase in "snappiness"

What the average users cares about is how fast their device/computer "feels". Throw 16 cores in your low end computer and the difference is completely lost on that user. More cores is awesome for those who need it but ultimately there will likely never be a point in optimizing MS word to run 4 threads.

 

[toyota]

I would love 6 or 8 real cores for my next cpu though. my 2500k hits 80%+ usage in a few games and even gets fully pegged in rare occasions. I do not like HT at all though and it looks like the top Ivy Bridge 1155 cpu will indeed be a 4 core with HT.

 

[BrightCandle]

The algorithms we use for multithreading basically fall into 3 general categories:
1) Scales near linearly with cores. So if we double the cores then performance should double as well. Unfortunately normally some percentage of the time has to be single threaded so even if an algorithm was 99% parallel with infinite CPU cores you can only go 100 times faster because the serial part will eventually dominate the time.

2) The calculation falls into a tree producing a log(n) response in increasing performance. Consider for example adding up millions of numbers in parallel, the optimal is to produce a binary tree and add leaves together and then adding the partial sums of those on and upwards. The maximum performance in that case comes from having N/2 cores, but they are all only used for 1 step of the algorithm and each level of summation reduces the parallel performance until it only uses 1 core. Although adding is too simple as thread intercommunication would dominate but you hopefully get the idea.

3) The third type is an algorithm where there is no known parallel algorithm. These are doomed to be serial. Consider for example the problem of encryption. Most systems today encode the beginning and then use the result of the first block of encryption to encrypt the later parts of the file. You can't load the entire file in and work on encrypting each part separately, you need to use earlier calculations and can make no progress before they have been calculated.

You equally can't make the following code parallel because D depends on the calculation of A.
A=B+C
D=A+1

In practice we are seeing a number of programs benefit from increased core counts, but even they are starting to suffer diminishing returns already. Video encoding is one of those activities that is often similar to encryption because we are looking for similarity in the image so you can represent that with a smaller amount of bytes. A lot of the current encoders are struggling to scale linearly even on moving from 2 to 4 cores. Sure they are getting faster but its likely a (1) type algorithm with the serial part being quite large, or potentially more like a (3) but with a bit of parallel work mixed in.

But with the silicon not scaling to high clock speeds and the additional transistors not really able to offer additional algorithmic performance how should they be spending the every increasing transistor budget? Right now it seems to me modern CPUs are cache memory with a 4 small CPUs attached. By percentage the amount of space the cores are taking is ever decreasing.

 

[tynopik]

how should they be spending the every increasing transistor budget?

gpu

 

[zlejedi]

I would love 6 or 8 real cores for my next cpu though. my 2500k hits 80%+ usage in a few games and even gets fully pegged in rare occasions. I do not like HT at all though and it looks like the top Ivy Bridge 1155 cpu will indeed be a 4 core with HT.

This assumes games will use those 6-8 cores.

Which is unlikely to happen unless next gen consoles have this many cores.

Just look how many years it took from Lost Planet and Q6600 (which we can consider first affordable quad core for masses) release to widespread use of game engines that use those cores.

And even now a lot of stuff doesn't go above 2 cores or uses like 20-30% from 3rd/4th vcores.

 

[toyota]

This assumes games will use those 6-8 cores.

Which is unlikely to happen unless next gen consoles have this many cores.

Just look how many years it took from Lost Planet and Q6600 (which we can consider first affordable quad core for masses) release to widespread use of game engines that use those cores.

And even now a lot of stuff doesn't go above 2 cores or uses like 20-30% from 3rd/4th vcores.

a few games can already use 6 cores but of course they do not really need those 6 cores. my point is that I have a 4 core cpu now and it already gets fairly close to fully utilized at times so I see no point in getting anything other than 6 or 8 cores when I get a new cpu.

 

[OVerLoRDI]

a few games can already use 6 cores but of course they do not really need those 6 cores. my point is that I have a 4 core cpu now and it already gets fairly close to fully utilized at times so I see no point in getting anything other than 6 or 8 cores when I get a new cpu.

Do you have examples?

 

[zlejedi]

a few games can already use 6 cores but of course they do not really need those 6 cores. my point is that I have a 4 core cpu now and it already gets fairly close to fully utilized at times so I see no point in getting anything other than 6 or 8 cores when I get a new cpu.

Or you could just go with 4 core that will have let's say 10-15% higher IPC with 10-15% more overclocking headroom due to mature 22nm process vs 32nm used in SB.

For me the most important lesson from a year when i had Thuban X6 was that less good cores >> many slow cores.

 

[cbn]

This assumes games will use those 6-8 cores.

Which is unlikely to happen unless next gen consoles have this many cores.

If Nvidia has their way the CPU cores race probably won't happen for gaming.

Google will be the "Game Console" and some type of "Denver" chip will the SOC powering it. <----I am expecting this to be Nvidia's attempt to put mostly GPU cores, rather than extra CPU cores, on die.

 

[cbn]

Intel Creates $300 million ultra book fund

Based on that, my guess is that "quad core" will remain the standard for the masses for many years to come. (Intel is responding to demand for smaller devices, rather than demand for Tower computers with more cores)

 

[TuxDave]

Adding more cores generally is a case of diminishing returns, except for in cases where loads can be distributed easily across many cores.
1 -> 2 cores: huge increase in "snappiness" of day to day tasks and real world performance
2 -> 4 cores: moderate increase in "snappiness" of day to day tasks and real world performance
4 -> 8 cores: basically zero increase in "snappiness"

What the average users cares about is how fast their device/computer "feels". Throw 16 cores in your low end computer and the difference is completely lost on that user. More cores is awesome for those who need it but ultimately there will likely never be a point in optimizing MS word to run 4 threads.

Funny thing is that I wish this laptop right now had 4 cores. Who knew my jumbo excel sheet would easily peg a 2 core processor?

 

[cbn]

Personally I'm not interested in more cores, and I'm also not so much interested in having a laptop CPU in my desktop (sub-95W).

If AMD and Nvidia can figure out a way to cool >300W GPU's then I'm ready for Intel and AMD to open up a product lineup that goes there too.

And do it the good old fashion way, keep it to 4-6 cores and just give me some nice high clockspeeds. With configurable TDP, if I don't want 300W then I can set it to 95W and have the clockspeed throttle itself accordingly.

My humble guess is that Companies licensing ARM CPU would need to start their own binning process in order to force pressure back on Intel.

(Although I wonder what kind of TDPs we would be talking about for these future Intel "ultra bins" you are referring to ? The trend seems to be towards lower TDPs. Maybe the top power consumption number would actually be somewhat reasonable in light of future process techs?)

 

[toyota]

Or you could just go with 4 core that will have let's say 10-15&#37; higher IPC with 10-15% more overclocking headroom due to mature 22nm process vs 32nm used in SB.

For me the most important lesson from a year when i had Thuban X6 was that less good cores >> many slow cores.

yes I am aware of that as you can tell since there is a 2500k in pc not a Phenom X6. again though for me to upgrade, its going to have to be more than just a tiny bump in performance on yet another 4 core cpu.

 

[LOL_Wut_Axel]

yes I am aware of that as you can tell since there is a 2500k in pc not a Phenom X6. again though for me to upgrade, its going to have to be more than just a tiny bump in performance on yet another 4 core cpu.

This. $180-350 Intel quad-cores are getting very long in the tooth. We've been using them for five years now; it's time to push higher.

 

[edplayer]

Ok so now its 227w a little better than the 270w you said in this thread it pulls at 5 ghz.

you might have a llc or pll voltage set high for it to pull that many watts,I will hook up a meter to my atx 12v line and ill give you my computer if it pulls anywhere near 227 or like you first said in this thread 270 watts.

Ill post up a video in a few days,just ordered a ac power meter where Ill plug the computer into it and see total power draw at 5ghz etc.

I know the software is not 100% spot on but its not 150 watts off

,I will hook up a meter to my atx 12v line and ill give you my computer if it pulls anywhere near 227 or like you first said in this thread 270 watts.

I doubt this is going to happen but let us see...

 

[IntelUser2000]

Haswell is being reported as a 30&#37; reduction in power as compared to Sandy Bridge,

30% reduction in STADNBY POWER, not ACTIVE POWER.

If Haswell will be 30% more power efficient than Sandy Bridge (~67W TDP vs 95W TDP), while having twice as many cores, that would be a pretty solid step forward in CPU tech.

Probably same 95W, but lower idle power. See above.

2x cores at same power: CPU design isn't magical.