Sunday, December 10, 2006

Torrenza: Why AMD may not have to introduce a new processor

Over some of the past entries I've commented on Intel's Conroe/Core2 design inability to scale performance with processor speed. I've also commented on the Athlon64 design which does scale predictably has the speed increases.

However, AMD has yet to introduce a direct competitor to the ultra-high end Conroe/Core2 designs, such as Kentsfield, and nor does AMD seem to be in a hurry to do so. Many AMD fanaticts seem to be convinced that AMD will answer Intel Shot for Shot, and fully expect Quad-Core to regain the performance crown. Many analysts seem to believe that AMD will release another version of the K8/Hammer design that will make up the few percentage points lost to Core2.

AMD may, but they don't have to. Back in 2001 and 2002 AMD was beginning to show off the architecture they called Hammer, a fully 64bit processor with amazing potential. Early tests showed that Hammer was putting out performance numbers on par with AMD Thunderbird and Intel Willamette and Northwood designs with a 1ghz clock advantage.

When the AthlonXp 3200 launched, AMD knew that the Xp line was running out of steam. The Intel 3.2ghz Pentium4 was the clear winner in a majority of benchmarks. Rather than continuing to answer shot for shot with Intel, AMD pushed Hammer, the Athlon64 processor. Remember what Intel initially said about x86-64? That 64-bit was useless and that nobody wanted or needed it? And what actually happened as time went on? Intel only began x86-64 support with Prescott in 2005, 2 years after AMD, and even then applications compiled for x86-64 specs under Linux wouldn't run with Intel's version.

AMD redefined the battlefield rather than continuing an arms race of who could get the fastest clocks out of the processor.

Just the same with Core2 and Athlon64. AMD has already been in this position before. Intel, while slow, is capable of competing. The Athlon64 would not always be the most powerful processor available, AMD has never said otherwise. Just as the Pentium4 3.2ghz went beyond the AthlonXp 3200, something would go beyond Athlon64.

And again, AMD has redefined the battlefield, this time with the processor inter-connection than with a new processor design.

Torrenza is basically cache coherent Hypertransport, allowing for processors or chipsets to be directly aware of what other processors or chipsets are doing. By opening up CC-HyperTransport, AMD can link their processors directly with other processors that may not be of Athlon64/Hammer design. As is, Sun Microsystems, HP, and IBM are behind socket compatibility, placing their processors into AMD sockets. Sun is known for it's Sparc Architecture and the most recent Niagara design. HP has the PA-RISC architecture and what is left of the Alpha Architecture. IBM is one of the primary backers of the PowerPC architecture, and while Sony may be failing with Cell in the Playstation3, the Cell design is not one that can be dismissed.

Now, let me try to put some context here:

In the common computer today it is taken for granted that not all processors process in the same manner. Some processors are better at some tasks than others. AMD and Intel typically focus on the Central Processor, the brains behind the computer. The central processor collects, collates, arranges, and directs the computers other systems. ATi and Nvidia typically focus on graphics, designing high performance add-in cards to increase visual display quality. Creative markets a sound processor that has a focus on producing the best sound possible. AGEIA markets an add-in card that removes physics computations from the central processor. BigFoot makes and markets the Killer NIC, a hardware networking card.

There is no question that each of these different dedicated hardware parts are better in terms of performance than integrated parts. The difference may be noticeable in only a few cases, but it is there.

Now, given that we take for granted that a computer can have a multitude of different tasks that are each served best by one architecture, is it really too much of a stretch to think that the same applies to the Central Processor?

Sun's Niagara design is a monster in several server tasks, but even with the most powerful video card in the world, don't expect it to be able to run Unreal Tournament 1999. It is not built to run that kind of code.

What Torrenza enables is multi-central-processor systems that don't have to be one architecture. Lets say that somebody running a webserver may also want to host a game server as well. Today, that means having to buy two different computer systems. One that is tuned to serve web-pages and one that is tuned to serve games. With Torrenza, that person could buy one server, equip one socket with an Athlon64, and the other socket with a Sparc Niagara design processor, and be able to do both.

That is how the battlefield is being redefined. AMD's Athlon64 is not the be all/end all processor design. There are some tasks that it just will never be as good at. With Torrenza though, consumers can use a completely different processor which can be stronger in areas that the Athlon64 is not strong in.

This is why AMD doesn't have to rush another processor architecture out the door, or ramp the speed up on the existing Athlon64 designs. A single server with an Athlon64 processor and a Cell processor inside would render Conroe processor designs irrelevant. Intel can have the single application performance crown, that won't help when the competition is able to answer Conroe processor designs with equal performance in similar tasks, and also give a devastating performance difference in tasks that the Conroe architecture is not designed for.

Now, I do need to state that I am intentionally trying to avoid the Operating System question in relation to Torrenza. There is only one Operating System available that could possibly pull this merger of architectures off, and its not from Microsoft. It may also be years before we actually see anybody putting anything like this together. However, consider this:

took 2 years to add x86-64 support into Pentium4
is 3 years behind AMD on delivering comparable x86-64 performance
is at least 5 years behind on a Direct Connect Like Architecture

By the time Intel gets their Direct Connect Like Architecture out the door, Torrenza could already be fulfilling the scenario outlined here.
Post a Comment