9950 AMD Overclocked it not what you think - Page 3

Last message on previous page:
I love how the forums on all sites just show how loyal intel fanboys are. Remember how they got caught blackballing AMD? Paying other companies to use their stuff instead of AMD's. There are no "fair" benchmarks out there these days anyways. when people benchmark different processors they use different mainboards and different memory and different cards and expect it to be fair. what a joke!

jed, my point is, there is an actual difference between what's being compared. AMD benchmarks go against higher clocked duals, whereas Intel is competing against its quads. AMD's have a reasonable advantage in future-proofing because they're quads, but Intel is already against quads, so it's not an issue of quad to dual, it's a question of C2Q vs. Ci7. In that question, AMD suffers from the same conclusion any dual vs. quad does, faster dual for now, quad for the future. Intel loses because they're both quads, and i7 doesn't show itself to be faster.

I hope you can understand that.

let me see i7 is faster than everything out there in regular and sever
apps. The only thing it don't win in is games against C2Q.
Now i do a lot more than play games on my computer, so it will be
some what the same in games and a lot better in everything else.

@jaydeejohn

I commend you on your coolness and calmness you've exibited throughout this entire thread, I first was attracted to it because I'm running a 9950 Phenom, as my choice of cpu, and I guess to a certain extent I am partial to AMD, though the majority will think I've lost my mind by not building an Intel machine instead.

I understand what you've been trying to say even though others just want to make their points as how it always seems to turn out when the Intel fanbois show up.

We overclockers are a minority in the computing realm as a whole, of all my friends, family, and aquaintences, computer modding friends, I'm the only one that overclocks!

I'm saying that because average Joe, doesn't even know what overclocking even is, so to them overclocking isn't in their vocabulary, what I mean is whats seemingly so important Intel vs AMD, doesn't mean a thing in the mainstream population.

I've built many computers for customers and never overclocked one for any of them, I've never been asked to overclock a machine for someone, a person has to know what overclocking is to ask you to do it.

My point is how fast is fast enough and when will any of us be satisfied with what we have???

My Phenom is not the fastest CPU, I new that before I bought it but it does what I want and need for it to do.

I do remember the past though and if AMD manages to repeat the past the Intel fanbois will be singing different tunes, like when the FX57 stomped Intel in the dirt, no fanbois can argue that point the FX57 ruled for about a year.

I hope AMD turns out to be a true Phoenix!

This is just a one time comment and a pat on the back for jaydeejohn!

Deneb is 7% higher in what, IPC? That might just be from tripling the L3 and raising it by some 200MHz. Not the actual die shrink, though that helps to make all this L3.

I sure hope Deneb offers more than 7% performance over Phenom, by increasing clocks and headroom; otherwise it would still follow the steps of Kentsfield, which in 1-2 months celebrates its second anniversary. From a bird's-eye view, the Deneb core looks identical to the Phenom core, so I'm hoping that means maximal effectiveness in ramping their 45nm process.

I'm hoping for more than the estimated 15% performance increase. To bump up performance a mere 15% over a year (no extra cores) worries me a lot - the competition isn't idling. They're not saving much die space with Deneb, either. Some die shot analysis this spring:

Barcelona 65nm 285 mm^2
Shanghai 45nm 243 mm^2 (15.3 mm^2 cores)
Nehalem 45nm 246 mm^2 (24.4 mm^2 cores)

Barcelona went after the money; had K10 been a desktop oriented CPU, I'm afraid AMD would have been deeper in the red, unless it conquered the enthusiast market over Penryn.

Intel heading for the server market is probably what's worrying everyone at AMD. Shanghai is supposed to go after the server market (it just came out), but when Nehalem comes out (in a few weeks), what exactly will happen to sales of Shanghai?

jimmy, I agree that i7 is a MT monster, and MT has its obvious advantages. Its just currently, its not well used, and as for gaming, that hurts i7's attractiveness to many a gamer. Weve had tons of threads asking which is better, a dual or quad for gaming, and it still is being asked, for over a year its been asked.

Now, when I say Im disappointed in i7 as far as gaming goes, thats overall, not particulars. In games that use MT, it looks to be the winner I thought itd overall be, but the ST is showing less than Id hoped for a brand new arch. Its no biggie, really, just unfortunate for us gamers for RIGHT NOW. Down the road? Who knows?

Looking at what Valve wants to do, creating incredible AI and such, which is primarily MT, it looks (i7) very promising, but again, thats then not now. The elevated costs of an i7 system inhibit its attractiveness even further, and its not just me saying this, but alot of people. I dont buy all the crap thats been said about its costs, but we can all agree its going to cost more than usual, and by that I mean, what weve been accustomed to.

As far as Deneb goes, I wish like everyone else, and also about i7, that we had more info/benches. If I were to JUST compare the IPC improvements between i7 and Deneb, then Id say its a scratch in most ST scenarios, BUT, again, i7 is a whole new arch, whereas Deneb is but a die shrink, with added cache, and beyond that, it could just be a dumb shrink, as we dont know.

So, when I say Im disappointed about i7, and have more enthusiasm for Deneb, its purely for AMD, and their users. I see the higher clocks as something thats truly needed, and from what we know, it looks like Deneb may deliver them. IPC, at least from what weve seen in ST looks like a draw, and thats good for AMD and their users, and thats ID AMD comes thru and prices Deneb properly, because if they dont, again, Ill be disappointed in AMD as well.

Just remember, Im not the only one that feels this way towards i7, and what Ive said about AMD should have no bearing regarding i7, as I know, you know, we all know (except a few mentalists, and not the kind that can move things) that Intel is leading the pace, and AMD, like youve said , has a long ways to go, but at least theyve got their speeds up, now onto IPC

I look at it like this. Way back when, everything ran on steam, until the combustion engine was made, and thats the last true dynamic change weve seen for many a use, be it cars, trains etc Changes like that dont come often, but when they do, people had to learn how to make diesel trains etc, so to, the SW folks will have to get on board with MT, as the single core cpu, single thread idea is dead, and its time to move on, its just going to tale awhile, because we have a distorted view at how often the public changes their computers, as we do it all the time.

If the public moves at a certain speed within a market, buying new goods in that market, companies move at the PUBLICS pace, not their own capable pace. Having said that, theres many knowledgeable people at Intel, and every tic and every tock, its what 2 years between tocks? So Im guessing, at Intels current and projected production speeds/capabilities, every 2 years they CAN allow for new arches, as sales limit them to this pace. Thats why Ive said i7 is before its time, by a lil bit. Thats why we wont see MT tomorrow, but after awhile. Doesnt mean Intel cant do more, its more business climate (sales etc) control its tech as well as its growth.

Ill point something else out. The home user no longer drives these sales/improvements/arch, but businesses do. Both AMD and Intel have realised this, and their current cpus show it. The home user used to see huge benefits from newer arch's, faster speeds , better IPC and so on, but no longer. We here want this, need it in some instances, but the average person just doesnt need everything both Intel and AMD are making today.

So, as business requires, or sees benefits from these improvements, IF they can be applied to DT, then we will also, if not, itll be more like what we see now, some good, some well never use, but its still there, or the cpus will still be able to do things well never use, and that itself sucks for us, but is good for businesses.

The Q6600 is a Multi Chip Module. It is not quad core jed.

http://www.youtube.com/watch?v=Ivx0XYMCZJw

OOPs

Thank you.

Im afraid as long as the Phenom B2 miasma persists (which it will for years to come, and thats assuming AMD actually delievers the goods with its upcoming releases), the AMD fanboys will grasp at any straw, no matter how unrealistic, with which to bolster the image of their "precious". As such, dont bother getting your hopes up that they will suddenly let go of the Henri Richard brainstroke "multicore for dummies" native quad shtick. Even though it every metric proves it is meaningless, even though in terms of yield its a nightmare for a company with manufacturing difficulties, it is the only thing the fanboys have, therefore they will continue to cling to it as a drowning man would cling to a life ring.

I'm an AMD person and I'll say this: Intel made the right decision in not going native quad before Nehalem. AMD probably couldn't go the route of Intel because of the memory controller being on the CPU and not the Northbridge, but I'm not sure of that. I vaguely remember reading about future AMD CPU's having two native quads packaged together.

The Q6600 etc. are quads. They are true quad cores. They aren't native quad cores, as that refers to AMD's Phenom and Intel's Nehalem. Not all AMD fans buy all the marketing, but I noticed back in the Prescott days that almost all Intel fans bought the marketing. Must be because Intel has a better marketing budget. All AMD had to defend Phenom B2's when they came out was spin.

Agree with your post, and AMD deservers credit for trying. The Phenom B3 is what Phenom should have been on release. No TLB errata and slightly better stock clocks and overclocking ability, though thermals are still high.

All I want AMD to do is improve on their last architecture. B2 had higher IPC than Athlon X2, B3 fixed B2 issues with some improvements and Deneb will be a die shrink that will improve things further. All in all, Phenom's a good budget CPU. My 8750 is competitive against similarly priced Intel quads and beats higher clocked Athlon X2's where it matters most to me.

The whole packaged vs. native, what is a quad core argument is a bit old. The SOI vs. HK has merits, and I can't wait to see the promised second version of Deneb using some kind of HK process. That should be a good step towards their next architecture in 2010.

Im sorry but I must disagree. The trend towards multicore was predicted years ago when it was realized that Moores law could not be sustained. This really shouldnt have been a suprise to anyone as possible reductions in lithographic nodes were and are known to be finite. The only possible way to continue increasing the number of transitors was to go multi core. And even this is finite based on surface area. To continue progress, the next step towards multilayer must either succeed or progress will stop and Moores law will become Moores failed theorum.

As for demand for dual core subsiding, quite the opposite is true. While it is difficlut for the "enthusiasts" who frequent hardwqare forums to understand, The vast majority of the worlds population niether owns nors has access to a computer. As always, people seem to forget that the market is constanly expanding. In order for manufacturers to remain healthy and viable, they must either expand into these markets, or fail. And they themsleves are quite aware of this and have been working to open those markets. With the natural expansion of the computer market into the low income bracket, demand for cheap, low to moderate power (computational) processors will (or would have prior to the recession) increase exponentially as that particular segment composes the vast majority of the potential total market, itself being composed of those who have no need of quad or even dual core processors. Obviously. Either production capacity must increase to meet the increase in demand the low income market presents, or prices must remain high and accordingly that market will remain untapped. The real race is to fill that market...high volume low margin, as the higher end markets, including both server and mobile are far more limited.

As far as the native design being far superior to the MCM, this is wholly incorrect. Both designs have advantages and disadvantages resulting in net equality on the manufacturing front. And frankly, Im quite suprised that you seemgly dont know that the true advantage of native is in fact a performance advantage. Natives chief advatage is NOT in manufacturing, but in bandwidth, simply because native dies readily support IMC whereas MCM does not. This is not an advantage in DT, but is an undisputed advantage in the server market. This performance advantage is so significant that AMD still holds a strong position in server while their DT market has been bleeding to death.

While MCM does require more manufacturing operations, exclusion of the MC reduces die complexity itself. Furthermore, MCM provides both significant reduction in waste and at the same time increases the total number of possible component combinations and thus the number of possible varients, resulting in broader markets coverage without the attendent increase in manufacturing/design costs native requires. Both AMD and Intel have already demonstrated these effects. In AMDs case, they have already demonstrated what happens to a native quad that fails QC. It becomes a tri core. While the AMD marketing wizzards have tried to 'invent' a market for tri cores (smartly but ineffectually), it really doesnt exist, and the sole purpose of a tri core is to recover the manfacturing overhead and material investment in failed quad core dies. Intel on the other hand has clearly demonstrated the manufacturing flexibility of the MCM by flooding the market with different versions of both quad and dual core. Frankly, too many AFAIK, but that was their choice, not mine, and an option afforded them cheaply only by use of MCM.

With the current direction of research and developement going on in mulitlayer modules/etching processes/multi logic cores, the day of the native core itself may be numbered. Should this research successfully reach maturity, one possible out come is independent logic cores covering all aspects of operation. Weve already seen this direction from both Intel and AMD with GPU/non CPU logic core on CPU die. If this continues and progresses to multilayer, it is entirely possible if not probable that IMCs as well as all other components will be segmented, and placed on seperate layers, allowing customization of other layers. By using MCM in multilayer configuration, manufacturers will theoretically be able to retain IMC in an MCM design, allowing maximum flexibility while retaining the most notable benefit of native coes.....not manufacturing but bandwidth. While this will further increase complexity of the manufacturing process, at the the same time it will increase manufacturing flexibility allowing all products of a generation to be produced on a single line regardless of core count and core variation, and do so with a single set of masks.

I agree with you. Intel has always had the vastly larger marketing budget, but back in the Prescott days AMD gained market share and grew a base of loyal customers despite Intel's massive marketing. Why? Because AMD had something more powerful than marketing; they had a better product.

Yips we have been over this many times. The biggest problem AMD had during that period that stopped them from taking more than 33% of the market was their lack of ability to produce enough chips. Thats the biggest problem for them. But now they have the foundry or whatever so next time, if the foundry system can produce more chips then they will gain more market share.

And wr, your pic doesn't work so Imma help you out:



Naitive hexcore. Pretty and shinny too.

The plant in Saratoga hasnt been built yet (and still not a 'done' deal) and AMDs 45nm isnt in retail yet. Additionally, producing chips is one thing. Producing chips that function to design spec is another, and AMD has demonstrated problems with this at 65nm, so there is no reason to beilieve they will not encounter difficulties at 45nm.

Not adding variables, not even remotely. Native without IMC is only 1/2 of the argument. Theres is nothing sensitive about it. Native simply has that flexability. MCM on the other hand is adifferent story. MCM does not readily lend itself to IMC (at this time) and it is precisely that which is the limiting factor with current MCM. Were it not, were Intel able to achieve the memory bandwith with MCM that AMD can achieve with Native, I doubt i& as we know it would exist. There would be no reason. Futhermore, I doubt Phenom as we know it would exist.

Add the manufacturing process which requires the least number of vairables is not always the most desirable if it limits options. One only need look at Henry fords production line, the basis for modern manufacturing to see this demonstrated. Were this true, there would be independat lines for the Ford Crown Vic, Mercury Marquis and Lincoln town car, yet they like so many other models, are all made on the same line as using common components reduces the number of lines neccessary to produce multiple products without stoping the line and retooling. While auto manufacturing is vastly different than CPU manufcaturing, the same basic principles of manufacturing apply. That is not "spin", it is merely the economics of manufacturing.

Higher margins does not mean margins are high enough. Even with a high margin, flaws still occur and can be estimated/calculated on a surface area percentage. By locking in to a native design, and incresing the surface area consumed by individual dies, the percentage of wasted dies increases. This is simple math, and not contestable.

jimmy if you look at that die shot carefully you can see Tron on his AMD bike ... being chased by Sark's henchmen.

Nice pic that ... very nice.

Someone should put a sticky at the top of the CPU area so people can contribute good pictures like this one.

Shiny ...

This is what disappointed me the most:

Both AMD, both using nVidia 7950GX2's (2006), both in 3dMark06

2005 era 939pin 3700+ (OC @ 2.8ghz) vs Phenom (10% overclocked)




unless your really into converting DVD's into MP3's, its THE SAME FPS in games!



.

DON'T STONE ME......I'M NOT A FANBOY!!!

I think it comes down to user needs and budget. For me, the 9950 is great. Does everything I need it too. Runs great FPS in games. Also can handle playing Blu-Rays on my 32" TV at 1080P while still running a processor hungry program on the main display. Granted I have the 4850 GPU. Sorry, I had to throw that in. It may not be as fast as a $329 C2Q, but for $175 not bad.

Far Cry 2 runs better on Phenom 9950 then Athlon 6400+, period.

You slap in a Q6600 on the same test and you'll get basically the same thing except slightly better-than-Phenom result in CPU test. The days of dual core gaming is over. New titles, such as GTA4 and Far Cry 2, will run better on more cores.

Also, there are reasons to use more cores. I usually use my computer as TV at the same time I read websites and posting crap. Extra cores do help when you have both TV running election report and HD movie at the same time while running spyware plus anti-virus at background.

My appologies for not responding sooner, I've been out of town on business.


First, you are quite correct,

but you said you think you get my point, and I believe you do. While the manufacturing line (etching, cutting etc) wont change, and MCM does complicate the assembly, even discounting higher geometric yield and increase of wafer area usage it offers, it allows the potential for exanded product lines/branches. To offer different product families with native, beyond binning speeds and fuzing cache or cores, (which are all just as applicable to MCM) you either have to open a new line, or swap a line by 'retooling" i.e changing out masks etc. Who knows, they may make that easier as time passes, but given the nature of the etching environment and process, it seems far easier/less costly to set a line and leave it for the time being.

Additionally, as I pointed out, with the litho wall fast approaching, the native MCM argument goes beyond the "now". Both AMD and intel have to choose between 2 options (that I can think of) to increases transistor count in the future:
-Go to larger dies, which means new, larger sockets which will eventually mean larger motherboards
-Expand 'vertically'; layered CPUs

Another option which can increase system performance
-Go "mainstream" multi socket - doesnt increase the number of transistors on a die, but gets more transistors on the motherboard.

2 additional options which dont increase transistor count:
-accept the litho node "wall"
-push on to optical processors or some other technology

-Now, Ive seen very. very little on optical processors. Not that theres a lack of information, just that I havent read it, so I can not debate from any solid ground on those....they may be the second coming, they may be failures, I dont know.
-I doubt either Intel of AMD will 'accept' the wall, as long as both are in business, but should one fail leaving the other a monopoly....well, you can guess the results there.
-I cant see them increasing die/package size to the point it requires larger mobos....but who knows
-Multi socket is a cheap way for the manufacturer, but not the consumer. Not an unattractive short term solution, but who knows
-Layered CPUs...High initial invesment in R&D, and a waste if they cant make it work. If they can, IMO, in lieu of optical, thats the future.

On the quad core becoming mainstream....are you saying they are supplanting the dual core? If so, see the previous post. I wouldnt count on that anytime soon, unless I was an enthusiast, and Im not.

On the scrap rate, its QED. (and you have to think in terms not just of wasted dies, but in wasted wafer area)

None of the following includes design or process flaws, just surface imperfection from contaminates. And, its purely for demonstration purposes and purely theory.

http://www.bit-tech.net/hardware/2 [...] _x3_8750/1

For an X3 @65nm

BTW, take it as you will, but these guys say "confirms" on the disabled quad core. I know, I know, but there are still those who deny it.

and to back that up, from Legit Hardware:

http://www.legitreviews.com/article/597/1/

So can we agree that the die size is 285 mm sq @ 65nm?


Now, on a C2D @ 65nm, from Anand we have

http://www.anandtech.com/cpuchipse [...] spx?i=2795

So heres a little graphical comparison



First, I cheated like a SOB in favor of AMD
Assuming for both AMD and Intel:
300mm wafers.
Edges perfect.
Cheating in AMDs favor: Zero loss cutting (dies abutted)
Intel: add 2.7 mm sq per die to account for cutting
For both, symetric distribution (no shifting of columns/rows), right up to the edges, which favors AMD as its die are almost square


AMDs die almost perfectly square @ 16.88 mm x 16.88mm
Intels die 10.52 mm x 13.63mm ( a little over )

Based on the die sizes, and cheating as stated above, this gives AMD 221 quad core die per wafer, and Intel 448 dual core or 224 quad core per die, from a 300 mm Dia or 70686 mm sq wafer. Just for arguments sake, assuming a .01% surface defect area for each wafer, that gives 7 mmSq. Pulling an average .4 mm radius per affected area for each defect (purely out of my behind) that results in 14 imperfections per wafer. For the moment, lets assume no cross die imperfections. This means you loose 14 quad die fromAMDs wafer, or a 6% waste rate. For the Intel, the loss of 14 die gives a 3% waste rate for dual core or 4% for quad, + 1 functional dual core remaining.

There is a formula for loss based on total surface area vs % bad based on sub section area, but I dont remember off the top of my head....been a while since I played with that area of calc, but I think it works out to 85% of total for a doubling of size...the probability of 2 flaws occuring in one area doubles for a doubling of the area, which reduces the loss, but is offset by the reduction in division which reduces the total # of sub areas. Perhaps one of our probability/stats members can drop us that theorum. That still leaves AMD @ a 5% waste rate.

Now, going back to simply geometry, lets talk about waste...unused area.

If we were to etch a single square die from a 300mm wafer, as large as possible, it would be 212.4mm per side for a surface area or 44986.4 mm sq. A waste of 36% of the wafer.
If were to etch multiple die at 1/6 that size (35.4mm x 35.4mm, (1253mmsq)) that would give us 36 die totaling 212.4mm x212.4mm. Additionally, we could fit 8 more die for additional 10026 mmsq., or a total of 55139 mm sq used, an additional 12% of the die used.
If were to etch our die at 1/12 that size, we increase our used area to 60151.7mmsq, or an additional 7%.

Simply put, because we are stuffing squares or rectangles into a circle, the more we can reduce the size of the of the square or rectangles, the more total area we can use. In the Intel wafer I used in the example, by offsetting the rows, we could fit an additional 4 dies for a total of 452. If I were to stagger/rotate, I could fit a fair amount more. Cant do that with the square 'native' quads. If we can get the die small enough, we can stagger rows, but that doesnt help nearly so much as stagger and rotate.



Whats all this mean? Simple. If AMDs margins are high, nothing. If their margins are against the bone, every sq mm counts. Native wastes more than MCM, both in flaws and unused wafer....even assuming identical manufacturing process and flaw %s.

The point I was making is that I strongly suspect that the additional labor to make an MCM will cost more than the increased wafer scrap would cost you. Labor and overhead are expensive and I would guess that they likely comprise a significant portion of the manufacturing cost for a processor.

I understand, but I disagree. I believe the reduction in waste (both wafer and and lost dies) coupled with the ease of design, plus the flexability allowing MCM to offer greater variety makes Native and MCM even in terms of manufactring benefits detriments.

In terms of performance, for now there is no comparison. That capacity of Native to easily accomdate an IMC gives it the advantages of the IMCs bandwidth. Another plus for native, if they chose, they could just as easily use an offdie MC....but then, why would they want to? MCM could use a IMC....but the cost/benefit ratio of trying to pull that of is prohibitive, and I think it'd wind up being another crapburst/phenomB2.

Ultimately, I dont see a lot of 'great' options down the road to increase transistor count without going to layered MCM.

I am curious what you think will happen when AMD and Intel hit the litho node wall.

Yep, wait and see. Still a lot of questions about what Deneb will be.

This paragraph needs more. It's just mentioning that i7 doesn't do too well in gaming, but not actually saying anything. And how come CrossFireX is a whole beach, but SLI is just a few drinks? This whole conclusion is very different from the one I came to.

Long story short:

Intel has done some shady business practicies, but AMD is ultimately responsible for its rise (Athlon, Athlon XP, Athlon 64, Athlon 64 X2) and its demise (K10). AMD shot itself in the foot.