Separate names with a comma.
Discussion in 'CPUs and Overclocking' started by MisterMac, May 8, 2012.
The folks who say "who needs more than 4 cores" arent very forward thinking...
Except some of those differences you state are completely meaningless for desktop environments, mainly double the memory bandwidth and extra L3 cache. Double the PCIe bandwidth matters only if you're running three or more graphics cards or you have two or more graphics cards and a RAID card(s). When all is said and done, perhaps 1% of enthusiasts will take advantage of those things you mention and don't need the much higher platform cost--hence why processors like the 2600K and 3770K sell a lot more than the 3820.
Going right into platform costs, here are some prices:
Core i7-3820: $310
ASRock X79 Extreme3: $210
8GB (4x2GB) DDR3-1600: $55
Core i7-3770K: $350
ASRock Z77 Extreme4: $140
8GB (2x4GB) DDR3-1600: $45
Core i7-2600K: $310
ASRock Z68 Extreme3 Gen3: $122
8GB (2x4GB) DDR3-1600: $45
Well, would you look at that? $100 less, or 21% less, for a platform that will deliver the exact same CPU performance in 99% of desktop applications all while consuming a decent amount less power. That, or $40 less for a CPU that has 5% higher performance clock-for-clock and additional power savings over the 2600K. That, and since the Mainstream platforms are only using half their available memory banks/slots (you need to get a $250+ X79 motherboard for 8 RAM banks), that means the platform is better in something more people will actually care about: the ability to add RAM without having to replace all of them. Gee, must be real difficult not to understand why it's a smarter choice to go with Z68 or P67/2600K or Z77/3770K rather than X79/3820...
To quote Tom's Hardware:
TL;DR: If you're gonna buy into the expensive X79 platform and you're not going with a Six-Core, you're wasting your money. If you're gonna go with a Quad-Core buy into the Z68/P67 or Z77 platform and if you're gonna go with a Six-Core buy into X79 and a 3930K. The -3820 is a completely odd duck in that it doesn't really fill any purpose yet you get to pay extra for it upfront and in your power bill, hence why it sells very little.
I see virtualization taking off even for the home market in the next few years. More cores is definitely in the future.
Show me the software that craves 6 cores ?
seti .. folding .. encoding .. a photoshop plugin ?
4 cores is more than enough for the time being .. let the software tinkers get a chance to catch on ... it will take another 5 years minimum (i think).
It's progression - it's NATURAL to expect MORE for the SAME PRICE range each generation.
are you inept? insane?
As virtualizing and multithreaded revolution takes up arms(It' should have already gone off by now imho) - more cores is more performance without reinventing the deep wheel.
this is not a "OMFG you have to PAY more for 6 cores, why?!" whine - this is we want progression, we want more performance as time passes on.
How can ANYONE disagree with this?
I find it amusing and backwards thinking - almost offensive.
Most people downright FLAILED amd for not really increasing perf\price from the previous generation but staying linear.
Yet Intel\AMD on the CPU side are allowed to not progress at all?
If we all had 12 cores nehalems atm - you seriously think games and anything else wouldn't be heavy threaded?
You think streaming and mobility devices would pop up if the radionet wasn't pushed on to edge, 3g, 4g?
It's so backwards to claim "we won't need more than a Quadcore" - well no shit.
Cause there's no reason for anyone to develop a product that requires more when we can't we buy more without selling our unborn child.
People should think about why games and programs dont use more threads than they do. Then they would know why you wouldnt need more than a quadcore for many years ahead.
Servers got it much easier, since you have multiple concurrent users for example. But it doesnt help much if you go from 4 cores to 32 cores. If 50% of your code is serial in nature due to the way your program function.
Its not for fun that Intel for example tried looking at the mitosis project. Basicly willing to use 8 cores to gain maybe 10-15% over 1 core with serial code.
TSX greatly improves multi-thread synchronization.
Furthermore, Haswell will have twice the throughput per core,thanks to AVX2. So no need to ask for more cores when you're getting a lot more performance out of a quad-core compared to previous architectures. ^_^
More cores is not always more performance without reinventing the wheel. If you understood programming, you would know that there are only so many threads you can make to perform a task before you either A. see no more gain, or B. code becomes way too complex to manage. Right now there are only a small selection of application which can use more than cpu 8 threads efficiently. And the other 90% of applications run fine on 4 cores and would not see any difference on 8.
So I must be insane. I dont think MOAR COREZ is the answer to everything. New Instructions, IPC increase, better cache subsystem, faster RAM is more important to me than more cores.
I honestly have no idea why you made that long post comparing the 3820 to the 2600K, which has been beat to death back in Jan. In this thread I was comparing the 3770K to the 3930K which is the 4 core and 6 core CPUs of choice right now. Your post does not seem to add anything useful to that discussion.
You have to look at it from Intels POV. How many mainstream users even really need four cores, let alone 6 or 8? Hardly any, so if you need that many cores you're not really mainstream and you need to go to the other sockets,
1150 is for the 99% of the people that don't need 6 cores. Many of these though will definitely be using the integrated graphics, hence the need for Intel to improve in this area.
It's all perfectly logical to me.
What is the price for a x58 setup in EU OP?
I dont think it is a matter of core count (I'm in the e-Penis gimme mah corez camp because I hold on to PC for a long time)
The OP is concerned that Haswell is sticking to status quo.
Since Haswell is only a year away (give a take a couple of months) the current core counts seem fine given the SW market in the mainstream class.
Probably the tick (or is it tock? I don't remember which is which) aka Haswell 1.5 likely to pop up in 2014 will introduce more cores in the mainstream.
For now and the near future (2013 is near future at this point), 4 physical\8 logical is sufficient for mainstream class PC's.
Those that want to get their core counts on can pay for enthusiast\workstation class parts.
I am a programmer, by hobby.
It's not my fault threading and locking - is made a complex scheduling hell with x86\ia 64 implementations in most modern libraries.
The fact asynchronous locking only becomes a problem if you don't write the basepillar well enough.
And yes, it's easily becomes complex and requires strict I\O control design decisions wether input is from pipes,memory or network.
However since java or dotNET's half arsed performance killing attempts still yield slighty on the bottomline people use that.
However if it was essential for performance(hi 12 cores), you'd see all the more options and well written choices for the "average programmer joe".
People would line up to provide expensive libraries left and right - to use in engines and god knows what else.
Trust me that if this was the groundcase - intel\amd would obviously improve all they can - but we'd see a core count race.
And as long as they could keep ringbus bandwith/communications up - it'd be a extremely easy shortcut way to performance.
It's the whole concept of evolution all over - give people more to use, more to learn, more to do - and people will do it.
Get your faster ram, your faster latency, your faster bus interfaces.
That's all pretty shit useless with slowass DISK i\o - if you ever wanna actually manipulate that data.
MOAR CORES and faster interface connects to storage will provide multiplicative larger efficiency than 2800 mhz ram or larger l3 cache.
Both for desktop responsiveness and multitasking "Experience" - and the numbercrunchers in datacenters.
I have to say I agree with Mister Mac, mhz stagnation + core count stagnation == meh. I agree with Edrick, new ISA extensions should make the new four cores devastatingly fast in some cases.
I won't apologize for Intel, though. It is only through creating a dissatisfaction with the current situation that change can be driven. Cheerily taking whatever Intel gives us is silly, we should expect more... Shareholders want blood from the stone, as stakeholders we should get our due as well.
It looks like AMD might have a target niche to hit if they can deliver since Intel is set @ four cores/eight threads for the foreseeable future.
1200 USD in Denmark\Scandinavia cheapest.
Good mobo, add 100/200 USD.
Prices for a 980/970 are same&same actually, 20-30 USD apart - 800 USD.
For some odd reason - no price difference from 980 to 3930k.
990x still selling for 1400 USD.
(Ironicly, 3960x is like 1325 USD ish - yea we're messed up).
You not only have to code for Multithreading but you also will have to code the applications for AVX/AVX2 and new ISAs in general.
Having a single core with AVX is nice, having a dual AVX core is better, a quad AVX capable core is even faster and so on and so on.
We need all of them, new ISAs, higher IPC, higher core count and programs coded for multiprocessing. We also need applications taking advantage of new ISAs sooner.
Yes I seriously think so, i would think you blew your money on next to nothing.
Writing "threaded" software is not *just* about designing your software "to be threaded" .. it is friggin hard given the current paradigm of programming languages .. you are looking at a real hurdle here. Shit has to get reinvented, the wheel has to get .. round again.
edit : point being, until you have a generic approach to utilizing n cores I believe we are kind of stalled here - what 'generic' software today utilizes 4 cores+ without a hiccup ? None. I wouldnt be surprised if we need a whole new kind of language to cope with this..
Are you a programmer?
Have you a designed a Asynchronous memory io for threading?
For controlling locks?
Then your entire post is useless and bs.
You do not claim fact on given scenario with a subjects you have no knowledge of.
It's mathematics, it's logic.
Differencet Scenario's different designs.
The base however is the bloody same no matter how you twist and turn it.
Oh god, please no Well, at least not via CMT. I wouldn't mind seeing a Thuban+2 design with Bulldozer improvements.
Part of the reason for the stagnation as far as cores/threads goes can be seen with the way Intel approaches server chips. AMD actually provides a better example of this because of how extreme the impacts are as you scale up in core count so let's take a gander at Interlagos:
16-core 8-module MCM Interlagos chip is only clocked at 2.1ghz. Though it's dirt cheap for what you're getting (core-to-$$$ wise), it also means they've had to decrease the clock speeds significantly to fit within a set TDP and provide at least somewhat decent perf-per-watt (that depends highly on workload, but you know what I mean). Keep in mind this is also an architecture that has a purposefully longer pipeline and slower caches that allow for higher clocks yet it remains difficult to attain higher speeds as you scale up in core count, MCM or not.
This ties in with all microarchitectures and means that given an equal die size and same architecture the higher core count one will provide lower clock speeds and thus lower performance at fewer threads. Not a big deal for servers but for the desktop it doesn't make much sense. So instead of giving us something that most users won't use at the cost of what users would use, Intel has opted to increase significantly the one area where they currently suck at and that's graphics. The other area that they suck even worse at happens to be drivers for graphics, so we'll see how that works out...
But every new instruction hardware set, requires a re-implementation in 99% of the cases.
Moar cores on a base modular build up of locking/threading access does not.
Optimizations of everything would come, naturally.
But what' gives most of the box ?
I am a programmer, by degree and by profession...
It's complex on any architecture. But fortunately Haswell adds hardware transactional memory to greatly facilitate it and increase performance at the same time!
The reality is that the majority of people have a dual-core CPU, and the number of people who still have a single-core is about the same as the number that have a quad-core. In other words, you can't cater for the 4+ core market any more than the single-core market if you want to create the best experience for everyone!
That said, we're certainly not at a standstill. Ivy Bridge brings quad-core to mainstream laptops, and Haswell will extend on that and brings us AVX2 and TSX to boot.
So don't underestimate Intel. They know what the market needs. If it's evolving slower than you expected, maybe you should adjust your expectation and also see how revolutionary AVX2 and TSX really are.
That's why SSD's are under heavy development.
I'd just like to throw out there that here in the states, you can build your own uber-mega thread box (dual socket LGA2011, dual six core SB Xeon w/HT) with 64GB ECC and all the trimmings for less than $3k. I know that sounds ludicrous, but rewind ten years or so.
@ Pelov -- haha I am thinking we'll know more about AMD's future with Steamroller. It sounds like they are going to have laser focus on fewer segments. If they can balance the clock speed vs thread count optimally and come out ahead of Intel in content creation, that'll be a win for some people.
All valid points, but completely besides the point - i'm making.
AMD's MOAR COREZ strategy - should have already payed of.
(If BD hadn't actually been worse than stars, but let's leave that).
A modular design paradigm of threads and available cores for resources would allow both for improvements from IPC and pipelines for decoders - but also allow fast free performance upgrades by shrinking and slapping another core on the pcb out of the box.
Assuming you'd have scaling nailed down that should give atleast 80% for each core added.
The reality is what you say - what i don't accept is why it's that.
As AMD wanted it to be - core count should matter.
Fully modular use of resources on a cpu - would not only bring another arms race (The ghz warz vs core warz) instead of Intel now sitting on it's throne expanding on all fronts.
Intel couldn't do both medfield & haswell GT3 - if they had to deal with 16 core desktop BD's in threaded world.
- I get paid as one, but that means jack.
- I am not quite sure I get this.
fx. i've designed heavily threaded code to move data from A to B, as to max out bandwidth and get the job done in x time, with all that that implies. I've worked with and against synchronization, locks and race conditions, meh, aside.
General synchronization issues aside, you're talking about transactional memory (in context of haswell), right? Since transactional memory is nothing new, then show we me the development platform that takes a generic approach to utilizing n cores. Is it there? outthere? somewhere? (and case in point your example was 12 core nehalem)
- ARH COMMON .. IF Anything I am ignorant .. and then pleaaaase educate me without the bs ..
- TO THE BEST OF MY KNOWLEDGE - therefor i am
... I read this paper on SCALA and their 'award winning' approach on threading of the future... and it wasnt convincing, to me atleast, to be a generic approach.
- that was redundant at best (sorry dude )