AMD CPU speculation... and expert conjecture

[hcl123]

Yes, but opteron solutions aren't really aimed at consumers...I suppose you could build one. Though, I really don't think the quad channel opteron controller would work in the FX series. Even though the 2 are of the same architecture, the 16 core opterons are a different animal. Integrating them into the desktop FX line would likely present engineering challenges and R&D costs of it's own. The benefits of which would have to be weighed versus the cost of losing business they might have gained.

Considering their current position, I think it would probably cost them less in the long run to just keep dual channel memory, because the percentage of consumer grade PCs with quad channel memory demands is likely so small that what little margin there is in it wouldn't recoup the costs from the R&D.

Yes but they could had trashed the Intel 6 core/ 12 thread "enthusiast" solution of Intel... 2x 3 modules/6 core chips on a MCM package, could go or surpass yet the 4.0Ghz barrier, power will be higher, but this would be a better solution than Centurion, a REAL 12 core G34 solution at 4 to 4.2GHz will be no match in multithreading for the intel solutions.

And the chipset could be the same FX990, the FSB would be the same Hypertransport link of server, that is, no need and not recommended to have a server chipset.

 

[8350rocks]

How about Microsoft shooting themselves in the foot with XB1? Anyone else see this?

http://www.tomshardware.com/news/MSFT-Online-Used-Games-Xbox,22987.html

 

[esrever]

How about Microsoft shooting themselves in the foot with XB1? Anyone else see this?

http://www.tomshardware.com/news/MSFT-Online-Used-Games-Xbox,22987.html

I think its a good thing for a good majority of people. Having to put a disk into the console to play it is so 8 years ago.

 

[hcl123]

I think that offering G34 Opterons as performance chips won't gain AMD any real money because they would lose the poorly-threaded, ICC-compiled benchmark suites widely used by lesser review sites even to the current FXes. AMD is probably better off in offering special tweaked chips like the FX-9000 Centuriion than in trying one more hurrah in bringing more larger-server hardware to the desktop.

Yes but those benchmarks are NOT REPRESENTATIVE of current software used, least of all with Linux where many of this systems would end up (FX could win there), and a workstation market is where there are more aware and knowledgeable users.

G34 around 220W would be a much better solution than Centurion now, it may not reach 5Ghz, but it would have 12 cores and the package be above the 4Ghz yet (2 FX6000 serie chips on a MCM, each above 4Ghz is quite possible attending the improved 32nm SOI of Richland). BETTER.

 

[hcl123]

Intel also integrated the VRM onto the same die, a very difficult thing to do. If you do the numbers, they run up 13% higher in power consumption with 12% increase in performance. I don't think that was appreciated by many, but its a necessary step in Intel's plans, and was going to happen sooner or later.

I see a 8% increase in general performance, but only an 1-5% in games. Moreover the chips run hotter now and OC poor.

Yes FailWell is quite disappointing for a tock... but each time i see a comment it grows in performance, now its 12% lol... soon it will be 15% lol... and not that the "commenters" are wrong, the hardware will not change, but the benchmark software will, yet carry the same name and version number; that is why md5 signatures would be so important to have less deluded audience.

An important reason are exactly the "critics" pointed to AMD, the front-end can't supply more, that intel decoder also reached the top, shown by Haswell having 2 more execution pipes than SNB/IB, theoretically it could be 33% ahead, but isn't... which proves that is not BD "design" that is wrong, its mainly a thing with x86 model that constraints.

Also the "hot spot" issues that was present in IB seems to got worst with HW... it might mean (for me is for sure) that intel doesn't have the best fab process around, and makes it worst when a process is tweaked for "low power", which makes it usually not good at "high clock high performance", but is used for all solutions. That is, the process "gets" comfortable only at those low TDP marks, which might indicate that a 84W part (lots of salt of how intel counts this) is already a factory OC part, meaning getting low power marks doesn't mean you have MORE headroom to much higher TDPs, only means different considerations and possibilities ( i think that is why AMD is playing with "Centurion")

ARM have quite low power yet apart from STMicro none passes the 2 GHz, simply because the process is not up to it (ARM on high performance SOI and possible above 4Ghz), its nothing about design.

EDIT: IMO where Haswell could shine is on benchmarks with Hardware Lock Elision and some of that other Hardware Transactional Memory features. And i say benchmark software, because none of the current software used out there will use it, not for quite some years, not if desktop/client software isn't pushed to quite more thread count in average. Those HTM features only makes sense on high number of thread, even speculative threads, not on low thread number and worst when the push "propagandized" seems to be for "single-thread" (single-thread was never about high performance, which shows that BD kind of uarches can have quite an advantage in future, they have already this consideration).

 

[-Fran-]

You raise a very good point hcl123, but I'm sure there's a very practical reason behind the G34 being left out for servers only.

It could be the incredible fail that QuadFX was, or maybe the fact that Intel's "Extreme" line sells little in comparison to the "mainstream" i series that it could not be worth the effort from all that execution.

I'm really not 100% sure, but it is indeed a very good question to throw out there.

Cheers!

 

[EXT64]

Some people have OC'ed current G34 chips to the mid 3.5 GHz range - and once there they are the fastest single (4P) node setup (faster than Intel's 2011 4P which has no OC capabilities). Unfortunately they are at the limit of motherboard power circuitry (entire system over 1300W) with current silicon; but if an improved process could get them closer with reasonable power consumption they would be amazing. Hopefully we will get that - plus even more performance per clock - out of SR.

 

[hcl123]

A little more poking at Hasfail: http://www.xbitlabs.com/news/cpu/display/20130606231316_Retail_Versions_of_Intel_Core_i_Haswell_Are_Hotter_and_Slower_Than_Expected_Report.html

So the early refresh will be right on the heels of Steamroller if the planning is unchanged... I wonder what AMD has under its sleeves, really.

Cheers!

Yes that article seems "ANOTHER" admission of NON-REPRESENTATIVITY. Intel has quite the resources to proceed with much more exhaustive "binning" strategies, which is usually, all this more extensive testing of most of the dies, a quite expensive undertaking.

So the "Review" parts might get the very best of bins of the product, not an average selection.... meaning not all dies end up the same, but for reviews will go the very best. So end users with intel usually will get worst, yet based on chart numbers they think they have the same( there will be very very very few lucky ones).

Not FUD but a reality to every IDM and another fallacy, that with intel is quite worst as it seems.

 

[hcl123]

Wow...4.3 GHz is all any of them can get?? That's sad...

Even though this generation is a failure, because it's intel, it will still be successful.

Haswell will be successful because the graphics are 40% faster than Ivy and overall they consume less power than Ivy, allowing to enter even lower power form factors.

What some niche overclocked pre-built shops can do is entirely irrelevant at under 0.1% of the marketplace.

Well not quite that irrelevant. I remenber the issues with some MacPros and some users returning their laptops because of heat problems: potentially you can OC quite high, the most problem is getting the heat out of the way as fast as possible.

Low power and high-clock high performance you cant have both at the same time... as simple as that.

^^ strangely, the cpu support list says fx4350 is supported (listed as 125w) yet none of the other 125w fx cpus are supported...
edit: vrm look anemic... and it's cheap asrock...
however, i suppose higher end mobos with better vrm can support as long as bios support is there (wikipedia says cpu pin counts are the same while socket hole numbers are different...)

I pointed that out...the 4350 is the least power hungry SKU of the 125W TDP chips.

Yes and you can't see that by the rated TDP. Usually a chip consumes much less power than the rated TDP that is only a theoretical max TDP for which the vendor "certifies" a chip. Many things have influence, the software load have influence, if it triggers all the subsystems of the chip intensively or not, even the cooler have influence, the lower temp a chip runs, the lower the TDP "wasted" tends to be.

IBM chips have quite high possible TDPs, yet they don't even rated them like that(edit lol), usually is the all power consumption of an all system. And in this department, Intel usually is also very good, specially at medium rated configs, but at the end of the day is not a CPU chip alone that makes a lower power consumption system.

I think for more demanding systems AMD is on par, but there less expensive parts tend to be worst than intel.

CONCLUSION is not a CPU chip that makes you green or not, as example even LCDs, if not LED, there are quite hungry beasts even above the 300W (30" ones). Then count the all chipset and all the pheripheral and expansion where a single GPU alone can be above 300W to (and i've seen many with 2 and some even with 4), so counting only the CPU chip for power is a fallacy.

As i pointed above, you can even rate 300W, yet typical usage be quite below

Example (LostCircuits was always the best at this, perhaps that is why it doesn't get sponsoring lol )

even the FX8150, that suspected not very good hybrid of 45nm and 32nm, courtesy of GloFo, doesn't reach at load its rated 125W.

http://www.lostcircuits.com/mambo//index.php?option=com_content&task=view&id=105&Itemid=42&limit=1&limitstart=5

http://www.lostcircuits.com/mambo//index.php?option=com_content&task=view&id=105&Itemid=42&limit=1&limitstart=6

(windsor K8 is still the champion, 25W typical usage at load, rated 65 W i think)

 

[blackkstar]

Yes that article seems "ANOTHER" admission of NON-REPRESENTATIVITY. Intel has quite the resources to proceed with much more exhaustive "binning" strategies, which is usually, all this more extensive testing of most of the dies, a quite expensive undertaking.

So the "Review" parts might get the very best of bins of the product, not an average selection.... meaning not all dies end up the same, but for reviews will go the very best. So end users with intel usually will get worst, yet based on chart numbers they think they have the same( there will be very very very few lucky ones).

Not FUD but a reality to every IDM and another fallacy, that with intel is quite worst as it seems.

The real question is: do better binned chips turbo better, thus running at higher frequencies when compared to stock chips?

If a well binned 4770k can run closer to 3.9ghz under more load, and 3770k falls to 3.5ghz, that's quite the clock speed advantage, and it doesn't show up in any specs as the advertised base and turbo speeds are the same.

 

[hcl123]

You raise a very good point hcl123, but I'm sure there's a very practical reason behind the G34 being left out for servers only.

It could be the incredible fail that QuadFX was, or maybe the fact that Intel's "Extreme" line sells little in comparison to the "mainstream" i series that it could not be worth the effort from all that execution.

I'm really not 100% sure, but it is indeed a very good question to throw out there.

Cheers!

QuadFX shouldn't had to be a failure. The management of AMD was clearly a failure. QuadFX was scrapped, because that was a time when AMD was considering to sell its chipset division, and with that scrap HyperTransport( meaning now HSA chips would not be possible), seamless multi sockets would be scrapped (depend on a few suppliers for this for server), while client would have to rely on general 3th part offerings based on what they would offer, meaning usually single socket PCIe all around
(if they had sold the chipset division, there wouldn't be AMD by now, no place to put a CPU or an APU, safe the very worst, and blame it on the chips not chipsets lol)

Just to imagine how much better Hypertransport (torrenza interconnect, based on work of Jim Keller) was for those multisocket approaches compared with the clumsy shared FSB topologies of Intel... yet Intel had one double socket for client while AMD had none !!!

The propaganda (most sites that most of you point now indicating that intel FSB is better and if not, not worst LOL )... the end of cash... the market distortion that lead to the famous law suit against intel that intel bought out for peanuts...

Elas... G34 is quite a much better MCM than any FSB could had been, or even intel QPI can imitate. G34 is left for servers in a logic consequence of all that... markets that AMD already lost or scrapped. But if they want to try a come back of sorts then G34 is much better than AM3+ for sure, Centurion or no Centurion (the future is multithreading not single-thread), specially if rated TDPs is not an issue, as it usually isn't among OCers.

 

[hcl123]

The real question is: do better binned chips turbo better, thus running at higher frequencies when compared to stock chips?

If a well binned 4770k can run closer to 3.9ghz under more load, and 3770k falls to 3.5ghz, that's quite the clock speed advantage, and it doesn't show up in any specs as the advertised base and turbo speeds are the same.

Not only possibly turbo better, but also have less rated TDP. As i point usually typical usage is quite lower than the rated at possible worst loads, the problem is that if you don't have a binning system you might end up selling quite a lot of hot pigs on the bunches, akin to sell oranges for apples. End users usually none the wiser couldn't really tell the difference, but the OEMs would scream like maniacs.

So the better your binning the much more accurately you can lower your "certification" in the direction of the typical usage loads. The worst binning the much more careful you have to be for not hear OEMs screams( a larger margin). At the end of the day, none of the IDMs exhaustively tests and tweaks ALL chips, and the power rating is only an approximation by surplus of reality with more or less error margin.

Usually with some tests around, what is measured is the all power of a system(from the wall), or more parts than a CPU alone. Usually Intel wins, but that is not exclusivity of the CPU chips, other circuits count here, and a difference in memory or other peripheral or expansion can make a lot of difference. And this "measured" powers, if anyone payed attention, invariably is quite higher than the rated power of the CPU chips.

EDIT: Worst of all is "measuring" by a software alone approach... quite a fallacy...

EDIT 2: Another fallacy is the "linear" extrapolation related with performance and high clock. There is a point where you can push the clocks quite higher than a base reference, yet only get a minimal performance improvement, because the chip was not designed, specially circuitry FO4 and fab tweak, thinking about higher clocks. I tend to believe with reason that Centurion is quite a good factory OC, i've seen tests of BD and Vishera where the % of performance even raises above of the % of clock raised(example; clocks raise 20% yet performance raises 23%). The big and slow L2 of this designs, getting untangled, seems to me to be a good explanation. So Centurion is well imagined, if a G34 even better IMO.

The other part related with performance, is the fab process and clock, it might even be possible to OC a chip as high than another process chip that is tunned for high clock, but then your chip that at "certification" had quite a lower TDP, now at equal high OC have quite higher TDP... and heat derived from power wasted of higher general impedances of that lower power process is quite an issue... simply can't have the cake and eat it at the same time, a fact of life very annoying for those that want to have everything.

Some people have OC'ed current G34 chips to the mid 3.5 GHz range - and once there they are the fastest single (4P) node setup (faster than Intel's 2011 4P which has no OC capabilities). Unfortunately they are at the limit of motherboard power circuitry (entire system over 1300W) with current silicon; but if an improved process could get them closer with reasonable power consumption they would be amazing. Hopefully we will get that - plus even more performance per clock - out of SR.

I don't think that would be that high for a "client" chipset like the RD990 series. But neither because of a server chipset alone could it be, even if those include kitchen sinks lol . Typical 990FX is about 15W, the southbridge depends, more the power of the of a G34(could be 220W). 1300W might be but for a full server with 4 sockets.

Yet you could put some little more on those chipsets yet get half the power if at 28nm, IIRC the 990 series is made at 55nm TSMC... quite old... 28nm i think could get more yet easily be half the power.

Want real low power, the worst is the expansion, take out anything that you don't use, have 2 of something only because it transition from an older systems, usually is the real hog.

 

[anxiousinfusion]

Rumor is Kaveri will be capable of ~1.05 Tflops. Where would this stat stand if used in dual graphics with a 7750?

 

[noob2222]

Now, the question is, because it was ready, Intel released HSW, even with a few problems, and is it design or fab?
Not counting the USB

I figured Intel would get somewhat close to their claims. what a disaster.
http://www.anandtech.com/show/6355/intels-haswell-architecture/2

Focus on power: FAIL

The more I see haswell, the more fail I see coming true in broadwell. 2014 is already slipping off the roadmap for broadfail.

http://hothardware.com/News/Intel-Broadwell-Reportedly-Delayed-On-Desktops-Until-2015-Reality-Likely-More-Complex-/

 

[noob2222]

^^ If AMD can execute properly, they have some time, Intel is slowing down. SR could be the only exciting thing in 2014.

 

[de5_Roy]

i'm hoping for jaguar successor and kaveri refresh a la richland in h2 2014.
i read that jaguar left quite a bit 'on the table' to get to market quickly (may be the consoles).
and considering glofo's past history, kaveri might deserve a richland-like refresh...
both good.

 

[juanrga]

Also the "hot spot" issues that was present in IB seems to got worst with HW... it might mean (for me is for sure) that intel doesn't have the best fab process around, and makes it worst when a process is tweaked for "low power", which makes it usually not good at "high clock high performance", but is used for all solutions. That is, the process "gets" comfortable only at those low TDP marks, which might indicate that a 84W part (lots of salt of how intel counts this) is already a factory OC part, meaning getting low power marks doesn't mean you have MORE headroom to much higher TDPs, only means different considerations and possibilities ( i think that is why AMD is playing with "Centurion")

You make a number of important remarks. The problems of Ivy were due to fabrication. But from the Xbitlabs link given before:

Another concern for PC makers is that retail versions of Intel Core i “Haswell” chips when overclocked are around 15°C hotter than pre-production samples as well as hotter compared to previous-generation Core i “Ivy Bridge” products that were made using the same 22nm process

This suggests to me that the particular problems of Haswell are due to the new design, specially with how they regulate Voltages now.

Now, the question is, because it was ready, Intel released HSW, even with a few problems, and is it design or fab?
Not counting the USB

I figured Intel would get somewhat close to their claims. what a disaster.
http://www.anandtech.com/show/6355/intels-haswell-architecture/2

Focus on power: FAIL

The irony is that Haswell chips with GT3e are being massively rejected by OEMs due to power consumption:

OEMs don’t seem all that impressed with GT3e, as it's power hungry and expensive. We expect only a tiny number of notebooks will come with GT3e.

http://blogs.nvidia.com/2013/05/qa-why-gamers-still-need-a-discrete-gpu-with-haswell/

 

[JAYDEEJOHN]

Well, pretty much everyone has mostly agreed that Intel has the best process, and their binning is just a matter of point and choose, or at least in the past, there really wasnt much diversity in their chips, as they could all reach a certain level of OC, while a few would go lower or higher.
It appears this iteration/gen is the exact opposite, where it struggles to meet last gens clocks mostly, while some do, and some dont even make what may be considered 200Mhz less than last gen.
Of course the improvements outshine these declines, but it is topsy turvy from what weve come to expect.

My question is, with have v regs on chip, is this why?
I dont believe we will see these problems in mobile, other thatn the GT3e, but thats more a package, and expensive one at that, and is mainly the reasons for the OEMs to not go for this sku, where theyve probably avoided similar setups in the past.

 

[de5_Roy]

afaik, intel openly dislikes overclocking and their focus with haswell is on laptops and hybrids. so far i've read about poor overclockability related information. yet to read anything on retail laptops apart from iris pro reviews. from the anandtech review - it looks like intel hasn't started to think and function like a company that makes graphics processors (may be a management issue? they say that intel has good engineers...?), they're still filling empty die area. there seem to be some glaring shortcomings on intel gpu(gt3/e) design which amd and nvidia do not have. if apple hadn't demanded, intel probably woulda stuck with hd4600 and lower.

i don't think overclocking intel cpus will get any better from now on, but that's in the future (did i think the same thing on 3770k launch?....deja vu...?). i sorta don't find haswell's oc(clcokrates) that surprising. always had suspicion even by my limited knowledge. higher transistor density, higher heat generation per area, 3d transistors etc were right in the front. i am interested to see how the latest info on the gap between the heatspreader and the die pans out.

the article on the oems rejecting gt3e is from nvidia, whose gt650m was intel's target in the gt3e related p.r. nvidia's p.r. is not gonna sit idle and let intel take shots at their gpus. nvidia is losing gpu marketshare to intel and amd for a while, it's important for green team to make it's presence known. in the article, mr. haas is taking shots at all processor graphics as well as bragging about higher nvidia discreet gpu marketshare on intel-based notebooks (which have higher marketshare than amd-based ones), and throws in a 95% number in a suspicious fashion. :lol: it's funny to me because in the event that intel really reaches discreet gaming gpu level perf (overall) nvidia might be the first one to go... amd has already started coupling their apus with radeons and is leveraging their own laptop-friendly technologies.

seems like intel and amd are implicitly deciding on their individual price segment and cementing their positions. intel with $200 and higher in dt, amd in $200 and lower. that's great for budget users who are looking to buy new pcs (kaveri, kabini) and upgrade am3+ pcs. in return, intel is locking amd out of ultrabooks. which is still good for amd, because they have kabini. i woulda said kaveri as well and have little doubt about it's performance but... glofo....

 

[Cataclysm_ZA]

The irony is that Haswell chips with GT3e are being massively rejected by OEMs due to power consumption:

OEMs don’t seem all that impressed with GT3e, as it's power hungry and expensive. We expect only a tiny number of notebooks will come with GT3e.

As de5_Roy mentioned, Intel only made GT3 and GT3e on Apple's request because they were losing out to Nvidia on the Macbook Retina - which is why Iris-equipped chips are so expensive. In fact, I don't think they even cost that much to begin with, as Intel's monopoly in the high-end allows them to dictate pricing. I can't wait to see how Apple's going to react with the GCN parts that now constitute the mainstream HD8000 mobile lineup, because the last Radeon GPU inside a Macbook or iMac was from the VLIW4 HD6000 series. Although the GT650M is enough for the Retina Macbooks, going forward more power is needed.

 

[Cazalan]

Focus on power: FAIL

SA had some power numbers that showed a decent improvement while also offering 30% faster graphics in CS:Go.

The power focus really is for lower performance modes, not when maxed out.
They want to be able to say longer battery life while web surfing or watching videos.

The CPUs may have higher TDP but the motherboard should have less work to do. That's how the overall power numbers are down. As the nodes shrink further this becomes an absolute necessity as having the regulators on the motherboard will just be physically too far from the transistors that need consistent voltage to operate.

 

[-Fran-]

If Intel made GT3e for Apple alone, then they're very stupid IMO. So far, MacBook Pro sales have been going down fast. Apple is competing with themselves using the iPad and iPhone as part of the ecosystem. For the Desktop line up, they don't have any issues, since it's a whole different market there.

Even though it's from nVidia and I still take it with a grain of salt, it makes perfect sense. You don't want power hungry, easy to get hot, chips in your designs for portables. You can put all the big numbers you want, but if you cannot scale it down to a decent power envelope, it's a moot point!

Like I said somewhere else. I'm eager to see the mobile versions of Hasfail. I'm pretty sure we'll get a nasty surprise there. Not bad for AMD, though. The catch up distance would be shortened by a lot.

Cheers!

 

[griptwister]

^^ If AMD can execute properly, they have some time, Intel is slowing down. SR could be the only exciting thing in 2014.

Considering Broadwell supposedly got delayed and is expected to release in 2015 instead.

 

[JAYDEEJOHN]

I look at it this way.
Good enough has been here for awhile.
This is the market, average Joe etc.

AMDs improvements may elevate them to the close enough segment, this is for OEMs, where pricing matters as well as perf and availability, which stung AMD in the past, but no more.
That leaves power/mobile segment, where Intel is king, tho AMD is catching up here as well.
What happens when power/Wh's are also good enough?

These are the paths, markets and their desired needs/wants.
Altogether, AMD seems in fine position.
But, as densities become more a player, this too could halt progress on DT, and only server type chips which gain more from top end chips, which also brings in the margins, where we may see a split greater than weve seen before from DT to server, as its possible densities may require something more exotic than weve had on the DT models, and those margins may not support such chips for DT, as is my worry for design vs fab.

 

[MU_Engineer]

Yes but they could had trashed the Intel 6 core/ 12 thread "enthusiast" solution of Intel... 2x 3 modules/6 core chips on a MCM package, could go or surpass yet the 4.0Ghz barrier, power will be higher, but this would be a better solution than Centurion, a REAL 12 core G34 solution at 4 to 4.2GHz will be no match in multithreading for the intel solutions.

And the chipset could be the same FX990, the FSB would be the same Hypertransport link of server, that is, no need and not recommended to have a server chipset.

The problem is that multithreading is not really a problem for AMD even with the current 4-module monolithic CPUs. They do very well with the multithreaded "productivity" programs. Their issue is that they don't do so well in poorly-threaded games and benchmarks, which to add insult to injury, are often compiled with Intel's ICC as well. You won't get any higher clocks or better single-threaded performance out of G34 compared to AM3+ and in fact it would likely be *lower* clocks because of the added complexity and power consumption/thermal dissipation of the second die. It would thus be "worse" in the eyes of the shill reviewers like A*******h than the AM3+ FXes because productivity/multithreaded doesn't matter (you can/should do the former on a tablet and GPGPU is the Holy Grail for the second) and all that matters is getting 260 fps in a game at 1024x768 rather than 240 fps.

Yes but those benchmarks are NOT REPRESENTATIVE of current software used, least of all with Linux where many of this systems would end up (FX could win there), and a workstation market is where there are more aware and knowledgeable users.

G34 around 220W would be a much better solution than Centurion now, it may not reach 5Ghz, but it would have 12 cores and the package be above the 4Ghz yet (2 FX6000 serie chips on a MCM, each above 4Ghz is quite possible attending the improved 32nm SOI of Richland). BETTER.

I agree, for somebody like you or I who mainly run "real" programs on Linux, a super G34 (or 2 or 4 of them) would be much better than the Centurion becuase it would boost multithreaded performance a lot and that is what matters. Tech sites are paid for by traffic. The big popular sites are all predominantly gaming/overclocking/benchmarking oriented. There is only one I can think of that even touches on Linux, and that's even a partially subscription funded site (Phoronix.) Unfortunately many people seem to care about the benchmark bragging rights a lot, so the fake/canned benchmarks really DO mean something to them.

Some people have OC'ed current G34 chips to the mid 3.5 GHz range - and once there they are the fastest single (4P) node setup (faster than Intel's 2011 4P which has no OC capabilities). Unfortunately they are at the limit of motherboard power circuitry (entire system over 1300W) with current silicon; but if an improved process could get them closer with reasonable power consumption they would be amazing. Hopefully we will get that - plus even more performance per clock - out of SR.

The current G34 chips actually operate up to speeds of 3.5 GHz sustained at stock and can use CPB to nearly touch 4.0 GHz. You are thinking about the older Magny-Cours chips which were initially offered right around 2 GHz but could be overclocked in certain Supermicro boards to the low-mid 3 GHz range. The current 6200/6300 Opterons likely could go past the mid 4 GHz range if you have enough cooling and your board's VRMs are stout enough. But few have tried to overclock them since the famous overclocking BIOSes only support Magny-Cours. You could grab one of the many South Korean Opteron 6200/6300 ESes off Ebay and overclock them in software though. I am almost tempted to do that but am too chicken to drop $150+ on a chip that might be a dud.

You raise a very good point hcl123, but I'm sure there's a very practical reason behind the G34 being left out for servers only.

It could be the incredible fail that QuadFX was, or maybe the fact that Intel's "Extreme" line sells little in comparison to the "mainstream" i series that it could not be worth the effort from all that execution.

I'm really not 100% sure, but it is indeed a very good question to throw out there.

Cheers!

QuadFX shouldn't had to be a failure. The management of AMD was clearly a failure. QuadFX was scrapped, because that was a time when AMD was considering to sell its chipset division, and with that scrap HyperTransport( meaning now HSA chips would not be possible), seamless multi sockets would be scrapped (depend on a few suppliers for this for server), while client would have to rely on general 3th part offerings based on what they would offer, meaning usually single socket PCIe all around
(if they had sold the chipset division, there wouldn't be AMD by now, no place to put a CPU or an APU, safe the very worst, and blame it on the chips not chipsets lol)

Just to imagine how much better Hypertransport (torrenza interconnect, based on work of Jim Keller) was for those multisocket approaches compared with the clumsy shared FSB topologies of Intel... yet Intel had one double socket for client while AMD had none !!!

The propaganda (most sites that most of you point now indicating that intel FSB is better and if not, not worst LOL )... the end of cash... the market distortion that lead to the famous law suit against intel that intel bought out for peanuts...

Elas... G34 is quite a much better MCM than any FSB could had been, or even intel QPI can imitate. G34 is left for servers in a logic consequence of all that... markets that AMD already lost or scrapped. But if they want to try a come back of sorts then G34 is much better than AM3+ for sure, Centurion or no Centurion (the future is multithreading not single-thread), specially if rated TDPs is not an issue, as it usually isn't among OCers.

QuadFX failed becuase Windows XP hated NUMA and decreased the performance of two FX-7xs to less than that of one chip by itself. The QX6700 didn't have to use NUMA and thus did better. Also the FX-7xs didn't overclock well and Kentsfield was pretty well known for its overclocking prowess, rubbing dirt in AMD's eyes. If you were testing a QuadFX today vs. the QX6700 and had them at stock, the results would be much different as Windows 7/8 support NUMA much better than XP ever attempted to do so. The QX6700 would still likely have a little edge but it would be pretty close and not the embarassment that it was during the initial launch.