Hardware Analysis - AMD�s 90nm process, a success story?

Re: AMD�s 90nm process, a success story?

Eildert Slim — 2004-12-10T00:37:57-05:00

Here's some better proof:
http://www.reed-electronics.com/semiconductor/article/CA456681...latestNews
http://www-03.ibm.com/technology/power/newsletter/august2004/f...k_look.pdf
The first one shows a picture of AMD's 90nm process, the second of IBM's. As you can see, there are a few distinct differences between the two.

Re: AMD�s 90nm process, a success story?

Brandon Gates — 2004-12-03T17:59:03-05:00

How about AMD's latest roadmaps to back things up? How about some extremely low temps of a system I'm running? And yes, there will be barriers as the process does shrink; this I am aware of. But in the article, the APPLE processor's issues are cited to assume the AMD's will have exactly the same problem. Unless AMD changes their roadmaps, then I can't say they do or do not have problems. I'm mainly trying to point out that the new cores are to be implemented 100% into the lineup, and so far I don't see any problems.

Re: AMD�s 90nm process, a success story?

Sander Sassen — 2004-12-02T20:54:51-05:00

And you back that all up with a press release that was issued in Januari of 2003? C'mon, think about it for a minute, we're almost two years further now, that press release is pretty outdated. I suggest you contact some of the engineers or people working in R&D at the fab in Dresden to get an accurate perspective of what's actually happening with AMD and IBM's 90nm process. You can rest assured that neither has any time to or desire to re-invent the wheel and there's a huge overlap in technology used, that's one of the benefits of the joint venture. As mentioned the PowerPC processors manufactured by IBM and the Athlons by AMD differ in architecture, but not so much in process technology used to manufacture them. Sure, there are differences, but they are similar enough to validate the claims I make, if you have information or sources indicating otherwise feel free to post them here, with references.

Re: AMD�s 90nm process, a success story?

Eildert Slim — 2004-12-02T20:40:05-05:00

Collaborated, yes. Worked together on, yes.

But they certainly do not use the same process, as the article assumes. The pressrelease clearly states AMD had already almost finished 90nm development by the time they started working with IBM. Unless you claim AMD threw their own development completely overboard, of course.
I think AMD did some cherry picking from IBM's 90nm technology for their own. There was a $30M something deal involved with that, if I recall. I would assume they picked the best parts - not the troublesome ones.

Re: AMD�s 90nm process, a success story?

Sander Sassen — 2004-12-02T19:20:33-05:00

Oh, rest assured it is not an error, nor needs correcting, I have recently discussed this subject with Dirk Depta at AMD's Dresden manufacturing facility in Germany and he acknowledged that both companies have collaborated on different levels and have co-developed on 90nm as well, part of that collaboration included the introduction of 300mm wafers and SOI on 200mm.

Re: AMD�s 90nm process, a success story?

Eildert Slim — 2004-12-02T19:12:47-05:00

Furthermore, I notice you still haven't corrected this error: "What we�re looking at here is that the 90nm SOI process developed by AMD and IBM [...]"
There is no such 90nm process developed by both companies - that's not until the 65nm node. From the AMD pressrelease I linked to before: "The agreement includes collaboration on 65 and 45nm (nanometer; a billionth of a meter) technologies to be implemented on 300mm (millimeter) silicon wafers."

Yes, IBM has helped AMD with 90nm technology, but it is still based on Motorola's.

Re: AMD�s 90nm process, a success story?

Eildert Slim — 2004-12-02T15:25:19-05:00

I recall THG having very similar problems with PIII Katmai at 600MHz and Coppermine at 1GHz both requiring much better cooling than their slower brothers. That is what happens if you push a core to speeds it can't really handle anymore - it doesn't seem to be 90nm specific.

But at 2.8GHz Prescott already produces more heat compared to Northwood. As frequency goes up, the difference actually seems to decrease! Which means the problems IBM has with the 970FX are entirely different from those of Intel with Prescott.
Actually, HardOCP mentioned how raising the voltage on a 90nm A64 didn't really seem to help overclocking. Which contradicts with your 970FX graph. And at 2.2GHz Winchester uses less power than Newcastle. Which means it's can't have the same problem as Prescott.
So, in conclusion:
1) Intel has a technical problem with Prescott.
2) IBM has a different technical problem with the 970FX.
3) AMD may or may not have a technical problem with raising Winchester speeds, but certainly not the same one as either of the first two companies.
And so far there is very little reason to assume AMD has a technical problem with 90nm at all.

Re: AMD�s 90nm process, a success story?

Sander Sassen — 2004-12-02T14:27:58-05:00

You miss the point, the problem with 90nm is that is has a breaking point, which wasn't there with previous die-shrinks. Both Intel and IBM/AMD acknowledge that it is there and when you reach that point power leakage occurs which significantly increases power drain and heat production with very little return in clockspeed. Traditionally a die-shrink meant a smaller die and less power drain at the same clockspeed, but the potential to scale upwards in clockspeed. With 90nm this only works up till a point, beyond that there's diminishing returns that increase power drain and heat production that cause for high clockspeeds to work but not be a viable solution due to the excess power leakage.

Re: AMD�s 90nm process, a success story?

Eildert Slim — 2004-12-02T13:59:50-05:00

IBM and AMD don't use the same 90nm technology: http://www.amd.com/us-en/Corporate/VirtualPressRoom/0,,51_104_...96,00.html
"The agreement includes collaboration on 65 and 45nm (nanometer; a billionth of a meter) technologies to be implemented on 300mm (millimeter) silicon wafers."
AMD has gotten help from IBM for 90nm, but they had done most development on that already by themselves.
So by comparing the 970FX with Athlon64 you're talking about two different designs, using different processes. Doesn't make the comparisson - or the entire article - very meaningful.

Apart from that, AMD is obviously a little late with the transistion to 90nm. And it takes time to convert an entire fab - which has to remain in production - to a new process. For AMD the main benefit of 90nm is lower production costs, at the moment they don't need extra clockspeed or cache to compete with Intel. Also, their first batch of 90nm chips is 512kB L2 only. AMD's first aim with 90nm seems to be making S939 higher volume and mainstream, as that benefits them most.

Traditionally, Intel has had 3 advantages from any new process:
1) Lower production costs
2) Less heat output at the same clockspeed.
3) Higher performance per clock, usually from adding more cache.
This has been true since 250nm -> 180nm. And 1) and 2) have been true for an even longer time.
Until Prescott, that is. Fortunately for Intel, reason 1) alone is enough to justify moving their entire production to 90nm. But something clearly �s wrong.

Btw, can you think of �ny example where raising voltage in order to increase frequency d�dn't increase heat output a lot more? What matters is at which frequency the design starts needing significantly higher voltage. And that is very much a design issue - pipelining, anyone?

Re: AMD�s 90nm process, a success story?

Jeff G — 2004-12-01T23:48:43-05:00

Are you SEROUSLY suggesting that AMD is incapable of making faster processors? You do know that neither AMD nor Intel immediately ship the fastest processor they're capable of making don't you? Why should they if the competition isn't making anything faster? Why not milk all you can from the 130nm 2.6 GHz processors that you can while people will still by them since there's no cooler running 90nm 2.6 GHz processors AVAILABLE TO BUY.
Xbit Labs recently did a review of AMD's new 90nm processors and showed that they overclock fairly effortlessly to 2.6 GHz with only an 8% increase in voltage. At 2.4 GHz they're only generating about 50 watts of heat. I doubt you'll EVER see a single core 90nm Athlon-64 generating 100 watts of heat.
Are you aware that Intel's problem with the Prescott is NOT solely due to the 90nm manufacturing process? They made other architectural changes to the processor that increased it's power requirements. Also don't forget that Intel is not using SOI, a technology that reduces current leakage, and AMD is.
You think AMD is going to run into the same heat problems as Intel? You're forgetting one HUGE difference...

The Prescott, when compared to a Northwood at the exact same clock speed runs significantly hotter.

The Winchester, when compared to a Newcastle at the exact same clock speed runs slightly cooler... even with a 40% smaller die through which heat is dissipated.

So... while AMD may still be perfecting the 90nm process and working out the kinks, it has nothing to do with heat. They may indeed run into a clockspeed wall because of the physical limits of the silicon transistors as they did with the Athlon XP... but that's an entirely different matter... NOT the same problems Intel is having with the Prescott.

Re: AMD�s 90nm process, a success story?

The Real Robert Jones — 2004-12-01T23:18:09-05:00

Hahaha, all this and 60 nm processors come out mid-2005. Oh joy! I can't wait! (please don't flame me on what I say, they are indeed coming out in mid-2005, I have sources, and I'm not a little 10 year old who posts stupid stuff )

Re: AMD�s 90nm process, a success story?

Govan Greaves — 2004-12-01T22:53:55-05:00

Well given the performance stats we are seeing yes it is a success and considering that amd is planning to move to dual core cpus by mid next year there isnt any real overwhelming need to bump the cpu speed up by much more. They have already stated that the dual core machines will have cores running at a slower speed to reduce the overall heat of the chip. The fact that they are 2 cores instead of one is what will provide the extra bump in speed to continue to maintain its lead over intel because intel is somewhat behind with their dual core stuff.

So there yah have it. Why go to 3ghz and strain the process any further than really needed?

Re: AMD�s 90nm process, a success story?

Everett Williams — 2004-12-01T19:00:40-05:00

Your logic on marketing is essentially correct, if a bit incomplete. I believe that the 130 to 90 transition is also the final changeover from 200 to 300 mm in wafer processing. That is a huge incentive to make the changeover, because the cost to process each chip on the larger wafer is less. Now, I will be the last person to tell you that I can resolve the Gordian knot represented by supply, demand, and capacity, but I suspect that if AMD holds this speed advantage for a year or more, their market share will steadily if not necessarily spectacularly increase. If they can maintain that very visible market advantage while markedly reducing their production costs with both the 90 nm feature size and the 300 mm wafers, then they can again reduce price while maintaining profits....increasing market share...and the wheel rolls on.

Now, I would say that the speed/heat/voltage barrier at 90 is there, regardless of what is done with the technology, so rather than preparing folks for the lower speeds, it is a natural result of the technology that will force the companies to the multi-core chips to maintain the price performance curve. As a certain author used to say, TANSTAAFL.

Re: AMD�s 90nm process, a success story?

Thomas Perry — 2004-12-01T17:40:32-05:00

AMD is a business. Maximizing profit at a certain level of productivity sometimes forces less than ideal circumstances for the consumers. AMD needs only to stay a speed rating or two ahead of Intel to win the race. Jumping any higher will mean that AMD has taken their product to EOL well before Intel has a processor to fight back. Then when Intel matches speed in 6 months, AMD has no more headroom and must design a new product. You would want to stretch out the life of a product as long as you can staying just 1 step ahead. That gives you more time for economies of scale to kick in and more time to develop a new product for replacement at EOL.

Remember the old sayings that if AMD and Cyrix has not existed, Intel would still have us using Pentium 166 processors?

My guess is that AMD is not going to more chips on 90nm because it is not needed. They maintain a rather good healthy product and a rising market share. AMD knows that it will never remove all of Intel's market share and no matter how much better their products are, Intel is built into the mentality of the consumers. Moving to 90nm may increase chips per wafer, but if you do not have a larger market share demanding the capacity, it is a waste.

As a consumer, the process changes do not add value to a sellable product. Your major manufacturers do not list the processor's design specs so it does not matter if the speed ratings match. Some list cache sizes, but rarely do you see processor design specs.

Since you titled this story as a question, I would go ahead and say that 90nm is needed for dual core processors. AMD is capable of reaching desired clock speeds with 130nm. The lower speed 90nm processors are characteristic of the side effect of dual core chips, which is that they run at a slower clock speed. Would it be beyond reasoning to say the slower 90nm chips are done to help prepare the engineering for a dual core set and to help verify yields?

Anyone's guess could be right or maybe a combination of guesses are correct.

Re: AMD�s 90nm process, a success story?

Everett Williams — 2004-12-01T17:37:16-05:00

Brandon,

What is the need here for calling someone a fan boy. Yes, AMD does get more per cycle out of their processors than does Intel, but I just happen to be very interested in keeping AMD and Intel both in the business. If either one controls completely, my bargains will go away, as will the current rate of technological change. The point of the discussion here is that both AMD and Intel are hitting a speed/heat/voltage barrier at 90 and below. The barrier may be slightly different for each process and company, but it is definitely there for all of them. Yes, the barrier will drift downward with process optimization and with innovation, but I believe that it is inherent to the current technology, which I don't see being entirely replaced anytime soon.

Given that, we have to find different ways to use the current technology and we have an obvious candidate with which the industry already has significant experience...parallel processing. Putting multiple cores on a single chip is not new either, but it is new to mass production. It reduces the cost of parallel processing by reducing many of the things that are required to support multiple, separate processors on the same motherboard.

For the moment, it looks as though AMD has an advantage until the multi-core chips hit the market. Since Intel already has it's hyper-threading model, they may have an implementation advantage and a small performance advantage based simply on HT in two cores. AMD has an advantage in that they have a very efficient and modular architecture in place for interchip communication and for memory access. AMD does not need massive amounts of cache to maintain performance, leaving more room for processing and inter-core communication elements. The race should be very interesting, and as long as neither competitor is eliminated, shifts in balance should only benefit us...the consumers. I'm just glad that I have a ticket to the fight.