Prescott: the answer to AMD's Hammer?

Introduction

In the last few months, computer related articles that didn't speak of Hammer, the so called revolutionary new CPU from AMD, were rare. Boasting 64bit instructions and HyperTransport, the new chip will make the Pentium a thing of the past, so it's said... There could be no greater mistake. Intel has more tricks up it's sleeve than a pokerplayer in a classic western movie; it's just that they've never been good at creating hypes. We hereby invite you to read this editorial, and meet: Prescott.

The power of PR

Processors come and go fast, much faster than several years ago. Can you remember the Pentium 200MHz's time at the top? Or the Pentium II 450MHz? In the 'old days' there'd be a new CPU every six months. Nowadays, you are lucky if you brand new state-of-the art processor isn't considered crap two months later.

It's all caused by the competition which formed in 1994, when Intel decided other manufacturers couldn't longer copy their designs and started using the name 'Pentium', the sudden change came as a hard blow for many companies depending on Intel-clones. One of those companies was AMD. However, unlike many others, AMD decided not to watch passively, but to fight back. Their technicians explored the outer limits of the 486 core while others worked on a new design to compete with Intel's new brand of CPU's. The result was called K5, who, like its successor K6, never became well known. Everybody wanted a Pentium, you know, the one from TV, with the funny tune. Pentium crushed everything that stood in its way, and the word 'Pentium' at some point meant the same as 'computer'.

But now I've got one... Dutch commercial for Laser computers

While AMD gathered more and more knowledge, partly by experience and hiring other companies' designers, they started to plan a surprise attack. Ex-Alpha designer Dirk Meyer was hired by AMD for a very ambitious project: to launch the seventh generation of AMD's CPU's before Intel could release its counterpart. Not many believed the rumours to be true, but at the end of 1999 the impossible happened; Pentium III was slain by K7, while the Pentium 4 was miles away. To make things worse for Intel, AMD mastered the tricks of public relations. K7 was called Athlon and AMD quikly gathered support. Especially Tweakers were very happy with the Athlon's overclocking potential. Although that's only a very small group, many of these held positions in which they could advise others. The word spread quikly, Pentium finally had a real competitor.

In the two years that followed, Intel suffered from a lot of bad luck. The market for semiconductors crashed and the space that remained, for the once-so-glorious company, was taken by AMD. AMD beated Intel at the race for the 1GHz processor. The Pentium 4, who was to turn the tide, was troubled by a messy contract with Rambus. The Pentium 4 core had to endure a lot of negative publicity, mainly because it couldn't show its full potential. AMD however, was racing upward. The company opened a new factory, to keep up with the growing demand, while its CPU's appeared not to have any kind of trouble keeping up with Intel's. AMD's PR-section re-introduced the PR-rating-system to convince everybody of Athlon's power. AMD became more popular with every day that passed and nowadays people are still convinced AMD is winning the CPU-race.

Intel underrated?

So, while Intel had quite some tough years to look back at, AMD used to expand its share on the CPU-market for a while. The discussions about what happened if Intel didn't make these mistakes, still continue. What's done can't be changed, so we'll just forget about that. Back on topic: those who say that AMD won are right. However, the victory isn't more than a personal one. Intel suffered some big blows, mainly on the psychological side, yet the powerbalance has hardly been changed.

Currently, Intel's share of the market is as big as 81,9 percent. That figure has been larger, but not much. On top of that, Intel is getting profit from her sales, AMD is losing more money each year. If we look at figures from the first quarter of 2000, we can even see improvements on Intel's side. Their profits in the last year, tower above AMD's with a whopping 830 percent, while the average for the entire period rates 740 percent. By producing a wide variety of products, Intel has managed to keep its average turnover at 90% of the average from the 21st century. AMD has just an 80% average. We can conclude from these figures that Intel has quickly recovered from the year of disaster, 2001.

Both sides saw a cut in profits. Intel pumped more money than ever into research, while the prices for its products became smaller and smaller, thanks to the competition with AMD. They haven't seen a green figure for a full year, mainly because of that same competition. There's only one side in this conflict that is really profiting: the people who need some more MHz.

			AMD turnover		Profit		Intel turnover		Profit
	Q2 2002		600		-185		6300		446
	Q1 2002		902		-9		6800		936
	Q4 2001		951		-16		7000		504
	Q3 2001		766		-97		6500		655
	Q2 2001		985		17		7060		854
	Q1 2001		1189		125		6700		1100
	Q4 2000		1175		178		8700		2200
	Q3 2000		1206		263		8700		2900
	Q2 2000		1170		250		8300		3500
	Q1 2000		1092		189		8000		3100

Those who keep calling AMD will put an end to Intel's profits when they release the Hammer, just forget how big Intel is. Intel has roughly one hundred establishments, spread across the globe. At twelve of these, you can find the so-called Fab's: huge, completely automated halls, many times cleaner than the rooms where doctors do their surgery, containing equipment worth billions of dollars. Two more of these facilities are currently being built, while there are some other factories that are just used for testing purposes. On top of that, the company holds a load of factories used for assembling and many more offices and laboratories, all in all containing some 80,000 workers.

AMD has two factories, but only one of those is capable of producing the Claw- and SledgeHammer, therefore making it impossible for AMD to take a big share on the market, no matter how popular the new chips will be. This problem will eventually solve itself; when chips become smaller, as they do with every new generation, one factory can produce more of them. One facility is more vulnerable than twelve, though. One fire or flood can cripple AMD's productions of Hammers for months.

Some people say AMD has an advantage by being small and agile, while Intel is big and slow. They couldn't be more wrong. Intel Research has some of the biggest, best financed and most modern facilities in the world. Together with other companies and world-leading universities they work on everything from software to nano-technology. The CPU-developers thankfully use the technology that evolves from those years of research.

To prevent people from becoming lazy, there are two different teams. The team in Oregon is, in a friendly way, constantly competing with the one in Santa Clara. While the first is trying to improve the current core, the second is already working on the next generation. Every team tries to implement smarter solutions than the other does. That's not easy, as the teams can see exactly what the other team is doing. Apart from the ones in Oregon and Santa Clara, there are also teams in Dupont, Washington and Bangladore.

AMD's risky Hammer

"Intel may have more money, laboratories, people and technology, but Hammer is more innovative than anything Intel has." These pronunciations aren't unusual lately. Are they true, or is it a perfect example of AMD's marketing? In 1999 they said that the K8 core would be built from scratch, but reality seems to be different; the K8 is definitely an extended K7 design. The team, who also designed the K7, took a risky decision by integrating the memory controller with the processor and applying HyperTransport on-die. If you exclude these gadgets, then you'll see that there are less radical changes in the core.

The x86-64 story, what has a lot of attention, is a huge kick for every marketeer, but technically spoken it's nothing more then extending the registerbanks and busses inside the processor and making sure the new instructions are recognized. Hammer will be 20% to 30% faster than an Athlon at the same clockspeed. This is achieved by a lower latency of the on-die memory controller, which easily could be merged with the Thoroughbred or Barton. The front-end has two new steps for decreasing decoding, distributing and predicting time for executing multiple commands. The back-end, the area for die-hard calculations, hasn't been changed. On behalf of the radical changes on the new platform and the extra set of x86-64 instructions, AMD decided to introduce the Hammer as an 8th generation processor.

Since time and budget were AMD's limited resources, they put an exemplary achievement with the Hammer. It isn't a complete new processor, but a "re-fit" of the K7; refreshed and ready for the future, but still an Athlon on the inside. Using older technology in new products isn't a sin. All renowned companies, such as IBM, Alpha and Intel, follow similar strategies. Nowadays, it isn't a problem to give a processor brute strength. The challenge is to optimize the usage of that strength.

The K8 platform uses concepts which aren't seen on the desktop, and AMD uses it for its marketing of the Hammer. It isn't important that its an Athlon from the inside and you'll never hear any complaints about it if its gaining performance. It's still risky though; the Hammer is the next "big thing" and if it isn't going te be as expected, then you'll see some image sinking. Postponing the release of the ClawHammer until 2003 is a bad sign, and then we aren't even talking about ongoing releases from Intel yet. Also while AMD is waiting to release the ClawHammer, the current Athlons are increasing in speed. Barton, with 333MHz FSB and 512KB L2 cache, is a possible threat to get even with the Hammer.

The question is if the production of the Hammer wil take a favourable turn. Is it possible for the K8 to gain such high speeds, or is it necessary to tweak the design (just like they did with Thoroughbred) before you achieve success? Rumours say that the low yields are responsible for the current postponement, and the yields are an advantage to the Pentium 4. AMD could use the Hammer to get even with the Pentium 4 and compete with the successor of the Pentium 4.

Meet Prescott

We would be lying if we'd tell you that Prescott is a revolutionary design. The core is based on the P7 design that made its first appearance around the end of 2001. Named 'Willamette', it came in 1.4 and 1.5GHz. Since then some improvements have been made: the socket changed from 423 pins to a smaller version with 478 pins. The amount of L2 cache was doubled, the die-size is reduced to 0.13 micron and the FSB has increased from 400 to 533MHz, to produce the core we now know as Northwood. The first chip to boast this core was the 2.0Ghz Pentium IV, which was launched early this year. Now, some months further in time, there have been some tweaks in the steppings and the P4 3.06GHz is knocking on our door.

Thanks to some tweaking-fanatics (yes, you even find them in Intel) we now know that the current core will easily go up to 4.0GHz, possibly even 4.6GHz. That would mean a part of the core would be running at a rate of 8 or 9.2GHz. Of course, heat problems will arise, but some good cooling could solve those and anyway, what would you rather have: a lot of heat to get rid off, or a core that just doesn't go any higher?

Extreme cooling, would this be healthy for you?

But Northwood won't be with us for ever, some more releases in 2003 and the core will have to make way for a newer design, codenamed Prescott. Just like AMD's K8, the Prescott will contain some new features. Here's what we can expect to find:

HyperThreading

Hyperthreading is a much talked about feature, but all this technique does is use the optimal amount of power available by acting as if two threads are running at the same time. Of course, this won't give you the same speed as a dual-CPU configuration, but a 20 percent inprovement in multi-tasking environments will be possible. Those of you who have been following the reports lately will have noticed that HTT isn't new: the Xeon series has it and soon the Northwood will support it too. Prescott, however, will probably contain a improved version of HTT. It's sure that some minor glitches will be removed and it's said that the CPU might even have four threads ready for usage at any time, thus using its resources even better than the two-threaded version does. It's also possible that only the server-version of the Prescott, codenamed Nocona, will support this feature.

LaGrande

The name LaGrande was first heard at the Intel Developer Forum. It will be a part of the CPU that's especially security-minded. While the exact purpose of this function remains vague, it appears to be a hardware solution for encryption and/or hashing. There's nothing new about that, there are plenty PCI-cards that help to speed up the RSA algorithm, but it will be the first time such a thing will be implemented on such a wide scale. LaGrande will give us an advantage wherever we find encryption, and that's not just for secure data transfer or storage. Digital autographs for instance, could draw benefit from LaGrande. Also, when tracking changes in a system, a lot of principles from cryptology are used. The system that Microsoft invented, called Palladium, will probably use this function too.

LaGrande is a good idea, certainly when viewed from a PR-viewpoint. The more money is spent online and the more important security becomes while accessing data from the Internet, the bigger the demand for hacker-proof security systems will be. The Prescott can be introduced as a device that will make online shopping safer, improve privacy and make computers more reliable. It's not that such things are impossible without Prescott, but thanks to the RC5-64 project we know that the Pentium 4, when compared to the Pentium III and Athlon, isn't that good at working with RSA. It might be better at better algorithmes, but the fact remains that when security grows, the CPU's resources must too. LaGrande should help there, so maybe our Dutch Power Cows will profit from it too.

SSE3

The Pentium has MMX, the Pentium III has SSE, the Pentium IV could welcome SSE2 and Prescott will get SSE3. Although a lot of people doubt the use of these improvements, in practice, they do have their advantages. Not all software supports it, but those that do will get a nice performance boost. Drivers can also use the extra instructions, something nVidia thankfully implements to give their Detonators some extra performance from time to time. The exact benefits of SSE3 are not clear yet, although a part of it has been designed to support a new type of speach recognition that Intel is developing.

The platform

The FSB of the Prescott will be running at 166MHz and will be quadpumped like any other Pentium 4, which produces a datarate of 667MHz. Anyone who knows that the bus is 64 bits wide, can calculate that approximately 5GB/s can be pumped to the processor. Looking at the bandwith, this is a perfect fit with the two DDR333 channels or PC1200 RDRAM. Although it's actually only a 25% increase of what the 533MHz Northwood bus can deliver, it's important that the multiplier can be reduced, because this will always have a positive effect on the latency. The increase in distance, measured in clockcycles, between memory and processor can be partially compensated by the implementation of cache. Willamette had 256KB, Northwood doubled that and the Prescott will probably get a full megabyte: more then was available to any desktop computer so far. In comparison: ClawHammer will initially receive 256KB cache and a memorybus of 2.67GB/s.

Those who want a Prescott processor in their system can choose between several different chipsets. Intel itself offers the Springdale, the followup to the Granite Bay. Granite Bay, officially known as E7205, is a chipset that still needs to be brought out, and that supports dual channel DDR266 and AGP8x, while the southbridge will be equipped with Serial ATA 150. Springdale will add three things to that list; a 667MHz FSB, support for DDR333 memory, or DDR-II if the market is ready for it. It will also feature a special link to the northbridge that will provide gigabit ethernet and will relieve the southbridge of that burden. The southbridge ICH5 will have opportunities for eight USB 2.0 ports, eight PCI masters and Wi-Fi. VIA in the meantime works on a pair of Prescott chipsets: P4X600 with dual channel DDR333 and a version with DDR-II P4X800.

The answer to HyperTransport

Of course also server-editions of the Prescott will become available, in the form of Nocona. Since the Itanium-serie has been using the Pentium 4 bus since the McKinley core, the chipsets that where developed for the Itanium 2 (and later) can also be used by processors like Prescott and Nocona. The most interesting candidate for such an exchange is a version of E8870. By using a system with Scalability Ports and I/O hubs, with this new generation chipset a network of chips can almost be built in the same easy and flexible way as with the HyperTransport based AMD-8000. The major difference between the two is that AMD has placed HyperTransport and the memory controller on the processor itself. This ensures a reduced latency and probably also cheaper mainboards, but it will probably be a less flexible solution when you're thinking about an upgrade. In addition Intel can still use simple north- and southbridge mainboards, whereas a Hammer mainboard cannot ignore HyperTransport.

Yamhill
Officially it's still being denied, but it has been confirmed by so many sources that there's almost no possibillity for doubt; Intel has developed a form of x86-64 technology. A two percent increase in die-space would be sufficient to expand the core with 64 bit possibillities, just like AMD did (although these probably aren't mutually compatible). It's unlikely that this techique will be enabled in the first generation of Prescotts. First because it would not be good for IA-64 and second because it had to be announced already if it were, this in order to give developers the chance to adapt their software. Should the situation occur that 4GB of memory isn't enough for desktops, before the Itanium for home users is a reality, then Intel is prepared to extend the 32 bit series.

0,09 micron technology

AMD showed a couple of weeks ago that they are capable of making the 0.13µ Thoroughbred-B. Fab30 in Dresden switched from 0.18µ to 0.13µ technology but it wasn't easy: AMD cooperates now with UMC and Infineon hoping that switching over to a new process in the future will go smoother. Intel stays number one in production capacity, research staff and budget, which results in a six months leap of production technology. When AMD introduced their 1.8GHz Thoroughbred, and thus showed their 0.13µ technology for the first time, Intel showed their first working model of 0.09µ chips. Intel plans to build three or four 90nm production labs within one year, while AMD expects their first 90nm sample in the next six months.

Prescott has twice as many transistors as the Northwood, which means that all of the new features (and maybe even unannounced features) and extra cache will fit into the core. It has more than 100 billion transistors where the Northwood has 55 billion. If the Prescott would be built on 0.13µ technology, then it would be twice as expensive in production. Don't even think about the large cache Xeon MP, based on Nocona, which has 330 billion transistors. The Px60, which is used by Intel for the 0.13µ process, has gates of 130nm by 70nm, The successor of the Px60, P1262, will have gates of 90nm by 50nm. The transistors will be twice as small. Exact figures can't be given, but it's certain that the Prescott will be smaller than the Northwood and maybe even smaller than the ClawHammer, because the latter is bigger than the Athlon.

A closeup of the 0.18 micron Pentium 4.

Strained silicon

Because the Prescott is intended to reach the 6GHz the production process will involve more than only smaller transistors. One of the other 'features' is strained silicon, the counterpart of SOI. A transistor basically only does two things: first of all, it doesn't let power through when it's closed. Second, it has to have the lowest resistance possible when power goes through. Resistance and power causes heat, the number one enemy of every processor. The SOI technology, which AMD is going to apply on the Hammer, is meant to prevent leaks that causes the heat. Intel didn't choose SOI, because they think that SOI has disadvantages which will not be resolved for a few years. Strained silicon is used to reduce the resistance. In 0.09µ an extra layer of interconnects has been added. These thin wires that connect the transistors will have an extra wire: seven wires instead of six, so they can work with the huge ammount of transistors.

SiGe

Prescott will also be the first Intel chip which uses SiGe (Silicium-Germanium) technology. By mixing germanium with silicium, when the wafers are being produced, favourable properties arise. Germanium is a better semiconductor then silicium; it's suitable for speeds above 200MHz. The only problem is that it is useless when it hits seventy degree Celsius. With more than 100 billion transistors working at several hundred gigahertz, smaller than a flu virus and just a simple fan for blowing the heat away, this is not an acceptable limit. Don't forget that when you read the temperature with Motherboard Monitor, you measure the outside temperature and not the temperature on the processor itself. The SiGe mixture is an attempt to combine the properties from germanium with silicium: high frequency network chips from IBM showed us that this strategy works. The Prescott will work at 1.2V and, according to intel, in the future the voltage could even be lowered further.

Upcoming generations and conclusion

Prescott is part of a strategic plan to bring the Pentium 4 architecture up to ten gigahertz, and Intel already has a very good idea about what techniques are going to be used to achieve that goal. Although plans as these are very easily subject to changes, it seems like the chip planned as Prescott's succesor will be the first chip on earth to run at ten trillion cycles per second. The CPU, codenamed 'Tejas', will be built by using 65nm technology, enabling the use of even more transistors that will allow the CPU to go higher than ever and with the use of less power than current processors need.

Tejas will be equipped with so-called Depleted Sustrate Transistors, that leak ten thousand times less power than the current technology does, and will boast the first implementation of SOI. Those features, together with some others, will bring the ten gigahertz limit within range before the end of 2005. The early rumours tell us that Tejas will have a 1.2GHz FSB and will be partially compatible with IA-64 instructions. Ten gigahertz won't be the end though; Intel expects to hit twenty gigahertz a few years later. Double and triple gate transistors built at 0.045 and 0.032 micron still have a lot of potential and when those are used to their maximum, nano-technology will give us another performance boost.

Conclusion

After reading all the info in this article, you probably realize that Prescott is going to be a big and important release for Intel. In fact, the new CPU will be equal to the Pentium as Hammer is to the Athlon. There are strong indications that Intel plans a major PR-offensive to make the public aware of all the facts we just summed up in this review. There's really only one logic way to show people that a CPU is ready to take on its competitors and thats by giving it a new name. Prescott will probably be introduced as Pentium 5, something Paul Otellini as good as confirmed at the Intel Develloper Forum:

La Grande that will be implemented in its next generation Prescott processor. [...] "La Grande is not a Pentium 4 product. It will be a next generation architecture after Pentium 4," Otellini explained.

With a new logo and a new commercials loaded with expressions as 'safety', 'performance', 'low power consumption' and of course the well known amount of Gigaherz's, Intel's PR-department is hoping to leave behind the P4's bad name, that it got directly after its launch. The P5 will indicate how far Intel is ahead in technology, and give the sales-boost that the 0.18 to 0.13 micron swich could not give. Of course, Intel could drop the entire 'Pentium' name, but that would not be a smart move; the name is too well known and a fifth line of CPU's - together with Pentium, Xeon, Celeron and Itanium - would only create confusion.

Who knows the secrets of this building.

The ClawHammer's delay is very bad news for AMD. In his overview of the Hammer, Femme Taken already concluded that AMD was fairly safe until Prescott would launch. However, things have changed. The first thing that must be considered is the upcoming release of even faster Pentium 4's (the 3.06GHz and 3.2GHz). Secondly you should keep in mind Intel's decision to activate HyperThreading in the Northwood core. Hammer will be launched at the start of the second quarter of 2003, right in the middle of 3GHz+ hyperthreading Northwoods, only a bare few months away from what could become Intel's biggest ad-campaign ever. Those who have been 'in the business' long enough might remember the Pentium II that, with only one die-shrink and SSE instructions added, suddenly became Pentium III. Was that necessary? No. Did it work? Yes: in financial terms the Pentium III era (1999-2000) was Intel's most profitable period ever.

SledgeHammer (Opteron) and ClawHammer (Athlon 64?) will at least be a nice victory for AMD. The company has done everything it could to keep Athlon in the race against it's major competitors. 64bit instructions will be a good option for servers and workstations containing more than 4GB of memory. The new platform and the release date, which is still quite some time before Prescott, will add to AMD's market share. The most important thing for AMD, however, is that the Hammer will bring in more money per sold CPU, giving the company some breathing space. Despite all this, experts still think AMD won't be profitable in 2003 either. That just shows how hard AMD needs Hammer.

Meanwhile, Intel recovered from its errors and is clearly planning to stay ahead by using superior production technology. Prescott's new features may be less revolutionary than Hammer's, but a combination of a big PR-budget and the new name 'Pentium 5' will almost certainly ensure a victory for Intel. When the dust has settled, in about a year, people will conclude that not much has changed. Not in the power-balance and not in the performance difference between high-end AMD and Intel CPU's. 2003 will be an interesting year, though. Both companies will do their very best, and we don't have to choose sides until we need a new processor...