AMD Predicts End of 32-bit Processors

DDumitru writes "Infoworld reports that AMD predicts it will stop producing 32-bit processors by the end of 2005. By depending on price cuts for Athlon-64 and Opteron, AMD is predicting that it's sales of 32-bit CPUs will fall off and obsolete 32-bit systems in less than 3 years. This is either a push forward, or a tactic to try to capture the 64/32 bit standard leaving Intel in the rear. Or it could just be hype." I'm not in a hurry to ditch any of my 32-bit machines, so long as I get them replaced by 2038.

Comments

You know.....

[inteller]

.....forgive me for being captain obvious, but my old A7M266 board runs just fine with XP 2000+....they can continue dropping the price on these suckers so that eventually I can max it out with 2600+s and also plop two MP 2600+s in my A7M266-D.....I don't even use half the capacity now...I'll be blown away when I plop those in for $50 each in a year or so.....keep predicting the demise AMD, it's all fine with me.

Re:You know.....

[addaon]

You have to realize there's a certain minimum cost to retail processors, the same way there is to, say, video cards. You can't sell an AGP video card for $30 MSRP. The cost of packaging, shelf space, support, etc. outweighs the cost of hardware by so much at that level that, even if the hardware were essentially free, the product price would be around $30 or so. Ditto for processors; once a processor goes below about $40, it seems like it's no longer worth producing; the cost of packaging and so on outweighs any possible further reduction in price. On the other end, at $5 or so they become worth producing again (see PICs and such) because, at that point, no packaging or support is expected... but don't expect to ever find a nice $40 CPU to put in your motherboard.

Re:You know.....

[randyest]

Why is that? Generally smaller geometries result in lower yields. It's worse when the process is first rolled out, and yield does improve as a process matures (which 90nm has by no means done yet), but in general, a larger geometry process results in higher yields.

If you care, this is because defect densities in the silicon remain relatively constant, and although die sizes may be reduced somewhat with smaller geometries (not as much as you would think, though, due to wiring density not scaling linearly with gate size), the odds of a defect being fatal (i.e., falling into one of the increasingly dense "wrong spots" on the silicon) increase exponentially with gate size decrease.

Re:You know.....

[ivan256]

I run my A7M266-D, with 2 Athlon MP 1800+ processors, plugged into a Kill-A-Watt meter. All signs point to it being cheaper to buy a dual opteron over the not so long term simply for the power saving features at idle. I figure that the new machine would pay for itself in less than a year. I'll be buying as soon as there are new opterons readily available so the price on the old ones will be low.

Perhaps you should try being captian not-so-obvious and look and see how much it costs to leave that dual Athlon MP, which uses about the same amount of power at idle as at full utilization, on all the time. Also, if you don't need the capacity, why did you buy it in the first place?

Re:You know.....

[tomhudson]

poster wrote:

You can't sell an AGP video card for $30 MSRP

Why not? GeForce 64meg w. tv-out are retailing for about that now.

Ditto for processors; once a processor goes below about $40, it seems like it's no longer worth producing;

You can get a motherboard w. duron 2000 cpu for $100 CAN (about $75.US). Split the cost in 2 - half for mb, half for cpu, we are under the $40.00 price point already.

Same w. CD-ROMS at $20.00 new, etc.

Re:You know.....

[Eccles]

Why is that? Generally smaller geometries result in lower yields.

Yes, but the greater number of chips/wafer almost always more than compensates. Remember the increase in the chips/wafer is the square of the inverse of the decrease. I.e., if you go from .15 nm to .1 nm, the chips/wafer increases by (.15/.1)^2 = 2.25 times as many chips.

Re:You know.....

[shepd]

>(turns out that msi doesn't have a bios that works with the Barton 2500+ at 333, you have to up the voltage (voids warranty) and use kingston memory - but kingston says they no longer support this (brand-new) motherboard, because they've had too many problems with it)

That's odd. I've sold a lot of those boards using PC2700 RAM without any problems at all. Perhaps kingston is supplying garbage memory. I don't know. I do this without making a single change to the BIOS, except to set the CPU to the proper speed.

You must use an AGP card for video to get this to work if you are talking about PC3200 RAM. AOpen, using the same chipset, warns that you cannot use such highspeed memory and expect stability with the onboard video.

It's all pointless anyways as the speed increases are infinitessimal compared to the heartache of an unstable system.

Ask anyone selling the low end stuff their return rates and compare it to the high end parts. There is a difference. I've seen it as a retailer for both end of the spectrum.

(BTW: Some PC Chips boards that don't include hacked and pirated BIOSes have been known to have bad IDE controllers that randomly corrupt data, due to their false advertising. I've been bitten by this a few years ago. Never again.)

PC Chips mode of business:

- Fake Cache
- BIOS Hacking (as above)
- "Customizable" chipsets (translate to: We use our inhouse garbage chipset with broken IDE support and remark it to whatever you like)
- Fake CPU speeds (*STILL!* K7SOM/K7SEM users beware)
- Paper thin PCBs (my experience)
- Website served from 56k modem (or so it seems), written by ESL students.
- No support after you buy it (If you get more than 1 working BIOS update you are so very lucky)
- Above is likely due to them PIRATING BIOSes from other boards (own experience, not unusual)
- Deceiving naming of products (That's a 666 Mhz VIA board)

There's also unverified rumours that the owners of PC Chips were into CPU remarking.

Think about that next time you buy PC Chips (or, *shudder* ASRock or ECS [same company, PC Chips has a need to change their name often, I wonder why!])

Oh, and before you think I'm just doing this for my business, I try to add more margin in for the junk parts as I know I'll just need to deal with returns right away.

bah

[fjordboy]

you young-ins and your 32 bit processors...I'm using 16 bit and I have no plans on upgrading now. You'll be back...

Good news/bad news for AMD

[burgburgburg]

Good news: I agree in the quickly coming obsolescence of 32-bit chips.

Bad news: I'm getting a dual 1.8Ghz PowerMac G5, baby! Yeah!

I have to go lay down now.

Massive gains in cooling tech?

[PierceLabs]

Unless AMD or someone else has a massive gain with respect to being able to cool these monster CPUs along with shrinking the die so that they are suitable in general consumer electronics, I don't think anyone is going to stop producing smaller/cooler CPUs. While the 64 bit chips are great and all, I just can't imagine seeing one in a phone (for example) or even a PDA in the timeframe that they suggest.

Maybe they meant to include that they won't go into these markets which limits their desire to produce low footprint, low heat chips.

Re:Massive gains in cooling tech?

[PierceLabs]

If you're going through all this trouble to down clock a processor, wouldn't it make more sense to just use something that runs at a lower clock? Its like putting a Farrari engine in a car that is only ever intended to move at 35 MPH - sure you could do it, but wouldn't it make sense to just get an engine more suited to the task.

adaptability

[DenOfEarth]

As a user of open source, I think this shouldn't really be a problem at all, should it? I mean, once gcc can compile 64-bit code, than we should simply be able to recompile all of our current apps for these new processors, shouldn't we? I'd be happy if someone out there could point me out as not being in the know...

Re:adaptability

[RevMike]

As a user of open source, I think this shouldn't really be a problem at all, should it? I mean, once gcc can compile 64-bit code, than we should simply be able to recompile all of our current apps for these new processors, shouldn't we? I'd be happy if someone out there could point me out as not being in the know...

The Linux-on-Alpha project already did all the heavy lifting needed to run 64 bit almost ten years ago. Linux and *BSD is already running 64 bit on PowerPC. Virtually any package you can download that has an active support community is already 64 bit ready.

Re:adaptability

[Neon Spiral Injector]

That pretty much is true, along with the fact that GNU/Linux has been running on other 64-bit platforms for many years now.

The biggest problem I'm having building a custom install for my AMD64 machines is the fact they have 32-bit compatiblity. On the Alpha is was easy, the lib directories contained 64-bit libaries, because the machines were 64-bit, period. But with the AMD64 the lib dirs are still supposed to contain the 32-bit libs, with the 64-bit versions installed in a lib64 variant. This causes problems because almost no libary packages are setup to compile twice once in 32-bit and then over again in 64-bit mode with a simple ./configure.

Re:adaptability

[Anonymous Coward]

I am not sure what you mean :
1) Gcc 3.3 actually compiles to AMD64 (i.e. Athlon 64 and Opteron)

2) Gcc 3.3 generates pretty fast code too, as you can see on spec.org where IBM submitted results obtained with an Opteron, Suse Linux and GCC 3.3 : the baseline is above 1100 whereas Apple said that a Pentium 4 at 3.06GHz only achieves 880 with GCC 3.3. So it seems that the Opteron is a better processor for people who use GCC.

3) I write this text on my Athlon 64 running Mandrake 9.2 RC1 for AMD64. I can tell you that there is litlle left to adapt because it works pretty well. In fact you could hardly tell the system is the AMD64 version...

4) running 32 bit programs works fine : I have to use Java 1.4.2 for Linux x86 since the AMD64 version is not available. To do that, you just run the 32bit Linux program and it works at native speed.

As a conclusion, the Athlon 64 is good but it is still expensive (I paid 450 euros for my processor).

I hope the editors realize...

[AKAImBatman]

I'm not in a hurry to ditch any of my 32-bit machines, so long as I get them replaced by 2038.

I hope the editors realize that 32bit processors CAN process 64 bit numbers. In Java, for example, the date is handled by a 64bit number that tells the number of milliseconds since Jan 1, 1970. Amazingly enough, that clock won't run out for another few billion years.

Oh, and most Unixes have fixed the time problem. The real issue is getting the programs to switch over to the new APIs.

Re:I hope the editors realize...

[the_mutha]

No, the REAL issue is memory space. 32 bit just won't cut it anymore for large database servers and the like, regardless of the movement for clustering.

The thing is, in the desktop it will take longer for machines to require more than 4/8 GB of memory.

16-bit? 16-bit?

[burgburgburg]

You're soft. Why, in my day, we sent rockets to the moon with 8-bit processors, and we considered ourselves lucky to have them.

16-bit? Why don't you just go lay down on a feather bed and let servants peel your grapes for you? Harumph.

Re:16-bit? 16-bit?

[Waffle Iron]

Looks like the Apollo guidance computers were 15-bits. From a random Google hit:

Each computer had two types of memory, erasable and fixed. The fixed memory contained the programs, constants and landmark coordinates using 36,864 terms or words, each of 15 bits length. That came to a grand total of 74 kilobytes of memory. The erasable memory, which was used to store variable data used in calculations or as registers for logic operations, had only 2,048 15-bit terms.

I remember reading elsewhere that the 36 Kwords of ROM were hard coded by hand threading the bit patterns with tiny wires and magnetic cores, and then they were sealed in a block of epoxy. Turnaround time to fix bugs took weeks.

Re:16-bit? 16-bit?

[shut_up_man]

Listen you cocky young whippersnapper, in MY day we broke the Hun's codes with 1-bit difference engines the size of buses, running on vaccuum tubes infested with poisonous moths bigger than your hand. We entered our programs with sledgehammers, completely from memory, during air raids, in the dark, while we were completely drunk!

And you tell the kids of the day that, and they don't believe you...

hrmmmm I don't have to worry about bits.

[zaqattack911]

I'm betting by 2005 my 32bit cpu well become self-aware and upgrade itself.

I can see it

[steveha]

AMD has said that next year they will be shipping Opteron chips that only dissipate 30 Watts. An opteron runs 32-bit code faster than an Athlon, and totally owns if you run it in 64-bit mode. If you could buy a chip like that, for the same price as an Athlon, why would you want an Athlon?

(If there was a problem getting a good motherboard for the Opteron, that would be a good reason to still want Athlons, but there isn't a problem. There are plenty of good motherboards for Opteron and Athlon64 already.)

steveha

32-bit processors will still be around

[dtjohnson]

I think that AMD is just saying that *they* will stop making 32-bit processors by 2005 but not that 32-bit processors will be dead by then. It is reasonable for AMD to end production of the 32-bit processors because AMD does not have a lot of manufacturing capacity and they will want to make their 64-bit processors rather than 32-bit processors with the capacity that they do have. Also, AMD's 64-bit processors are also better 32-bit processors than many of the 32-bit processors that have been sold in the past.

The 32-bit processors will obviously be around for a long time yet but they just will not be made by AMD. Intel will keep making them and probably other companies such as VIA and that chinese 'red storm' company (can't recall the name of it) will make them for many years to come. The old 16-bit 286 processors are still made today, even though Intel stopped making them years ago.

IPv6

[RevMike]

Finally the more advanced processors will become commonplace, if it is not just a rumor. Even if it is a rumor, it is bound to happen eventually.

64 bit PCs will be here long before IPv6 makes significant inroads in replacing IPv4. (Ducks out of way of ensuing flame war)

Re:IPv6

[IWannaBeAnAC]

Because the advantages dimishish. the progression from 4 bits -> 8,16,32,64 is not exponential, but an exponential of an exponential. The more 'natural' progression would be word sizes increasing by some constant number of bits every generation, which would imply the range of arithmetic (and memory sizes) increasing exponentially. But it is technically easier to have word sizes that are themselves a power of 2. We have now reached the limit where 32 bits is barely enough to address the amount of memory desirable even for a small machine, but with 64 bits increases that tremendously. If it took 30 years to get to 64 bits, it should take another 30 years to get to 128.

Of course, this is completely independent of things like memory bandwith tricks (I/O with multiple words at a time) and vector processing (operating on multiple words at a time). But that is orthogonal to the underlying word size. '256-bit GPU' is just a marketing term, that has nothing to do with the amount of memory it can address or the number of bits of precision of the arithmetic, which is what is really meant by the CPU word size.

Yes, but,

[gillbates]

The real question is how long will it be before the BIOS is 64 bit protected mode?

Probably never.

I still write in 16 bit assembly because the BIOS still runs in 16 bit mode. It would be nice if AMD broke backward compatibility for once and started off with 64 bit firmware so I could at least write 64 bit assembly. The mixed-mode stuff (16 and 32 bit) that I would have to do for OS programming is getting ridiculous:

• I could write in 32 bit mode, if I wanted to write drivers for every single conceivable piece of hardware out there. While this would be ideal, it is far from possible. And since the processor starts in 16 bit mode, I'd still have to write at least a little real mode assembly.
• I could still use 32 bit mode if I wanted to use a call gate to call the 16 bit code of the BIOS, or:
• I can write 16 bit code, call the BIOS directly, and only have to worry about the 64k and 1M memory limitations.

I don't like any of these solutions, but it's a lot easier to fit kernel modules in 64k than it is to write call gates for BIOS services. The reason why I like using the BIOS is because it is standard across differing computers - I don't have to write a different driver for every single video card and hard disk controller that I might come across. Plus, if it uses the BIOS, I can be reasonably certain that it will run on an arbitrary PC; I don't have to do any hardware probing or detection.

Well, it's a pipe dream, I guess.

Those of us who like to program their own hardware took a serious hit when the 32 bit OS became the standard. We either ended up jumping through hoops to use the 16 bit BIOS from protected mode, or we just decided not to use more than 1 megabyte of the machine's memory. If they had installed 32 bit BIOS's when the 32 bit processors came out, we would never have had these problems.

But no, we still have a 16 bit BIOS because the manufacturers are afraid that some fool might want to run DOS on their 3GHz Pentium 4 with 1 GB of RAM....

Re:Yes, but,

[dmayle]

I don't know why I'm doing your work for you, but you should try educating yourself.
Check out BIOS32 services. It's a 32-bit entrance for BIOS services, and it's an industry standard...
PDF link

Re:32 bits alive and well.

[njdj]

Hell 16 bit processors are alive and well!

most industrial PC104 form factor PC's are running 386 processors.

The 386 is a 32-bit processor.
(There was a later variant called the 386SX that used a 16-bit bus, but it wasn't popular, and anyway the CPU was still 32-bit).

John Titor?

[Aero Leviathan]

Oh.. so the 2038 problem is nearly fixed? So this John Titor guy is a fake, right?

Whew!

384,000km?

[billstewart]

Kilometers? In my day, when Americans went to the moon, we only had *miles*, and only had 238,000 of them....

Red storm rising (Cray & AMD)

[Space cowboy]

According to Computer Shopper, AMD and Cray are collabarating on a new chip interconnect method, which they claim runs 20x the speed of current solutions, called 'Red Storm' ...

Just a 'news in brief' item, so no real details...

Simon

Re:Ohh yea?

[Loki_1929]

"I was sure that Intel had announced plans to integrate it's 64bit "Yamhill" extensions into the Prescott chip which should be launched early next year."

They haven't, and they wont, unless the Athlon64s begin to eat away at too much of Intel's market share. Intel's problem is that if it releases 64-bit extensions in Prescott, it will be forced to do so for the Xeon line as well. The problem with this is that it would send 10 years of research and development, along with countless Billions of dollars down the drain. Itanium relies on one thing and one thing only: a need for 64-bit processing. The few Itaniums that are selling will be the only ones sold if Intel's customers can get 64-bits on a Xeon.

Unless Intel is forced to, it's simply not going to do "64-bits on the cheap". Intel has nothing to gain by announcing or implementing 64-bit extensions now; we don't even know if the Athlon64s are going to sell. On the other hand, they have everything to lose if they do open up Prescott to 64-bit quickly. What Intel is probably doing now is scrambling to come up with ways to sell Itaniums to its customers once 64-bit Xeons are available. Itanium's miserable sales to date show that it has enough trouble just selling it as-is. Take away the one advantage it does have, and they may as well take it off the table now.

Aside from that, we have absolutely no idea how well Prescott's 64-bit extensions perform. We have no idea if they've been perfected, nor what kind of problems Intel has with them. We have no idea if bugs exist in the instructions, nor what sort of benefit we'd see from them. They could very well be based off of AMD's own 64-bit extensions (Intel has licensed the technology), which would make it rather embarassing to release them to the public. It would also constitute a huge and embarassing shift in position on the topic of 64-bits for home; something Intel has publicly stated is not something the market currently needs.

I'm not saying Intel wouldn't possibly announce 64-bit instructions in the near future - it could be tommorrow for all I know. All I'm saying is that everything points to Intel keeping the 64-bit extensions under wraps for as long as absolutely possible. As of right now, the only 64-bit instruction sets we know of (when it comes to Intel/AMD), are EPIC for Itanium and the ones from AMD. It's doubtful that Intel has secretly developed a whole brand new instruction set just for the P4, and it's even more doubtful that they somehow rigged the P4 to use Itanium's 64-bit instruction set. Thus, we're left with Intel banking on limited adoption of AMD's 64-bit CPUs, which does not appear to be the case thus far.

P.S. All signs point to Intel releasing the first handfull of Precotts around Feb of 2004, with volume closer to early Q2. Solving voltage leaks that push your operating temp to extremes isn't something you throw a bandaid on; it's something you design around. I just hope, for Intel's sake, that they aren't rushing Prescott out the door as they have on other chips (P3 1.13GHz).

Re:a word from the Processor Growers Association

[WoTG]

I think 64 bit will take root at the servers and quickly force 64 bit processors out to the workstations. Even now, 4 GB of RAM in a server is a constraint in many businesses, and it will be much more pronounced in only a couple years.

Given this situation, Microsoft is bound to polish and market 64 bit Windows Server (which will soon be available for x86-64) because they know that 64 bit Linux distros are out or will soon be out. MS's place in the server world is far from secure, so it's in their interest to be as competitive as possible in this regard.

So, with servers (and high-end workstations) moving to 64 bit really soon, and the fact that one of x86's strengths is in volume manufacturing, the natural step (especially for AMD) is to move all chips to 64 bit in a fairly short time frame. Besides, the incremental cost per CPU is minimal, AMD claims somewhere around 10% die space. And, it takes more money to design and manufacture separate 32bit and 64bit chips than it does to sell a single product line.

I guess I don't think that any "killer app" is really required above and beyond what is out there already. Big fat database, email, and application servers that could use more than 4 GB of memory NOW.

Re:Amazing

[Loki_1929]

"This may have been true but the average user just wants it to look good on paper."

(Bob) "Hey, Bill. I just got a new computer!"

(Bill) "I was thinking about getting a new one myself. Did you get one of those new Pentium 5 ones?"

(Bob) "Nope, one of the tech guys at work was telling me how those are only 32 bits, so I got one of these Athlon ones. They're 64 bits."

In the battle of bigger vs better, AMD has 64, Intel has 32. AMD introduced the model numbers for precisely the reason you raised - the average user has no idea what actually influences performance. The new AMD CPUs are set to ramp up in clock frequency very quickly, which will cause the model numbers to shoot up fast as well. What you'll end up with is Average Joe Consumer looking at Athlon64 4400+ & Pentium(?) 4Ghz. AMD is likely to pump up the "Their's is 32, our's is 64!" marketing, and regardless of whether the consumer has any clue whatsoever what that means, it makes AMD look better on paper.

Re:Amazing

[Zathrus]

This may have been true but the average user just wants it to look good on paper

Ok, so go ask Average User how fast the CPU in the HP Pavilion 3000+ is. Odds are they'll say 3.0 GHz, which isn't true but is proof of AMD's success in "looking good on paper".

The issue that AMD has long had is poor motherboards. Via had a long, long time with poor chipsets/drivers which lead to crashes (this pretty much ended with the KT133A, but it's popped up every now and then since). They also had issues with MS not including the drivers for the chipsets with the OS (which is a death knell, especially for something like a motherboard -- the boards worked without the drivers, but they were dog slow). They also had some thermal problems, which were wonderfully overhyped by the hypemasters at Tom's Hardware (no, I won't provide a link -- if you don't know it you're better off).

Nowadays those issues are in the past. Nvidia has been producing rock solid motherboards for over a year now. Via has finally worked out its issues as well. Via even has chipset support in XP (and Win2k/ME IIRC). Anyone who spouts heat issues is an idiot -- Intel chips now have higher power consumption and heat dissipation than AMD does at the same effective processor speed.

AMD's had issues breaking into corporate PCs though, and still does. Most PCs sold for corporate use are Intel only. They've also had problems breaking into the notebook arena, and they're making a slow go of it in both areas. AMD has long had the enthusiast market, particularly the value-oriented gamers, but it's by no means a lock, and it's really not a very large market.

Re:How can you say that?

[CaptKilljoy]

How can you say that with a strait(sic) face?

Because he actually knows what he's talking about. Here's a hint: the market for 8-bit processors is absolutely enormous even today. There's even a small but significant market for 4-bit processors. Do a little research.

(I can't believe the parent post is modded +3, Interesting.)

Workstation memory and "Moore's Law"

[linux11]

The doubling every 18 months seems to apply to workstation memory just as much as it applies to CPU speed. This shouldn't be a surprise since more applications tend to depend on memory for speed than raw CPU cycles. After all, if a section of code/data needs to be swapped from disk back into memory, all a faster processor can do is more NOOPs or context switch to a different process. So, while around 1980 a nice home computer would need about 64k to be "beefed up," now that about 15 cycles of 18 month periods have passed we are seeing beefed up workstations having around 2GB of memory or 64k*2^15. Next cycle is 4GB which maxes out the address space for 32 bits. We nearing the point where "power" users will start expecting workstations with over 4GB of memory and that definately calls for phasing out 32 bits on the desktop/laptop.

*cough* paged *cough*

[Ayanami Rei]

You're still limited to 4GB-(kernel prot.) per process.