Want an unlocked 12 core Extreme Edition?

[ShintaiDK]

Exactly!. The point I was trying to illustrate is that parallelism CAN be usefully done, as it is in the GPU.

You cant compare GPU loads to CPU loads that way. Then I could tell you CPUs can already scale to unlimited amount of cores and the problem is solved. You just need an unlimited amount of concurrent users on the same system.

It simply doesnt work that way.

Example
Microsoft/Sony who hopefully know a little bit about game software development, chose an 8-core next generation console solution, even though, if they had really wanted, a (probably) similar performance dual core cpu (at a much higher frequency and IPC), could have been fitted (probably at greater expense).
This could mean that Microsoft/Sony are very confident that 8 cores (I know it is rumoured that less than 8 cores are to be used for gaming) can be effectively used to make modern/competitive video games.

Consoles dont use 8 cores for gaming. They can only use 6. The last 2 are dedicated for other tasks. And just because a game can use 6, doesnt mean it will.

Also you had 6+ threads on the PS3/Xbox360. That didnt change much did it? 7-8 years later and all we got is BF4.

 

[SOFTengCOMPelec]

You cant compare GPU loads to CPU loads that way. Then I could tell you CPUs can already scale to unlimited amount of cores and the problem is solved. You just need an unlimited amount of concurrent users on the same system.

It simply doesnt work that way.



Consoles dont use 8 cores for gaming. They can only use 6. The last 2 are dedicated for other tasks. And just because a game can use 6, doesnt mean it will.

Also you had 6+ threads on the PS3/Xbox360. That didnt change much did it? 7-8 years later and all we got is BF4.

Example of cpu parellism in action, with a number of different applications, scaling across as many as a million cores


I don't know how read up on google, their software implementations, and the hardware they use, to achieve the results, you are ?

They encountered exactly the same sort of problems that you have been raising.

E.g. They had a new e-mail and/or Search and/or facebook google+ and/or etc/etc.

It may well have worked just great on one computer, with one user, messing with one small file, on their $10,000 (or whatever they use) single user workstation, with massage chair room nearby.

But when they try to scale it up, to work with say (all figures huge estimates, and may be widely wrong) 100,000 computers, spread around their world-wide computer centres, they find that it usually bottle-necks at certain place(s).

(In many cases they have kindly released publicly available document(s), explaining how/why/what they did, to solve the various computing problems that they solved). Some of the stuff is commercially secret, as well.

One example that comes to mind, is that it needed extensive database access, to something which had to run on just one machine. I.e. 100,000 computers could run the software, but the database itself, had to run on a single computer, even if it had a few multi-core processors inside it.

They solved problems like that, by using a completely different database system, which readily allowed multiple computers, to work independently on the "massive" database.

------------------

The purpose of the above examples, are my attempt at trying to explain (to you) that multi-core computing problems "CAN" be solved, given the technical time and engineering.

I.e. Saying "Amdahls law" ... so it is absolutely IMPOSSIBLE for any (non-GPU) cpu software, to ever go above quad/4 cores, ever, in the next 2,000 years or so, may be missing out of the true capabilities of mega-multi-cored cpus/computers, which we are likely to increasing have available in the foreseeable future.

Question for you ?
If "Amdahls law" is so unyielding, how come google have been able to very successfully scale up their computer/software systems, to something like, millions of cores (I DON'T know how many cores, google has, in total) ?

 

[ShintaiDK]

Google depends on concurrent users. Aka millions of requests. Its equal to have a million users game a million games on a single shared PC with a million cores for example.

I dont think you understand Amdahls law. It doesnt say applications cant scale 100%. Its simply about scaling vs amount of serial code. Some applications will scale close to 100% due to almost no serial code. Rendering is a good example here. However it goes terrible wrong when you add serial code to the mix that most applications got and will always have.

 

[SOFTengCOMPelec]

Google depends on concurrent users. Aka millions of requests. Its equal to have a million users game a million games on a single shared PC with a million cores for example.

I dont think you understand Amdahls law. It doesnt say applications cant scale 100%. Its simply about scaling vs amount of serial code. Some applications will scale close to 100% due to almost no serial code. Rendering is a good example here. However it goes terrible wrong when you add serial code to the mix that most applications got and will always have.

I was expecting you to say exactly that.

What I'm trying to say is that there can be technical solutions (some of which may not have been invented yet), to these issues, which will minimise/reduce/eliminate these issues.

Consoles dont use 8 cores for gaming. They can only use 6. The last 2 are dedicated for other tasks. And just because a game can use 6, doesnt mean it will.

Also you had 6+ threads on the PS3/Xbox360. That didnt change much did it? 7-8 years later and all we got is BF4.

I agree.
Hopefully the new consoles will, at least slightly, change this, as their single core performance, is NOT very good (I am led to believe).

 

[Ajay]

I have a dual 12 core AMD system, and a 12 core, 24 thread Intel system, and a 8 core 16 thread Intel system, and they all use all the cores. I could use an infinite number.

Yeah, but where would you put them
Oh, nvm, just label them as 'lost socks' - there seems to be infinite storage for lost socks :thumbsup:

 

[Ajay]

Google depends on concurrent users. Aka millions of requests. Its equal to have a million users game a million games on a single shared PC with a million cores for example.

I dont think you understand Amdahls law. It doesnt say applications cant scale 100%. Its simply about scaling vs amount of serial code. Some applications will scale close to 100% due to almost no serial code. Rendering is a good example here. However it goes terrible wrong when you add serial code to the mix that most applications got and will always have.

Actually, some application have super-linear scaling (scale better than 100%). That's because one can look at many types of parallel problems as total work done compared to work done per unit of time. In other words, some problems are simple too big to be done serially (would take 100 years, etc.), but when done on massively parallel computers, they can be run (e.g. maybe only 3 months of run-time). It just takes time and talent to figure out how find a parallel algorithm that does the same thing (or close enough to the same thing) as the serial algorithm. I have a some good papers on this, but I didn't find them online in a quick search (found abstracts, not full text).

It's amazing how often we have the same conversations...


PS Not quite the same thing, but here's one article by Gustafson who's done allot of research here:
http://hint.byu.edu/pub/HINT/doc/Gus/Superlinear/Superlinear.html

 

[ShintaiDK]

Actually, some application have super-linear scaling (scale better than 100%). That's because one can look at many types of parallel problems as total work done compared to work done per unit of time. In other words, some problems are simple too big to be done serially (would take 100 years, etc.), but when done on massively parallel computers, they can be run (e.g. maybe only 3 months of run-time). It just takes time and talent to figure out how find a parallel algorithm that does the same thing (or close enough to the same thing) as the serial algorithm. I have a some good papers on this, but I didn't find them online in a quick search (found abstracts, not full text).

It's amazing how often we have the same conversations...


PS Not quite the same thing, but here's one article by Gustafson who's done allot of research here:
http://hint.byu.edu/pub/HINT/doc/Gus/Superlinear/Superlinear.html

As far as I can see, they simply increase the worksize to get more performance on the expense of memory. Thats nothing you in theory cant do on a single core with the same scaling. The worksizes in Linpack is a good example here. The more memory usage, the higher throughput.

 

[Fjodor2001]

4770k = $300 (close enough for this)

So price per core for 4770k = $75

Lets say yield is 80% for the 4770k die. (change this to whatever you want, just a wild stab at it)

Hypothetical 8 core cpu = 8 cores * $75 per core * (1 / .8) for decreased yield = $750

How would this ever be "affordable"?

I imagine this is why there isn't one on the market.

Obviously simplified, but you get the picture I think. You guys act like releasing a X-core CPU at a reasonable price is just a matter of slowing the clocks down, which in reality has almost nothing to do with it. Extra die size costs money. It reduces yield and reduces the number of dies per wafer. It will not end up being affordable without stripping down the cores' functionality to save die area. There are real examples of this, see AMD's 8 "core" CPUs.

Forget software, this is economics.

An affordable 8-core CPU is not going to give Intel 60% gross margin, so they aren't going to release one. Not with Haswell cores anyways.

At 14 nm you can fit twice the amount of transistors per die area compared to the 4770K which is on 22 nm.

Also an option is to make a CPU version that ditches the iGPU, thereby freeing up a lot of available die area.

I wonder what the size of a potential Broadwell 8 core CPU without iGPU would be on 14 nm, compared to the 4770K on 22 nm.

 

[aigomorla]

if it does come out it wont be 2000 dollars.

probably 1600 dollars? or 1500 dollars?

They will also cripple it from being allowed to be used in pairs on a dual cpu board.


But if it comes out, i see a shining upgrade path for me... otherwise im gonna see how long my current system lasts, and if it can last til Haswell-E

 

[Rakehellion]

I'd rather have the Xeon.

 

[tweakboy]

Looks like Francois from Intel is trying to gauge interest in an unlocked, 12-core Xeon (IVB-EP) for Enthusiasts. To express interest, apparently you need to re-tweet his post on Twitter:

https://twitter.com/FPiednoel/status/396656918975889408/photo/1

Or get a dual socket xeon board. THrow 2 ,,, 12 core thats 24 threads

24 cores and 48 threads HT.

and throw a Titan on top of that. Since you spent already 4k , throw in a gift for you,, like a new mouse or something.. gl

 

[Ajay]

As far as I can see, they simply increase the worksize to get more performance on the expense of memory. Thats nothing you in theory cant do on a single core with the same scaling. The worksizes in Linpack is a good example here. The more memory usage, the higher throughput.

Well, I linked that paper just to show what super-linear was. In any case, there is a limit with a single core where adding more memory will cease to cause improvement but, adding more cores and more memory will continue to scale.

Bottom line, parallel is awesome where it works well and developers have the time and resources to push it to the limit. Unfortunately, that just not the case on the desktop for most programs.

Most irksome to me is that in single player mode most games just allocate a single thread to all 'AI' tasks, even if there are multiple factions. As such the 'AI' stays dumb and harder levels just me more enemies with more resources (and 'cheats' as well) - instead of more complex strategies. So single player is dying out in many mainstream games because they can just switch to multiplayer and have you fight a real learning machine (aka, a human brain).

 

[NTMBK]

I wonder what the cost of making this SKU would be? Surely it would just be changing which parts of the chip are fused off (fusing off the QPI links, leaving the multiplier unlocked) and printing a new box? There's obviously the opportunity cost of selling them to us plebs instead of into high-end workstations or 2P servers, but if they're selling them at the same price as the Xeon then that's essentially negligible. (And Intel are hardly capacity constrained right now.)

How many would they actually need to sell, to make it worth their while?

 

[ShintaiDK]

I wonder what the cost of making this SKU would be? Surely it would just be changing which parts of the chip are fused off (fusing off the QPI links, leaving the multiplier unlocked) and printing a new box? There's obviously the opportunity cost of selling them to us plebs instead of into high-end workstations or 2P servers, but if they're selling them at the same price as the Xeon then that's essentially negligible. (And Intel are hardly capacity constrained right now.)

How many would they actually need to sell, to make it worth their while?

Relatively cheap, if it was to run like an existing product. However with the E5-2697V2 being 2600$ and 2.7Ghz with 3.5Ghz turbo. The market is simply minimal in the desktop segment. You need inventory, support, warranty for 3 years etc etc. And at what price? 1500$? 2000$? 2500$? I cant see many buyers for it. And the PR would be a disaster, when review sites would show this CPU being slower than a much cheaper 4960X in all gaming.

 

[Fjodor2001]

I just want an affordable 8 core mainstream K processor. Too much to ask?

+1

And if priced by die area it should also be affordable. But the problem for Intel is that it would cannibalize on its other higher end segments then I guess. That's the beauty of monopoly (well, close to), when you can make such marketing decisions.

It's also why I'm hoping we'll see a 6/8 core Kaveri later in 2014. It's what might push Intel to release 6/8 core mainstream CPUs too.

 

[ShintaiDK]

+1

And if priced by die area it should also be affordable. But the problem for Intel is that it would cannibalize on its other higher end segments then I guess. That's the beauty of monopoly (well, close to), when you can make such marketing decisions.

It's also why I'm hoping we'll see a 6/8 core Kaveri later in 2014. It's what might push Intel to release 6/8 core mainstream CPUs too.

Its called capitalism and you are just not willing to pay what it cost. Instead you attack the lower products that capitalism and free market gives you. CPUs have not been cheaper in history.

 

[Fjodor2001]

Its called capitalism and you are just not willing to pay what it cost. Instead you attack the lower products that capitalism and free market gives you. CPUs have not been cheaper in history.

The free market and capitalism doesn't work well in a monopoly situation. That is why for example the U.S. has the antitrust law. The trick for a company is to stay as close to monopoly as possible without violating those laws. It for sure does not benefit the customer though.

 

[ShintaiDK]

The free market and capitalism doesn't work well in a monopoly situation. That is why for example the U.S. has the antitrust law. The trick for a company is to stay as close to monopoly as possible without violating those laws. It for sure does not benefit the customer though.

Answer me this, do we or do we not have the cheapest CPUs ever recorded in history in relation to the price index?

Its an endless loop with you because you feel entitled to a cheap 8 top performing 8 core on the mainstream desktop platform.

 

[Fjodor2001]

Answer me this, do we or do we not have the cheapest CPUs ever recorded in history in relation to the price index?

Its an endless loop with you because you feel entitled to a cheap 8 top performing 8 core on the mainstream desktop platform.

I'm not sure what the point of that question is. Of course electronics hardware becomes cheaper (or faster, better, for the same money) over time due to technological advances. It's not a limited resource like oil or gold.

But that does not prove we do not have a close-to-monopoly situation on the high end mainstream desktop market, or that Intel is not making use of that situation to segment the market in the way I described.

Let me put it this way: If AMD would release a mainstream 8 core Kaveri based CPU with single core performance close to Haswell, don't you think Intel would release a similar CPU soon after that? And don't take the Bulldozer FX CPUs as an example of why that does not have to be the case, since their performance per core does not match Intel's equivalent.

It's an endless loop with you, because your are constantly striving technologically backwards in your wish for as few cores as possible, and refusing to acknowledged the close-to-monopoly situation that Intel has and how it uses that to benefit the company's profit but not the end-user.

 

[ShintaiDK]

Again you make up the assumption that AMD would release one and evil Intel is keeping you from your cheap 8 core. Yet AMD do not want to, because the market for one is too low. Even if AMD performed the same as Haswell, they are more interrested in selling better IGPs just like Intel is, since this is where the demand is besides lower power consumption (AMD want to target a max of 65W after Kaveri.). The amount of people that want the same as you is simply too small to make a product for on the mainstream platform. So what you demand is for a company to make a bad economic decision and punish the wast majority, just so you can get your personal wish that you feel entitled too come true.

 

[Fjodor2001]

Again you make up the assumption that AMD would release one and evil Intel is keeping you from your cheap 8 core. Yet AMD do not want to, because the market for one is too low. Even if AMD performed the same as Haswell, they are more interrested in selling better IGPs just like Intel is, since this is where the demand is besides lower power consumption (AMD want to target a max of 65W after Kaveri.). The amount of people that want the same as you is simply too small to make a product for on the mainstream platform. So what you demand is for a company to make a bad economic decision and punish the wast majority, just so you can get your personal wish that you feel entitled too come true.

Well, first we have no idea what AMD plans for Kaveri, and whether there will be 6/8 core versions coming later. Also, they are in a tight monetary position, so they cannot do everything at once.

The comment that the market is too small for an 8 core mainstream CPU you have no proof of either, simply because there is no such CPU available so we cannot see what the actual sales volume would have been.

And you cannot take the non-existence of such a CPU as a proof of that there is no demand for it either, simply because we have a close-to-monopoly situation where Intel can choose to segment the market at will to maximize profit, despite that consumers actually would have preferred a different setup and product portfolio. If we would have had a functioning market with perhaps 4-5 close competitors in the mainstream x86 CPU area, then I'd agree with you.

 

[ShintaiDK]

The comment that the market is too small for an 8 core mainstream CPU you have no proof of either, simply because there is no such CPU available so we cannot see what the actual sales volume would have been.

How is the FX8xxx series selling again?

 

[Fjodor2001]

How is the FX8xxx series selling again?

Did you even read me earlier post saying "And don't take the Bulldozer FX CPUs as an example of why that does not have to be the case, since their performance per core does not match Intel's equivalent."?

 

[ShintaiDK]

Did you even read me earlier post saying "And don't take the Bulldozer FX CPUs as an example of why that does not have to be the case, since their performance per core does not match Intel's equivalent."?

Neither would an 8 core Intel vs a quadcore Intel.

 

[mrmt]

It's an endless loop with you, because your are constantly striving technologically backwards in your wish for as few cores as possible, and refusing to acknowledged the close-to-monopoly situation that Intel has and how it uses that to benefit the company's profit but not the end-user.

You are right when you say that we would be seeing more desktop products if AMD had a better foundation to compete, and Shintai is right when he says that the kind of product you want is not feasible anymore.

The kind of product you want just isn't feasible anymore. Both companies give us their server rejects. AMD started doing that circa 2009 with Thuban, and Intel went fully that route with Sandy Bridge. Take for example, the FX 9590. Do you think AMD really cared enough about their consumers to give them a goodbye gift for AM3+ socket? No, they don't. FX 9590 is just a high leakage part that had no place as a server die, or even FX 8350 die, so they binned it as a 9590 and dumped on the market. Crappy silicon sold as a premium product, a nice marketing coup if you will.

But what if AMD had a meaningful server market share, making enough money to develop new products? Then you'd probably see that kind of product, 8-10 Steamroller cores and the die-salvaged parts. Same as before, better bins for server, rejects as the FX line.

But... think about it, 2-3 years ago AMD used to have some 650MM for SG&A expenses and only three product lines to develop (big core, small core, GPU), and what we are seeing today on the market is the product of that time. Today the same company has just 450MM to sustain on top of the three previous lines an ARM product line and Seamicro. Do you really think they can develop the old lines with two thirds of the previous budget, not counting two more product lines to feed?