First complete review of Haswell i7-4770K

[Arachnotronic]

Yes yes, big words and sounds from financial speculators on stock market forums with an axe against AMD and an Intel stock portfolio, we heard them before a thousand times just before they'd hit the sewage drain.

Yeah, I also own a bunch of AMD, grimpr, so not exactly rooting for Intel to squash AMD. Although tracking through your post history, it seems you're a huge AMD fan. Care to explain your motives? Mine are simply to seek the truth.

As an Intel shareholder as well, I don't even think about AMD as competition - that battle has been over for a while. Intel's problems focus primarily on the Pandora's Box that ARM unleashed on the market by getting the Android ecosystem to fully support the architecture, and then by allowing Android to gain much more device unit share than Windows PCs ever did. This was a mistake, and Intel "missing" the first round of mobile isn't about whether they can get in and take meaningful share, but more about having to fight a huge uphill battle, and having to fight competitive, well-capitalized, and well-run firms such as Qualcomm (which I also own) and Nvidia (again, own it), rather than simply fighting AMD.

I think AMD can find a niche for itself, but the days of it being a high end competitor to Intel in any sense of the word while remaining economically viable are winding to a close. ARM in Android sort of restricts this niche a bit, though. Android tablets are superb and the ecosystem there only gets larger over time. Microsoft will need to vigorously defend its turf (and this is tough)

By the way, I say all of this with a short position in ARM, an upside bet on Microsoft, and no position in Google.

I tell it like it is (as I see it - my perspective may not always align with a more objective reality), whether it "suits" me or not. You've been goading me to come out and "admit" my stock positions for some time (that I have never kept secret), and you've been EXCEEDINGLY rude about it, so congratulations, I took your troll bait and wasted a lot of time doing so.

Now kindly go back into the Troll Cave from whence you emerged and let's get back to talking about these potentially game changing results for the PC gaming industry now that Intel has a decent IGP with GT2. And yes, we should also talk about the potential impact on AMD - burying your head in the sand doesn't change reality

 

[SiliconWars]

The gap continues to shrink at the high end.

There's no doubt that AMD's GCN is still a superior architecture, but how long does this superiority last? How much room does AMD have to beef up GPU performance in "Kaveri" given that it is moving from 32nm -> 28nm, which doesn't exactly buy a lot of die space/power savings at a given performance design point?

The move to GDDR5 will open up the gap considerably because AMD is hellishly bandwidth bound on APU's. A 7750 should be around twice as fast as a 6670 DDR3 (that intel claims GT3e is comparable to). Kaveri will fall short of the 7750 but not by much.

 

[ShintaiDK]

The move to GDDR5 will open up the gap considerably because AMD is hellishly bandwidth bound on APU's. A 7750 should be around twice as fast as a 6670 DDR3 (that intel claims GT3e is comparable to). Kaveri will fall short of the 7750 but not by much.

Besides TDP to reach a 7750 on 28nm. That GDDR5 is a big problem.

The GT3e still uses DDR3 DIMMs

 

[SiliconWars]

The 7750 draws less power than the 6670 so there will be no problems with 512 GCN SP's on 28nm.

 

[inf64]

One thing is for sure: Kaveri will be considerably larger die than Trinity/Richland. Hopefully they will manage to increase the performance while maintaining similar power draw levels as Richland parts. This would result in big perf./watt jump.

 

[AtenRa]

AMD's graphics advantage just died.

now that Intel has a decent IGP with GT2.

Decent iGPU at 720p LOW settings were CPU plays a big role almost one year later than Trinity??? Richland will be decent in 1080p (Trinity already is, in a lot of games). You have much to learn young one, I would wait until a more in depth Review comes out and then draw any conclusions

 

[LogOver]

Decent iGPU at 720p LOW settings were CPU plays a big role almost one year later than Trinity??? Richland will be decent in 1080p (Trinity already is, in a lot of games). You have much to learn young one, I would wait until a more in depth Review comes out and then draw any conclusions

Slightly increase number of EUs (25%) is almost enough to kill AMD graphics advantage in desktop. And we even didn't see GT3e performance numbers. Let alone notebook segment where AMD basically didn't have advantage at all.

 

[mrmt]

One thing is for sure: Kaveri will be considerably larger die than Trinity/Richland. Hopefully they will manage to increase the performance while maintaining similar power draw levels as Richland parts. This would result in big perf./watt jump.

Won't that kill whatever chances of AMD making money with the thing?

 

[SiliconWars]

No because higher performance parts command higher prices.

 

[mikk]

Considerably larger? Sure Kaveri has 512 Shaders and maybe 3 modules? On the other side it is a shrink to 28nm, only half-node though.

 

[R0H1T]

Slightly increase number of EUs (25%) is almost enough to kill AMD graphics advantage in desktop. And we even didn't see GT3e performance numbers. Let alone notebook segment where AMD basically didn't have advantage at all.

And yet you're claiming this based on what ~ "total blank" besides Kaveri coming out within the next year of this Haswell release, me thinks you don't look twice at what you post :hmm:

 

[SiliconWars]

Considerably larger? Sure Kaveri has 512 Shaders and maybe 3 modules? On the other side it is a shrink to 28nm, only half-node though.

It'll be bigger because of the extra module, but I doubt it'll go over 300mm2 as an example. This should be fast enough to get close to the current FX 6300, plus ~7750 level graphics (with higher TDP obviously no escaping that) at a smaller die size than the FX 6300. I'd guess AMD will target closer to $180 for the high end chip.

 

[galego]

First of all, that paper doesn't seem to be received well, or perhaps even accurate; it doesn't seem to get cited anywhere.

Secondly, do you even understand what is being proposed in that generalization? It says that Amdahl's law holds true unless you have a very specific case in which the speedup is theoretically indefinite (you can continue to parrallelize and get better results) given that the dimensions of the problem being solved increases with the number of cores the problem is being solved on. Otherwise, Amdahl's law holds true: the maximum speedup possible is roughly equal to the inverse of the portion of code which remains serial.

Can you name any code which falls under the first category (the dimensions of the problem can expand indefinitely)?

I don't know why all this sudden emphasis on the Amdahl's law. There are many publications studying this law and its limitations, when it applies and when does not. You can also find several generalizations of the Amdahl's law. I gave a reference that is free, but you can pay for others if you want

http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=135776&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D135776

Let me quote a part of this recent work (Bold face from mine):

http://www.sciencedirect.com/science/article/pii/S0743731509000884

History seems to be repeating itself. Two decades ago, mainstream vendors, worried about the pessimistic implications of Amdahl's law, made parallel machines only with 2 to 8 processors, such as the IBM 7030 Stretch Data Processing System and Cray Y-MP

[...]

Some others follow up with more limitations of multicore scalability based on Amdahl’s law [12]. We apply scalable computing principles that emerged in the decades following Amdahl’s 1967 work to multicore architectures and the hardware model proposed by Hill and Marty. Our study shows that multicore architectures are fundamentally scalable and not limited by Amdahl’s law.

The article proposes another generalization to Amdahl's law with speedups above those predicted by the Amdahl's law.

The future is "moar cores"

AMD's graphics advantage just died. Broadwell will read the eulogy.

Well AMD death has been announced a trillion times before, nobody really cares by the trillion plus one.

Its always better to make a new socket, since you can overcome the design limitations associated with an old socket.

No "always". Sometimes a new socket makes sense, sometimes it is just a way to force users to buy a new mobo.

 

[ShintaiDK]

Amdahls law says nowhere that multicore systems cant scale. But software is what prevents scaling. A handfuld of applications scales essentially perfect. But the wast majority doesnt, far from it.

It seems you completely missed what you linked. But simply took something out of context that fitted your agenda. The article is simply about hardware limitations in scaling.

Moar cores is a dead concept. Just like speedracer designs. And why AMD is essentially dead, and "ATI" is the only remains. How ironic.

No "always". Sometimes a new socket makes sense, sometimes it is just a way to force users to buy a new mobo.

Only to sacrify the benefits of a changing sockets to the other 99% buyers.

 

[NTMBK]

Moar cores is a dead concept. Just like speedracer designs. And why AMD is essentially dead, and "ATI" is the only remains. How ironic.

Increasing core counts is far from dead. Why do you think Intel is going from 8 to 12 cores on Ivy Bridge E? Why do you think they're introducing TSX, which only exists to make writing scalable multithreaded software easier? Why do you think they're bringing out a 62-core part in the Xeon Phi?

Increasing core counts is by far the most feasible way of improving overall CPU performance right now. This is especially true when you have a good Turbo implementation- that way you can still get respectable performance in lightly threaded situations.

 

[ShintaiDK]

Increasing core counts is far from dead. Why do you think Intel is going from 8 to 12 cores on Ivy Bridge E? Why do you think they're introducing TSX, which only exists to make writing scalable multithreaded software easier? Why do you think they're bringing out a 62-core part in the Xeon Phi?

Increasing core counts is by far the most feasible way of improving overall CPU performance right now. This is especially true when you have a good Turbo implementation- that way you can still get respectable performance in lightly threaded situations.

So servers with the ability to thread per concurrent user. Or HPC tasks that naturally scales makes "moar cores" a better concept for average Joe?

TSX is not a saviour either. From Intel themselves:

Intel TSX targets a certain class of shared-memory multi-threaded applications; specifically multi-threaded applications that actively share data. Intel TSX is about allowing programs to achieve fine-grain lock performance without requiring the complexity of reasoning about fine-grain locking.
However, if there is high data contention the algorithm would need to change in order to have an opportunity for high scalability. There are no magic bullets that can solve the problem, since true high data contention implies that the algorithm is effectively serialized

 

[NTMBK]

So servers with the ability to thread per concurrent user. Or HPC tasks that naturally scales makes "moar cores" a better concept for average Joe?

Or workstation applications aimed a single user trying to get work done.

TSX is not a saviour either. From Intel themselves:

Oh, I'm well aware that it is no silver bullet. But it is aimed at improving thread count scalability, which demonstrates that Intel care about it and think that increasing core counts is a valid approach.

 

[tweakboy]

Ivy E will pown 2013.

 

[mrmt]

No because higher performance parts command higher prices.

Why does that smells like Bulldozer part II? Instead of going more efficient, they are going for more die size and more thermals, but now coupled with a huge GPU that will be a bandwidth hog.

 

[exar333]

Why does that smells like Bulldozer part II? Instead of going more efficient, they are going for more die size and more thermals, but now coupled with a huge GPU that will be a bandwidth hog.

Been about 8 years since AMD commanded 'higher' prices...

 

[Arachnotronic]

Won't that kill whatever chances of AMD making money with the thing?

Exactly my point. Larger die, no pricing power due to better GPU, and outdated process node compared to what will surely be 14nm competition at the time.

This will not be pretty any way you slice it.

 

[inf64]

It'll be bigger because of the extra module, but I doubt it'll go over 300mm2 as an example. This should be fast enough to get close to the current FX 6300, plus ~7750 level graphics (with higher TDP obviously no escaping that) at a smaller die size than the FX 6300. I'd guess AMD will target closer to $180 for the high end chip.

3M Kaveri will get close to FX83xx performance wise (not FX6300 ).
iGPU will certainly rely on memory BW but will have some uarchitectural tweaks to alleviate the problems (somewhat). Die size will be larger than Richland for sure. Whether it gets over 300mm^2 or not is anyone's guess( AMD knows ).
Price wise it will be closer to FX6300 or FX8320 since x86 performance will be higher than that level (both ST or MT) and iGPU will be making lower end dGPUs obsolete.

 

[galego]

Amdahls law says nowhere that multicore systems cant scale. But software is what prevents scaling. A handfuld of applications scales essentially perfect. But the wast majority doesnt, far from it.

It seems you completely missed what you linked. But simply took something out of context that fitted your agenda. The article is simply about hardware limitations in scaling.

Agenda?

I don't know what part of what I said or of the quotes reproduced in my post you are rejecting. Do you accept the existence of generalizations of the Amdahls law or still believe it is an universal law?

Moar cores is a dead concept. Just like speedracer designs. And why AMD is essentially dead, and "ATI" is the only remains. How ironic.

Ironic? What is really ironic is how some people say me about how Intel will release 8-core chips to beat AMD and in the same thread others say how moar cores is a dead concept.

And yes, all us know that AMD "is essentially dead". It has been killed/death many times. Not a problem here, AMD is alive.

 

[joshhedge]

Agenda?

I don't know what part of what I said or of the quotes reproduced in my post you are rejecting. Do you accept the existence of generalizations of the Amdahls law or still believe it is an universal law?



Ironic? What is really ironic is how some people say me about how Intel will release 8-core chips to beat AMD and in the same thread others say how moar cores is a dead concept.

And yes, all us know that AMD "is essentially dead". It has been killed/death many times. Not a problem here, AMD is alive.

Intels four core chips w/HT already beat AMDs eight core chips.

 

[NTMBK]

3M Kaveri will get close to FX83xx performance wise (not FX6300 ).
iGPU will certainly rely on memory BW but will have some uarchitectural tweaks to alleviate the problems (somewhat). Die size will be larger than Richland for sure. Whether it gets over 300mm^2 or not is anyone's guess( AMD knows ).
Price wise it will be closer to FX6300 or FX8320 since x86 performance will be higher than that level (both ST or MT) and iGPU will be making lower end dGPUs obsolete.

Have we seen any confirmation of there being a 3-module Kaveri? The roadmaps I saw all said 4 cores i.e. 2 modules.

And I can't see any way that AMD could fit 6 Steamroller cores and a 7750 level GPU onto the same die without hitting Sandy Bridge E level of die size... This seems like wishful thinking which has been repeated so much that people assume it is true.