AMD Downgrades Bulldozer Transistor Count By 800 Million

Robadob writes "It has come to light that AMD PR had originally reported that the new Bulldozer processor's transistor count was 2 billion. AMD PR are now asking reviewers to correct this count to 1.2 billion from the original amount they provided ~3 months ago."

Comments

It was a rounding error

[Anonymous Coward]

The FPU in these chips rounds 1.2 billion up to 2.0 billion.

Re:It was a rounding error

[FatdogHaiku]

We are Pentium of Borg. Division is futile, you will be approximated!

Serious policy changes here ?

[unity100]

With the new ceo ?

Normally the route of a true american corporate cultured corporation would be to deny everything and fool everyone and rip as much cash as it can. Until they were confronted at courts.

But now, amd marketing is rather needlessly contacting reviewers to make corrections, while taking a hit in p.r.

But is it really a hit ? Coupled with the fact that the new ceo kicked a lot of marketing staff, this tells me that the new term in amd is going to be a term reminiscent of early 90s in technology - a responsible era in which corporations have actually manufactured useful gadgets and sold them honestly, trying to get the edge on each other through tech - not with filthy dealings or deceit (hello intel and the bribery verdict)

Re:Serious policy changes here ?

[mug funky]

what about examples from outside the USA? BP stands for British Petroleum, don't you know?

some of the food scares going on in China are just scary - melamine in milk, cooking oil salvaged and re-refined from retail liquid waste... of course, people got the death penalty for the milk thing.

Re:Serious policy changes here ?

[fnj]

Er, well, actually, BP doesn't stand for British Petroleum any more. It officially stands only for BP. In fact, they have tried to make people think BP stands for Beyond Petroleum. OK, this sounds like an academic point, but actually it illustrates the very practical point that all these corps from whatever country are international and indistinguishable nowadays.

They just lost my business!

[Anonymous Coward]

I'm paying for *transistor count*, not GFLOPS!!!

Priorities, people!

Re:They just lost my business!

[vipw]

With Bulldozer, you're not going to be happy either way.

Re:LOL

[Anonymous Coward]

Isn't this better? Doesn't this mean that the processor is smaller and more efficient?

Transistors are HUGE!!! You ever see a transistor radio?

I guess I just don't understand hardware. It just doesn't make sense to me. How can these fucking weird-ass CYLINDERS that are painted in stripes, plus weird flat "wires" on a green piece of whatever-the-fuck, actually do something? It makes NO SENSE.

Why?

[RobinEggs]

I understand the importance of truth in advertising, but is this information meaningful, or just an insignificant correction? The magnitude of the difference alone doesn't automatically make this an important story, or the exposure of some big, inexcusable lie by AMD.

What's the true relevance of transistor count? If I see two processors with identical performance and power efficiency but radically different transistor counts do I have any real world incentive to select one over the other? I mean, presumably the one with fewer transistors in roughly the same die space might overclock better, might have a longer MTBF, etc., but beyond that should I care?

Or did timothy post this just to keep up the fanboi flame wars?

Re:Why?

[fuzzyfuzzyfungus]

I suspect that transistor count means different things to different people.

The PR numbers provided for tech journalist previews and fan-wank benchmarks are pretty much just noise: If the number is big, you'll see a couple of sentences about "zOMG 2 Billion! motherfucker is a BEAST!". If the number is small, you'll see a couple of lines about how 'the foocorp design team was heavily focused on optimization for this generation'. The only thing the end customer will care about are the benchmarks at the end.

For people attempting to glean financially useful clues about a company's process strength or design prowess, or ability to hit some thermal target in the upcoming product cycle, transistor counts are likely much more relevant; but are also rather less likely to depend on PR numbers(actually reverse engineering a modern x86 chip would be Serious Business; but just paying somebody to crack the top off, get some die shots, and provide good ballpark numbers on transistor numbers and allocation between cache and various functional blocks should be relatively cheap compared to some of the moves you might make on the basis of such information...)

It seems bafflingly weird that PR would provide a number so grossly wrong, since the fanboys and the haters basically make no real use of the number and the people who really care should be able to easily detect a lie of that magnitude; but I'd be somewhat surprised if the original PR numbers meant all that much.

Re:Why?

[rbmyers]

Bulldozer has apparently been a disappointment. Why? One wonders. One suspects that AMD encountered some late-in-the-game, unexpected, and very unpleasant surprises. Maybe 2 billion is closer to the number they wanted it to have. Sounds more and more like the NetBurst Story. Intel never did figure out how to add in enough transistors to make the design work well without breaking the power budget. Robert.

Re:Why?

[icebike]

Or maybe 2billion is what the thing actually has, but in order to get it to work they had to abandon a significant portion even if they still exist on the chip.
It wouldn't be the first time unused banks of memory were left on chip but simply abandoned.

Re:Why?

[msauve]

"It seems bafflingly weird that PR would provide a number so grossly wrong"

Not really.

Marketeer: How many transistors in the new chip?
Engineer: We're up over a billion now.
Marketeer: Ok, thanks. 2 billion.

Re:Why?

[mobby_6kl]

Or maybe 2B is what Bulldozer actually has, but their new PR team came up with the idea: claim it only has 1.2B transistors, so even though it still sucks, at least it doesn't need 2 billion transistors to do it.

Re:Why?

[RobinEggs]

Transistor count is closely tied to cache size. This CPU just went from "Extreme Edition" to "Celeron" to use Intel terminology.

Alright, but doesn't that response just transform my question about transistor count in the whole processor into exactly the same question about transistor count in the SRAM? If the cache size and performance of the whole unit are reported accurately, should real people care how many transistors there are?

Is there some kickass use case for a chip with a SuperPi score of X, a SPEC score of Y, a 6MB cache, 8 threads, 2.9 GHZ clock, and 2 billion transistors that totally falls apart on a processor with the first five traits but only 1.2 billion transistors?

Re:Why?

[poity]

PR department playing telephone
First person: "We have 1.2 billion transistors"
Next person: "Wow, that's about 1 and a quarter billion transistors!"
Next person: "Wait, 1 and quarter billion transistors? That's almost 1.5 billion!"
Next person: "Holy shit, 1.5 billion transistors? That's nearly 2 billion!"

Re:Why?

[marcosdumay]

"I suspect that transistor count means different things to different people."

No, it means nearly nothing to anybody. The closest one I've seen is another answer to this thread linking it to cache sizes, but even then, people measure caches on bytes, not transistors.

Buyers want software performance, measured by benchmarks, cache size + instruction throughput, or any other functional metric. Engineers care more about hight level units, except where they optimize deeper, fabs care about die area. Nobody cares about how many times a poly line crosses over a crystaline line.

By the way, that is probably the reason such a huge mistake in the number could be made. Nobody cared.

Re:Why?

[realityimpaired]

What's the true relevance of transistor count? If I see two processors with identical performance and power efficiency but radically different transistor counts do I have any real world incentive to select one over the other? I mean, presumably the one with fewer transistors in roughly the same die space might overclock better, might have a longer MTBF, etc., but beyond that should I care?

If you can find one processor with 2 billion transistors, and another with 1.2 billion, and they both draw exactly the same power requirements, performance, instruction set, and have the same heat envelope, then either somebody in marketing is lying, or somebody is doing something horribly wrong. On the one hand, it should not take 2 billion transistors to do something that can be done with 2/3 of that, but on the other hand, if you have 1/3 fewer transistors, you should experience an according decrease in heat leakage.

That being said, numbers like transistor count matter to enthusiasts. These are the same people who used to spend $1000 to buy the 3.8GHz P4 chip instead of the $300 3.4GHz chip. For some folks on the market, bigger == better, and there's no point in trying to explain to them that they can accomplish the same job with a less powerful chip. Some people build/buy/upgrade computers so that they can brag about being more powerful than their buddies, and when you couple that mentality with a disposable income, well, you know the rest.

For somebody like me, it's not likely to make a big difference. But I'm typing this on a laptop that's powered by a Celeron U3600 ULV chip... a dual core 1.2GHz processor that's designed for low power consumption, not high performance. It's been a long time since I have built/bought a high performance system, and I'm unlikely to get back into that game for a while: I gave up on computer gaming years ago. The one thing in this announcement that may give me pause next time I build a system is that a revision from 2 billion to 1.2 billion transistors probably means some kind of manufacturing problem that they thought they could overcome, but are now not thinking they can. Even if that's not the case, AMD is going to have egg on their face a while over this one.

Re:Why?

[Macman408]

It seems bafflingly weird that PR would provide a number so grossly wrong, since the fanboys and the haters basically make no real use of the number and the people who really care should be able to easily detect a lie of that magnitude; but I'd be somewhat surprised if the original PR numbers meant all that much.

IANALE (I am not a layout engineer), but it's my understanding that it is not an easy task to actually figure out how many transistors are contained within a modern chip. The CAD tools used aren't anything like Photoshop, where you can pop up an info window and see how many pixels it has. There are many different pieces within a chip - some might be standard library cells (like building a chip out of legos, rather than making a custom injection mold and filling it with plastic - where each building block is a few to a few thousand transistors (or more?). Other parts might be a full-custom layout, where somebody hand-placed every transistor to get the highest speed, lowest power, smallest area, or some combination of all of those. The chip might also include some hard macros, IP that is purchased from another company (like a memory controller or power manager) and just plopped onto the chip, with no insight to what is actually inside. There are hierarchies, and some parts (like cores or cache sub-blocks) are replicated a couple times, or a couple thousand times.

So it's my indication that any time you hear one of these numbers, it's really just an estimate anyway. Probably some engineer at AMD heard the 2B number after PR trumpeted it, thought it sounded a little high, and found a mistake in somebody's estimate.

Or, maybe more likely, marketing just made crap up without actually asking any of the engineers. That happens a lot too, and it pisses us engineers off to no end. At least when they do it after the product is made, PR has to fess up. When they do it before a product is finalized, it usually means engineering has to scramble and actually make it do whatever marketing promised.

Re:Why?

[AmiMoJo]

It's like broadband, you get "up to" 2 billion transistors, or more precisely any number of transistors between 0 and 2,000,000,000.

I offered to pay them "up to 2 billion Euros" for one but they declined.

Re:Why?

[AmiMoJo]

I wonder if they count non-functional ones too? A typical mid-range graphics chip is the same as a high end one except that some of the parts failed factory tests and were disabled. There are 1000 stream processors on the silicone, but only 800 of them actually work.

Re:Why?

[drinkypoo]

IANALE (I am not a layout engineer), but it's my understanding that it is not an easy task to actually figure out how many transistors are contained within a modern chip.

I find this beyond difficult to believe due to the way computer chips are designed today, which is to say, entirely on the computer. You can figure out how many gates there are from the netlist.

Re:Why?

[slew]

IANALE (I am not a layout engineer), but it's my understanding that it is not an easy task to actually figure out how many transistors are contained within a modern chip. The CAD tools used aren't anything like Photoshop, where you can pop up an info window and see how many pixels it has.

Actually, it is really easy to figure out how many transistors there are. Generally, you run both LVS (logic vs schematic) and DRC (design rule checking) tools on the final layout data base. These tools look at all the transistors in the layout data base and compares them to the original design (LVS) and to make sure that the active areas of the transistors are spaced out accordingly to make sure they can be fabricated into masks that yield (the base rules are provided by the silicon foundary often augmented by the library provider). After you run these tools, AFAIK it just tells you how many transistors it checked. You might tell the tool to skip some parts of the design (say like rams) in the final netlist, but generally the parts you skip have been run through the tools before hand or in parallel.

Of course running these tools takes a long time, and sometimes they are not finished running before the chip is fabricated for the first time. Sometimes, you send off the design or tape it out, and then you kick off running these tools so the chip starts to go to the mask-maker whist you are still running the LVS/DRC tools, but generally you know the results of LVS/DRC before you go into production (or your company probably isn't meeting your ISO 900x certification requirements). If nothing else, the fab will make you sign a yield waver if you don't run DRC (basically, they won't guarantee any working parts).

Re:Why?

[tlhIngan]

Transistor count means a lot to the future evolution of the product. If it's lower than the competitor, then that means that (on the same process technology) you can fit more onto a wafer and so they'll be cheaper. A low count means that you can easily fit extra cores on a die. The transistor count also implies the transistor count per core, so a lower number means that adding a couple of extra cores is less expensive that previously thought so it's likely to happen sooner. It may also mean that they're under the transistor budget and can add some extra execution units to the next version.

That would be a problem if it was transistor count that was the big issue on CPUs.

Chips are divided into two arenas - silicon area limited and pin limited. Silicon limited is basically the more area you have, the bigger the chip can be, and this applies mainly to memory chips. Double the area, double the storage (the addressing logic is such a small fraction of the area it's actually negligible). The problem is, the cost of additional silicon goes up exponentially as yields go down the larger the chip is (there's a certain amount of defects per area, and increasing the area per chip increases the chance of one defect ruining the entire chip).

Pin limited is where the limiting factor is the chip packaging - the number of pins on the device. CPUs are in this category - they just don't have enough pins to handle all the functions they need to do. SoCs have it worse as a higher pin/ball density automatically means a more expensive PCB process needs to be employed, which limits circuit size due to expense.

The big problem with pin-limited chips is their transitor density is quite low as it's filled with random logic. In fact, the reason why the density is low is because there's no order to the logic (unlike a memory cell array which is extremely regular). Thus what takes all the space on a pin-limited chip isn't transistors, but wires! The interconnections between transistors gets pretty tight that transistors have to be spaced out in order for them to fit. It's why we have 11+ metal layers per chip these days. And these wires also mean that sometimes, we have to add delays just to let signals propagate across the chip.

Transistor counts really haven't mattered too much - we can easily fit 8/16+ billion transistors on a memory chip, so 1.2 billion isn't anything. In fact, most of that is probably cache that occupies maybe 10-20% of the entire die.

Slightly smaller disaster

[Guppy]

I guess the new figures make a little more sense. Bulldozer's performance was fairly similar to their previous (and smaller) Thuban Core, at 904 million transistors -- it was as if AMD decided to take more than half of their transistor design budget, heap it in a corner, and set it on fire.

Not helping their cause much....

[CajunArson]

So a few points about this rather bizarre announcement:

1. Unfortunately for AMD this does nothing to reduce the power consumption of Bulldozer which is higher than a 3960x at stock speeds. When you remember that over 1/3 of the transistors on the CPU (using the new 1.2 Billion transistor count) are in the L3 cache that only runs at 2.2 Ghz while the L3 on the 3960x runs at full-speed, you have to wonder at whether GloFo's 32 nm process has some fundamental flaws, or if AMD didn't listen to GloFo's design rules (or some of both).

2. AMD's and GloFo's combined marketing of their "gate-first" 32 nm process bragged loudly and repeatedly that gate-first (as opposed to gate-last used by Intel) gave 20%+ transistor density benefits and that Intel's process wasn't truly 32 nm. Well, when Bulldozer was reported to have a die area of 315 mm^2 and a 2 billion transistor count, this seemed like a justified advantage. Now, however, the transistor density of Bulldozer is lower than any other 32nm design from either AMD or Intel. Note: the same AMD PR guys that adjusted the transistor count confirmed that the 315 mm^2 die size is still accurate.

Rory Read is smart to shift the focus away from these unmanufacturable monsters and to put it on the next-generation of Bobcat and Trinity designs where AMD can actually leverage it's only real advantage over Intel: the GPU.

Re:Not helping their cause much....

[nadaou]

aka version 1.0 of the new design leaves much room for refinement but they couldn't wait any longer to ship it. News at 11.

Re:LOL

[Anonymous Coward]

Water, fire, air and dirt
Fucking magnets, how do they work?
And I don’t wanna talk to a scientist
Y’all motherfuckers lying, and getting me pissed.

well, theres only one thing for it.

[smash]

Who's gonna start counting? Methinks its a PR excercise for all the shit amd are copping not being able to best a sandy bridge quad core with 1.5x as many transistors and (according to AMD measurements) 2x as many cores.

Class action lawsuit please

[assemblerex]

If you bought a V10 car and it turned out to have a 4 cylinder, you'd be upset. No?

Re:Class action lawsuit please

[RobinEggs]

If you bought a V10 car and it turned out to have a 4 cylinder, you'd be upset. No?

Yeah, if it turned out it couldn't climb hills and had a 0 - 60 time of 17.1 seconds. If it performed like I wanted and happened to have only 4 cylinders I wouldn't care. Unless one of my primary 'needs' was for everyone to know I had a big-ass 'engine', if you know what I mean.

Put more directly, benchmarks and statistics are just dick measuring without some context.

Re:Class action lawsuit please

[SpazmodeusG]

They told you the V10 engine used 50 hours of labour in its manufacture but it turned out the V10 engine only used 25 hours of labour in its manufacture.

It's still the same engine in every way. Transistor count is simply a manufacturing detail.

Re:Class action lawsuit please

[O('_')O_Bush]

No it wouldn't. If you were comparing a 500hp 2l engine to a 500hp 7l engine, there would probably be a reliability correlation. Remember, # of cylinders != displacement. More cylinders means better throttle response, in general, but LESS reliability because of more moving parts. More displacement generally means more power for less stress. Take a look at some of the 3-4 litre ferrari V12s as an example of a small, high cylinder, high strung engine with the same power outputs as some of the much lower strung 7l V8s of muscle cars like the Ford GT.

Re:Class action lawsuit please

[zippthorne]

Have you ever bought a digital camera?

Re:Class action lawsuit please

[cynyr]

As long as it enables time travel and mows lawns i don't care wtf it is.

There's a new update

[Anonymous Coward]

AMD just clarified that Bulldozer does have 2 billion transistors after all, but only 1.2 billion work. Which explains something about its performance.

Re:There's a new update

[rust627]

and the other 800,000,000 are looking for employment elsewhere ?

Re:There's a new update

[ShieldW0lf]

and the other 800,000,000 are looking for employment elsewhere ?

No, they're just Occupying space.

Cache?

[gman003]

I recall seeing that the top Bulldozer only had 8MB L3 cache, which seemed a bit low - Intel's equivalent top-of-the-line desktop models reach 15MB, and the server models 30MB.

At first, I just figured they were targeting the middle price bracket, but then they priced against the high-end. So I would not be surprised if much of the missing (or disabled, if that rumor turns out to be true) transistors belong to the cache.

Re:LOL

[Anonymous Coward]

Whoosh!

Transitors Server vs Desktop

[Sollord]

The 16core server parts were listed as having 2.4billion transistors at launch so either the FX PR was wrong/confused and it really is 1.2Billion transistors as it they say it is now for an 8Core FX or AMD manged to bolts on an entire second 8core processor to the server parts with 400million more transistors.

Re:LOL

[hairyfeet]

From the looks of it Bulldozer is another Phenom I, where they have to use a generation getting the bugs out. Phenom II was and is an excellent chip BTW, sure its not gonna slaughter the latest and greatest from Intel but the dirty little secret AMD and Intel don't want to talk about is that for 95%+ of the users out there PCs have been "good enough" for quite some time. hell I'm the kind of guy that was building himself a new PC practically every year in the past, now my AMD Deneb quad is going on 3 years old and if I get that Thuban upgrade i plan to for my BDay i could easily see it lasting another 5 years, maybe more.

The smart move which I applaud the new AMD CEO for doing is cutting down on the desktop product to crank the living hell out of mobile because that is where the money is at right now. you look at Brazos and they have been selling out of those chips as fast as they can crank them and as an owner of a EEE E-350 netbook I can see why, 6 hours of battery life running full Win 7 HP X64, takes 8Gb of RAM easily, does full HD video without a stutter, low heat, and its a great little multitasker chip, running head and shoulder better than Atom and often beating ION at a lower price. The OEMs have taken notice it seems as i have seen Brazos in netbooks, all in ones, and HTPCs all over the place. Walking into my local Walmart the other day, a place that just a few years ago was strictly Intel land, more than 3/4ths of the laptop/netbooks and virtually all the desktops were AMD Fusion. I asked one of the guys I knew there about them and he said 'These things are selling like crazy, great for video and FB' which is of course where most folks are nowadays.

Frankly I think the path AMD is on is the smart one right now but its gonna have teething problems. They are in the process of switching their GPUs from VLIW to vector and looking at the chip bulldozer arch is really made for the new vector GPUs that simply aren't finished. by switching to vector you'll have a super FP that the CPU can hand off heavy math to when not in use for gaming while having a smaller FP on the CPU thus allowing more cores per chip. Like the switch to Stars its gonna take some teething pains to get everything switched over to the new designs and having GloFlo drop the ball certainly didn't help.

Personally I can easily see a day where Intel owns the top end and that is pretty much it, as AMD cranks out the chips for the low and midrange. Frankly the dual core Brazos is more powerful than 90% of the jobs my customers could come up with and I can see a quad version pretty much owning the low and midrange sectors due to the combo of price/performance and lower power. Just give them time folks, they are still cranking out Denebs and Thubans so I'd stick with those for now and let the new BD arch get the bugs worked out and by the time the chip after Piledriver comes out the boards and laptops will be cheap and plentiful. Until then just stick with Thuban, Mobile Phenom II and Brazos.

Re:Billion? Billion?

[GauteL]

Look. You need to check your facts before you appear even more stupid than you already look. The 486 processors from Intel had more than 1 million transistors when introduced in 1989. Do you REALLY think we've only increased the transistor count by 20% in 22 years? Moore's law would suggest that the number of transistors should increase by a factor of 2^22 ~ 4 million during this time, giving us 4 billion transistors. We seem a little short of this, but processors have definitely reached the 1 billion mark.

Looking at it in terms of density, the 486 debuted with a 1 micrometer process (1000 nm). Bulldozer uses a 32 nanometer process. Roughly speaking this should allow 1000^2 / 32^2 ~ 1 thousand times more transistors. No matter how you look at it, 1 billion+ transistors is about right.

You are looking at this in the completely wrong way. You are assuming (wrongly) that current processors have 1 million transistors and trying to work backwards using that false assumption. If you are still in doubt, I would suggest calculating how large a real CPU with 2 million transistors at 32nm process would be. Hint: you would perhaps be able to see it with a magnifying glass, so lining up 31.6x31.6 of those on one 1cm^2 die seems the right ballpark to me.

What's a transistor?

[scharkalvin]

Anybody that has ever looked at the schematic for a VLSI chip at the schematic level will have problems figuring out what the transistors are for because so many of them are actually being used as resistors, diodes, or capacitors. Many are bias regulators or interstage coupling voltage level translators. Transistors are the simplest things to put on an IC so there tends to be lots of them. The transistor count rarely translates into a true level of complexity for the device over all. Having said that the last time a transistor count on a microprocessor meant anything was with Motorola's first two major processors. The MC6800 actually had about 6,800 transistors. The MC68000 had about ... wait for it .... 68,000 of them!