CPU Stepping/Revisions

[sb.0326]

After doing a quick Google search I found that newer steppings often times yield better clock speeds and/or power consumption. What I did not find was a lot of details on what specific steppings changed or how they improved on the processor. Just that they did.

I have a Crosshair V purchased prior to Bulldozers disaster. Was one of those that had hoped it would live up to the hype. With that said I am currently happy with my Thuban at 4.0ghz honestly. Does a more than adequate job for me at the moment. But I will be looking to upgrade(duh) in the future.

My question is this.. What exactly are realistic expectations for Bulldozers future revisions?

Can they increase the speed of the L2?

Can they reduce the power/thermals of the chip significantly?

Can they increase performance outside of the traditional clock speed improvements you see with newer steppings?

Am I better served getting the last Bulldozer shipped or just getting a Piledriver based core?

I'm hoping that they can fix any or all of the above.. Cause.. That Intel 2011 platform is pretty sweet but oh so expensive..

 

[AtaliaA1]

Realistic expectations for the Bulldozer= Coming closer to Intels specs.

L2 increase= adding nand necesitate rebuilding the chip

Power/thermal reduction= again rebuilding the chip (smaller form factor/other type of metal)

Performance increase outside of clock speed= yes Instruction set similar to Intels

Piledriver or Bulldozer= Piledriver Better off getting an Intel i7 960 Bumping up the clock a bit and or using the XMP Speed boost in the Bios. Even the Boutique builders are still offering the i7's No significant improvement over the former chips. Some but not enoght to justify the expense. Take a look at the Reviews and comparisons.

 

[frostedflakes]

The difference between C2 and C3 stepping Phenom II chips would probably be a good example of what to expect. There was a tangible improvement, the C3 chips ran a bit cooler and overclocked a bit better at the same voltage (100-200MHz higher if memory serves, I think the C2 chips tended to max out at ~3.8 while the C3 chips could get up to ~4.0 on average). Nothing game changing, but they were better.

Piledriver is promising IPC improvements, but the community was promised IPC improvements with BD as well and we all know how that turned out. So I wouldn't get your hopes up, hopefully Piledriver will improve on some of the weaker points of the BD architecture, but I'll reserve judgement until reviews actually start showing up.

 

[sb.0326]

Kinda why I'm leery about what has been promised for Piledriver. Not sure why increasing the L2 speeds is out of the realm of possibility. Unless it's just cost prohibitive. And many places have pointed to the slow L2 speeds as "part" of Bulldozers problem.

 

[TuxDave]

I really wouldn't count on anything more than the best case 10% thermals and clock frequency. I highly doubt project management is going to have the team take the time required to repipeline anything instead of putting that time towards making sure it doesn't happen on the next project.

Most of the low hanging fruit has already been done prior to pre-production steppings and so I really doubt there's any non-fatal logic bug that they can fix that'll solve a ton of performance problems.

 

[LoneNinja]

C2-C3 Phenom II I'm not sure what they changed, but they managed to lower the required volt and TDP a little. The C2 965 was rated at 140W and C3 125W, overclocking improved too. Very few C2 chips get past 3.8Ghz, C3 usually reaches 4.0Ghz, and with the lower volt/tdp power draw was down some too. My Phenom II X4 965 C3 is running 3.8Ghz @ 1.4V and my X4 940 C2 isn't even stable at 3.8, instead it runs 3.6Ghz @ 1.425V and hotter too.

Bulldozer should improve with a new stepping, maybe we'll see 5.0Ghz overclocks or lower power draw, but it won't be a dramatic improvement. Process maturity will probably do more than a new stepping will.

 

[sb.0326]

Are the revision\stepping changes a result of strictly manufacturing process improvements then? So it's GloFo that makes these changes then and not AMD?

 

[Don Karnage]

BD's two main problems are Power Consumption and IPC. Lower the consumption and increase IPC by 20% and you have a winner

 

[Idontcare]

BD's two main problems are Power Consumption and IPC. Lower the consumption and increase IPC by 20% and you have a winner

I think what you are really getting at is that Bulldozer has one problem: performance/watt

To resolve this problem with bulldozer (changing the microarchitecture to the point needed for a 20% IPC increase would render the product something much different than bulldozer) they need to improve the power-consumption at any given clockspeed.

Increase the clocks, decrease power, or a combo of both.

And it is doable. AMD did it before, on 45nm, in comparing the initial release of deneb versus the followup release of thuban.

But it will still take a year, as it did with deneb and thuban, so there are no short-term miracles to be rolled out.

 

[Don Karnage]

But it will still take a year, as it did with deneb and thuban, so there are no short-term miracles to be rolled out.

Agreed. This isn't going to be a short term fix. What made Amd go in this direction anyway? Why didn't they just do a die shrink of Thuban and improve upon it?

 

[Idontcare]

Are the revision\stepping changes a result of strictly manufacturing process improvements then?

That is virtually never the case. Revision/stepping changes involve physical changes to the mask-set that is used in the stepper (litho tool)...those photomasks costs lots of money, they don't change them without good reason.

But other changes that do happen to the process integration - tweaks in implant for parametric yield improvements, tweaks in clean steps or deposition processes to improve functional yields - are done without requiring a new revisions/steppings.

Those things happen in the background so to speak. Sometimes, not frequently, the two will be done hand-in-hand. But those occasions are usually at the behest of a necessary (i.e. painful, expensive) change to the design rules that comes with a change in the underlying process integration.

The one example of this that I know you guys would more likely be aware of is the 130nm thoroughbred (T-bred) which had two revisions that differed markedly by the fact the second revision of the T-Bred (called Tbred-B) sported the addition of a 9th metal layer which required both new masks as well as new process integration.

There came to be two steppings (revisions) of this core commonly referred to as Tbred-A (cpuid:6 8 0) and Tbred-B (cpuid:6 8 1).[21] The initial version (later known as A) was simply a direct die shrink of the Palomino, and demonstrated that AMD had successfully transitioned to a 130 nm process. While successful in reducing the production cost per processor, the unmodified Palomino design did not demonstrate the expected reduction in heat and clock scalability usually seen when a design is shrunk to a smaller process. As a result, AMD was not able to increase Thoroughbred-A clock speeds much above those of the Palomino it was to replace. Tbred-A was only sold in versions from 1333 MHz to 1800 MHz, and was only able to dis-place the more production-costly Palomino from AMD's lineup.

AMD thus reworked the Thoroughbred's design to better match the process node on which it was produced, in turn creating the Thoroughbred-B. A significant aspect of this redesign was the addition of another ninth "metal layer" to the already quite complex eight-layered Thoroughbred-A.

So it's GloFo that makes these changes then and not AMD?

GloFo will make changes as a last resort if for some reason their process technology is failing to deliver the device paramterics (leakage, drive currents, reliability, etc) that they specified their tech node would deliver.

However, since the event of missing such key deliverables is extremely frowned upon by fabless customers in the competitive foundry industry, it is usually prioritized and resourced well enough by the foundry during the R&D phase of a new node such that key parametrics are delivered.

Not that it doesn't happen ever, but when it does it is a real deal-killer...IBM's foundry efforts never really recovered from the stumble they took with SiLK dielectric at 130nm in part because they completely eroded customer confidence in their ability to deliver on their node specifications and technology commitments.

So, given the penalty of failure, one can rightly expect GloFo to have resourced its R&D team sufficiently enough that any major changes to the process integration side of the foundry equation will be doable without requiring design rule changes or their customer's doing respins the likes of IBM's stumble with Xilinx and SiLK (in my opinion).

 

[Idontcare]

What made Amd go in this direction anyway? Why didn't they just do a die shrink of Thuban and improve upon it?

Well they did, if you think about it, in the sense that they did make Llano and it does have roughly 6-7% IPC improvement over the thuban microarchitecture.

But obviously they had no intention of positioning Llano as a thuban successor, as evidenced by the design objectives targeting mobility (power consumption, core counts, etc) at the expense of high performance.

IMO what happened to AMD's bulldozer aspirations, why it looked good on paper, good enough to convice AMD's decision makers to pursue creating it, is probably similar to what happened to Intel's Prescott aspirations and the reasons it came up short as well.

The bottom line is that designing multi-billion xtor CMOS devices is rocket science...this sh*t is hard, its not simply a matter of it being expensive and time-consuming, you can still throw billions of dollars at it and have a final product that fails to perform to expectation despite all the modeling and simulations beforehand.

Its the very same reason you only see two companies attempting to do it in the high-performance x86 arena. Its the reason Via doesn't even attempt to compete, and everyone else that still holds an x86 license considers it a non-starter.

Its a difficult thing to do correctly and to do in a way that still leaves you in a position to make money.

Even Intel got it wrong with Prescott, a chip that was a mere 125m xtors and a bush-league 122mm^2...compare that to the complexity of the challenge of getting a 1.2B xtor 315mm^2 chip functioning as intended

 

[sb.0326]

My god.. Thoroughbred.. That takes me back..

Idontcare, out of curiosity what sort of background do you have with this. For me it's just a hobby hence my ignorance in a lot of areas. And thank you for the detailed answer.

So I guess their is "hope" that Bulldozer can be fixed after all then.

Isn't AMD's and GloFo's relationship eroding as we speak though? Thought that Bulldozer was supposed to be very good on power draw. In part because of the shared internals. Other part manufacturing process upgrade. Is the failure on this end more on AMD's end or GloFo's?

 

[Idontcare]

Idontcare, out of curiosity what sort of background do you have with this.

PhD in Chemical Physics, BS in Materials Science and Engineering, ~17yrs at Texas Instruments involved in technology node development from 0.5um to 32nm, worked directly with node development engineers at TSMC, UMC, AMD, Chartered Semiconductor, freelanced as an industry consultant for tool vendors, Adjunct professor at University of North Texas, lots of patents, blah blah blah.

Basically its my life. If I wasn't extremely interested in this stuff then I would not have pursued a career in it, and if I didn't know what I was talking about then I would not have had the opportunity to have a career in it in the first place

The coolest experience for me though is in having the background and experience that I do which enables me to be able to provide answers to individuals who don't have the benefit of having had my career but are every bit as inquisitive and enthusiastic about technology and this industry as I was some 25yrs ago.

Only I didn't have easy access to a veteran of the node development industry, so I wasted a lot more time attempting to divine something about it from Computer Shopper articles and so on. It is immensely rewarding to me on a personal level that I can serve as a resource for this community in whatever capacity I currently provide.

(same reasons I suppose that attracted me to becoming a professor while also working at TI, that was a dual-position, but every bit just as rewarding as interacting with the enthusiasts on this forum)

For me it's just a hobby hence my ignorance in a lot of areas. And thank you for the detailed answer.

I was once every bit as ignorant about this as you are now. Ignorance is a temporary state on the learning curve. And trust me when I say that you could fill an internet with the stuff I am ignorant about

Isn't AMD's and GloFo's relationship eroding as we speak though?

This has me quite concerned actually. There is a definite chill in their relationship and such things rarely turn into water under the bridge, this will get worse before it gets better, unfortunately, IMO.

If you sum up the total net profits generated by the entire CMOS foundry industry since its inception you will find that TSMC alone has made 100% of all the profits. (no surprise since they are the Intel of the foundry industry, but they also started it)

Other foundries make money on occasion, but their losses exceed their profits overall when you sum it up.

With this in mind, GloFo faces a very difficult challenge in legitimizing themselves as a viable contender to TSMC for leading-edge process tech. It was absolutely critical that they not stumble on 28nm, they had to nail the execution on that to have any hope of generating the kind of confidence that is necessary to convince fabless companies to take on the risk of an unproven foundry and move away from TSMC for 22/20nm and beyond.

If GloFo has stumbled as badly on 32nm/28nm as some are saying they have then this could very easily snowball/cascade into a fatal misstep for them. Time will tell, as is always the case, but the earliest indications are not the kind that inspire confidence and that is very troubling to anyone who has skin in the game or has friends whose livelihoods depend on GloFo's ability to execute. (for me, personally, the impact on the consumer market is a distant concern versus that of the concern I have for the job security of people I know who are in harms way)

 

[sb.0326]

Well they say experience is the best teacher and few have ever done anything flawlessly from the get go in any field. So let's hope they now know that "this" part of it or "that" part just isn't working as intended so they can fix it and move on. They certainly aren't gonna just sit on their hands and say "well this is the best we can do".

As has been said we definitely need a competitive AMD and GloFo is one component of that competitive AMD. Least at the moment it is, unless they just pull up stake entirely and try to move everything over to TSMC.

Kinda odd though that such an expensive undertaking has such thin margins. One would think that being the ones that actually make cpu's and gpu's would yield better profits. Especially considering the cost needed to build new facilities. Though I guess that could be why margins are so thin. Constantly having to rebuild and upgrade everything every few years.

 

[VirtualLarry]

Though I guess that could be why margins are so thin. Constantly having to rebuild and upgrade everything every few years.

Just like consumers have an "upgrade cycle" - so do the FAB plants that produce the chips that you and I use.

And the cost of that has to be carefully amortized, so that those companies can make a profit and stay in business, rather than go broke and out of business.

Hopefully demand for desktop CPUs remains strong. I feel that Bulldozer is in a precarious position. Those in the know avoid it (unless there are specific requirements that call for a BD, I can't really think of one).