banker, Re: I think the RISC market is going away as x86-64 grows up. IMO, SUN is already planning for that eventuality.
I'd argue that Sun is planning for the death of UltraSparc, which is different than the death of high end computing.
Re: over time (5-10 years), x86-64 will take over more and more of enterprise computing. Just look at the growth over the last 5 years when all that was offered was x86-32.
64-bits wasn't the only thing preventing x86 from competing in the high end markets. It's a completely platform related issue. In the high end, everything is designed to work well with everything else with the highest uptime and the highest platform level performance. That means that memory bandwidth is better, disk IO is better, fabric interconnect is better, along with everything else.
The ironic thing is that Intel is behind in a lot of the things I mentioned, and in spite of that, their IPF architecture sometimes more than makes up for it. I can't even fathom the size of the step function in performance they'll get when they leapfrog in the next product generation with a better serial interconnect, as well as a better bandwidth and higher latency memory interface. At least PCI-Express is helping them today as far as IO goes, but it's too bad IPF won't take advantage of it in the current generation.
Re: Once SAP/Oracle/MS/et.al. offer their enterprise apps certified to x86-64 and OEMs offer x86-64 servers with RAS features, why would anyone buy anything else.
RAS is not a static feature set that Opteron can easily catch up to. Eventually, Opteron will have most of today's RAS, but Intel and other RISC vendors are already implementing tomorrows RAS in next generation products. It's an evolving technology, and it's not always efficient in terms of design time, validation complexity, die size, or power to implement tomorrow's RAS features in today's *commodity* CPUs. That's what product segmentation is for. You put those features into products that people are willing to pay for, and the increased margins allows you to afford to make that product. This is Business 101.
Re: You can get a 4-way Opteron that offers superior performance for less than a 2-way Power or IPF system. Similarly you can get two 2-way Xeons that offer higher performance for less than a single 2-way Power/IPF system.
Except two 2-way systems isn't always as good as one 4-way system, and the differences are more pronounced with more CPUs. Did you read what the NASA guys said regarding the Altix supercomputer?
Meanwhile, the Agency considered more ambitious approaches. Particularly intriguing was the notion of building a world-class supercomputer by November. One idea was to link thousands of dual-processor commodity servers into a sprawling cluster, but NASA quickly dismissed that approach. "We're trying to solve some of the toughest scientific problems in the world," says Jim Taft, task lead for the NAS Division's Terascale Applications Group. "We needed a system designed to efficiently execute the algorithms used in NASA's premier science codes, rather than one that would merely do well on artificial benchmarks."
Clusters of small commodity systems work fine if you have multiple applets each requiring their own CPU processing. However, with higher end systems, the applications have been optimized to work better with many CPUs under a single operating system image. This is not just in the HPC world, but includes commercial apps like Oracle or SAP.
Re: The change won't happen overnight, as chips take a long time to die, so don't expect everyone to knife their proprietary chips anytime soon. Nevertheless, x86 will one day rule the high-end of the enterprise market.
There are more business aspects fighting against this ideal than you may think. Unless the things that differentiate high end systems - things like RAS, uptime, and platform performance like I mentioned above - are sufficiently easy and inexpensive to implement, they'll never make their way into commodity platforms.
The x86 market is so competitive that product lines need to be refreshed on a yearly basis. If you miss the cadence, then stretching the current product line will become harder to do. If you want to consider some examples, recall how much of a tough time AMD had when K8 was delayed from 2003 to 2004. Stretching the K7 put them into one of the most direst in the company's history. Likewise, Intel cancelling Tejas because of its overly complex design means that they have to stretch Prescott for an extra year, during which it will be far less competitive, thus giving AMD the largest opportunity in the company's history. That's the kind of change in events that can happen when the one year commodity processor cadence is broken.
That's why Intel or AMD simply can't put the kind of high end features of tomorrow's enterprise server in today's x86 chips. Instead, they are left with putting today's technology in tomorrow's chips, because that's how technology waterfalls from one generation to the next. It starts out in the high end where it is complex and expensive, and eventually Microsoft integrates it, the industry comes out with a standardized spec, and the commodity processors can implement it within their one-year cadence.
For Intel, IPF is more than a different architecture; it's also a different business model. Intel doesn't update their IPF cores, except for once per year in a step-like function. When they broke their cadence with Montecito, they needed to fill the hole with a very underwhelming Madison 9M. Montecito returns the cadence to n+1 technology.
Years ago, Intel made the mistake of trying to do two things with IPF: make it a world class enterprise chip, and also use volume economics to drive it into commodity markets. I think Intel is realizing they can't have their cake and eat it, too. That's why x86-64 will be their commodity chip going forward, and they will have the ability to integrate today's enterprise technology into tomorrow's CPUs to grow their Xeon business. Meanwhile, IPF will get the enterprise level RAS features, first, as well as the platform level expertise of the various OEMs ( such as HP or SGI). It will differentiate from other offerings, just like IBM can do with theirs.
Intel's bet going forward is another interesting business risk: to share the infrastructure of Xeon and IPF in order to reuse micro-architecture and infrastructure design, thus lowering product development costs. This may benefit IPF and help it to succeed in higher volume markets, but it also might hinder the architecture by removing differentiating features. Which one it is will depend on the Tukwila CPU and what kind of features made the cut. I think it's interesting to watch.
News 
Market Data 
Discover 
AI
