Semi - Given the not-too-bad shipment numbers for the simpler socket 754 chip, this would seem to indicate some sort of yield problem with the more complex chip. They can make some, but not a lot.

I'm shocked...

Semi - I think it ended up as intending to be an eyesore. After what the Texas State Legislature did, I think intel wanted to remind them of it for awhile.

What did they do?

Bull and Dell have added Itanium SPEC scores:

http://www.specbench.org/cpu2000/results/res2003q3/

Sgolds - Recall that back in the 486 days, there was only cache, and it was off-chip

The i486 had 8K ondie cache. The i486DX4 had 16K ondie cache. The chipsets offered L2 offdie.

As for L1, L2, L3 etc, I think there are some highend chipset designs that even have an external L4.

Please Joe - give it a rest.

Paul - do you know if the P4 Xeons w/ L3 all have the same size L2, and if so, what size is it?

They are all 512K now.

Paul - I didn't know that, but I do assume they meant L2 because of the group it's in. Odd.

Yes it is. When Intel announced the product they said in their press release that they doubled the cache to 1Meg when anyone can do the math and see they tripled it. One possible explanation is that the L3 in inclusive so 512K is just a copy of the L2, effectively adding only 512K so the term doubling would be correct. Maybe someone else can explain this?

sgolds - We know from the release of the processor in April that what is really impressive is the scaling. The Opteron platform runs circles around the Xeon platform because as soon as you put in that second (or more) processor, there is no comparison.

That's not what benchmarks show. While Opteron does scale better in SPECfp_rate the other dual/SMP benchmarks, for the most part, go to Xeon.

Paul - The $690 3.06 Xeon has 1M L2 cache w/ no L3 cache.

This is puzzling. Intel doesn't make a Xeon with 1Meg of L2 yet we continue to see reference to a 1Meg cache product. The product does not say L2 but it doesn't say L3 either. I'm confused about why Intel does this.

AMD Clusters its Opteron Agenda

Aug 01, 2003 (Internet.com via COMTEX) --

AMD Wednesday began tooting its own horn when it comes to how well its new Opteron processors do in high performance computing (HPC) and clustering environments.

The Sunnyvale, Calif.-based semiconductor maker announced two major projects within a week of each other that would put its three-month-old chips in the top five fastest machines as ranked by the independent TOP500 List of Supercomputers.

AMD Enterprise Business Development Rick Indyke said the company is finding itself holding its own against its main rival's Itanium processors. The boast is based on speed, configurations and price comparisons.

"What we find there is the value for AMD," Indyke told internetnews.com . "We see ourselves fitting between Xeon and Itanium with additional benefits. For example, if a system spreads out over multiple processors with a memory intensive application, we do better. It really has been a case, by case basis though."

Two Supercomputer spots doesn't do much for toppling Intel's reign as chipmaker with the most TOP500 systems. But if the Opteron-based Chinese "Red Grid" and Japanese IBM eServer Linux units break into the Top 10, AMD could push out Intel's Itanium configuration at the Pacific Northwest National Laboratory.

Supercomputing aside, AMD says it has bigger fish to fry.

"Right now, early adopters are using the Opteron. They are taking the risk to get them into their platforms," Indyke said. "But what will validate it as a technology is moving it from the HPC market to the enterprise and the public sector, including the U.S. government."

Indyke says companies are really kicking the tires on Opteron and surveying the chip's claims of 64-bit processing and 32-bit backwards compatibility. Some are beginning to produce custom applications for the chip. As a validation, AMD says it has more than 50 "Proof of Concept" requests compared to 33 when Opteron first launched.

"What is interesting is the end user community is going to the manufacturer community and they are interested in what IBM is doing with AMD Indyke said. "We've got Oracle, IBM, Microsoft all looking at our chips to use on the Web server side, in development tools. People don't buy hardware for hardware's sake."

The company is currently in the middle of a 32 city road trip to educate companies about the benefits of Opteron.

When the chip was first being scrutinized back in April, analysts like Gartner's Joe Byrne told internetnews.com the people who are attracted to Opteron are the same that are attracted to AMD today.

"The beauty of the AMD strategy is that success is not contingent on how fast they run but on how they allow old code to run faster," Byrne said. "There has to be systems based around the Opteron and it has to be stable. Ideally the Microsoft and IBM would have compatible software and hardware available today. The second best scenario is to have support, which is what they did. It does look like they are being aggressive in pricing, but it will be awhile before we see the true industry impact."

Indyke says it won't be much longer, predicting major news in the next three months that will center on open source and Windows environments. Beyond that, Indyke says expect to see a lot of activity in the next six months.

"We have a number of bids we are working on including a couple with 512 node opportunity and at least two that are double that, all within in the United States," he said.

HP's 64-Way Superdome-Unix Combo Tops TPC-C Test

Aug 01, 2003 (ComputerWire via COMTEX) --

The techies at Hewlett Packard Co have been working overtime trying to squeeze every bit of performance out of the new 64-way Integrity Superdome machines that use Intel Corp's "Madison" Itanium 2 processors. HP announced yesterday that it had beat IBM Corp's best result on the TPC-C online transaction processing benchmark test, breaking the 800,000 transactions per minute (TPM) barrier for the first time.

HP broke the 700,000 TPM back in April when Windows Server 2003 was launched by Microsoft Corp. That test was also on a 64-way Integrity Superdome using the 1.5GHz/6MB L2 cache version of the Itanium 2 processors; this machine ran Microsoft's Windows Server 2003 Datacenter Edition and SQL Server 2000 Enterprise Edition. This time around, HP tested the same iron using its HP-UX 11i variant of the Unix operating system and the forthcoming Oracle 10G grid-enabled relational database from Oracle Corp. While Oracle 9i Enterprise Edition sells for $40,000 per processor with a 25% discount or $1.9m on a 64-way Superdome (using PA-RISC or Itanium 2 chips), Oracle 10G will, according to HP's TPC-C report, sell for $20,000 per processor. With a 25% discount, this puts Oracle 10G in the same pricing band as SQL Server on the Superdome machines. This is not an accident.

On the HP-UX/Oracle 10G test, the Superdome machine with 512GB of main memory and 64 Itanium 2 processors cost $4.36m. Main memory was half of that, Itanium processors were a quarter, and the rest was sundry peripherals and boxes. The 26.5TB of disk storage attached to this server had a list price of $5.16m, and HP-UX and Oracle 10G cost $1.43m. Add in application servers and three years of maintenance, and the whole shebang cost $13.27m. After a whopping 51% discount for a large systems configuration, HP was able to show the machine could deliver 824,165 TPM at a cost of $8.28 per TPM. HP's discounting has progressed from 38.5% on Windows boxes in April. On the latest Superdome tests running Windows, HP delivered 707,102 TPM at a cost of $8.44 per TPM. That configuration had a 36% discount. While the HP boxes running HP-UX and Oracle 10G can demonstrate a 17% performance benefit compared to the same iron running Windows 2003 and SQL Server, the price/performance only comes in line - even after using a much less expensive version of Oracle - with much steeper discounting. The funny thing is that the Windows Superdome had 10TB more of disk, and still cost less. That's the difference between using Compaq's Modular SAN Array 1000s on Windows and using the Surestore Virtual Array 7110 on the HP-UX box. The HP-UX and Oracle 10G performance probably depended to a certain extent on the Surestore arrays.

In late June, IBM put the 1.7GHz Power4+ processors in its 32-way "Regatta-H" pSeries 690 servers, and was able to crank through 763,898 TPM running its AIX variant of Unix and its DB2 database. The Regatta-H server had 32 1.7GHz Power4+ processor cores (with a 1.5 MB shared, on-chip L2 cache for every two cores), 512MB of L3 cache, and 512GB of main memory. The Regatta-H server cost $3.27 million, with $1.38 million going for main memory alone and $1.5m going for processors. AIX and DB2 for the server cost $632,725. The pSeries 690 server cost $10.7m at list price, but IBM tossed in a 41% large systems discount, which dropped the cost of the machine down to $8.31 per TPM.

IBM is said to be readying a benchmark of its own showing Oracle performance on this Regatta-H box, and the word on the street is that IBM will deliver a Power4+ processor running at 2GHz or faster before the year is out. Traditionally, IBM has announced high-end servers in September or October, and there is no reason to believe that this will not happen again. Such a machine could hit a performance of anywhere between 875,000 TPM to 890,000 TPM running DB2, and could break 900,000 TPM running Oracle.

ComputerWire News: Issue 4724, August 01, 2003

Neye_eve - were the benchmarks done in uniprocessor or DP configuration?

Both but it should also be noted that it was using Windows2000 Server. iHT works better under XP so the dual processor scores would have been even better for Xeon if they had used XP instead of W2K. As it was Xeon did very well anyway.

Joe - The ignore feature works just fine.

wbmw - I have, however, seen "intellectually dishonest" posts from several people who post on this board.

The ignore feature works just fine.

Ron - What name did you post under on SI?

Guys I suggest we get back on track before the moderators have to scold us.

ChrisC_R - I for one hope ChipGuy continues to post here. He's added very valuable technical insight. Too bad you've missed it.

Petz - So the AMD component is 1,058 2-way 1U Opteron 2 GHz servers (IBM's eServer 325), and the Intel component is 130 4-way (SIZE?) 1.3 GHz servers.

I couldn't find that Intel information in the link you provided.

sgolds - Now, arguing that AMD should get out of processors and focus on memory, for instance, may be an interesting argument but it isn't the one on the table in JerryR's post of yesterday. He said AMD should have focussed on mobile processors only, and that would be certain death because then Intel would raise desktop prices and lower mobile prices.

The general thesis remains. Had AMD not pursued a business plan of doing harm to Intel at all costs there would have been a number of business avenues open to them, Flash, uCs, PLDs, Comms etc. The risks to shareholders would have been lower and the potential for meaningful sustained profits would have been much greater, IMHO. We've discussed those options in depth and don't need to go into them at this time but the point remains that Jerry S. committed AMD to x86/64. It's too late to go back.

Petz - NOT! What Joe and I have just shown is that AMD is bombing where there is no anti-aircraft fire.

Again you show no regard for us shareholders. AMD needs to stop blowing themselves up and start thinking about markets where they can generate longterm profits. Forget saving the world from the evil empire and think about enhancing shareholder value.

Petz - In fact, AMD has never lost "billions." (Its loss over the last two years is 1.792B :)

Yes 8 straight quarters of losses going on 9,10,.... but add in the ~$800 million they paid for Nexgen and you have "Billions".

Its [Itanium x86 compatibility mode] performance in this mode is so slow (Pentium 2 - level) that the capability is useless.

So you admit your statement was wrong. Meanwhile I agree that Merced x86 performance was disappointing but please support your claim about Madison.

And you gave no answer for the 64-bit desktop market, since Intel's mouthpiece still claims it doesn't exist.

AMD doesn't have a 64 bit desktop solution either. Get back to me when it does.

Joe - There is no competition to Intel in microprocessors other than AMD.

So AMD has taken the role of being the competition at the expense of shareholders. That's the point. This is good for the market but not good for shareholders.

Intel can always depress prices in one segment, where there is competition, but at the same time, Intel can't depress prices in all markets, because it would result in losses.

Haven't you noticed, it's already resulted in losses for AMD. Why would this result in losses for Intel? They keep making money hand over fist despite AMD's price wars. AMD is blowing itself up in hopes that some of the blast will hit Intel. AMD is the suicide bomber of the semiconductor industry.

Petz - AMD's strategy for Opteron is a winner.

I guess that explains the billions in losses over the last 2 years.

For the first time they designed a product for a market that Intel was ignoring rather than trying to beat Intel in a market in which it was the leader. I refer to 32-bit compatible 64-bit capable servers. Intel has none and AMD does.

Wrong again. Itanium is 64 bit and it's fully compatible. It executes x86 binaries in hardware.

It is not at all the same strategy as K5, K6 or Athlon and anybody that can't see the difference is blind.

I can see oceans of red ink. I think you're the one who's blind.

Joe - I thought you knew it was low.

Then that's my estimate. It's low.

Joe - I don't know, it is just an assumption. I am surprised that you are taking issue with this.

I took issue because you said this

There has to be tons of idle capacity, fabs that are not fully ramped.

But you just answered. There doesn't have to be tons of idle capacity. You just assumed there was.

borusa - AMD is the #2 manufacturer of pc processors in the (known universe, all deference to SETI) world.

AMD is the most unprofitable CPU supplier in the known universe. AMD has gone nowhere in 20 years. AMD has the worst growth rate among x86 CPU suppliers. How does this server the shareholders.

Sorry, but competeing with Intel goes with that territory.

Competing with Intel only goes with the territory for companies that choose to compete with Intel. AMD made that choice. How does this serve the shareholders?

The processors both AMD and Intel make are suitable for many purposes from embeded, mobile, desktop, WS, server. Thats just the way it is with a well designed processor.

Designs are for academic discussions. Products are what generate profits. A great design can still be a bad product as AMD continues to demonstrate. How does this serve the shareholder?

sgolds - AMD's road is difficult, that I will grant. However, there needs to be competition for Intel in every major CPU market, or the price game for Intel becomes too obvious.

I think you just made Jerry R's point. You just described AMD's role in terms of battling Intel. You don't give making a profit so much as a mention. AMD's #1 responsibility should be to the shareholders, that's us. While the market needs something to counterbalance Intel, why do AMD's shareholders always have to be the ones to bear that burden? Why do we have to finance the holy war? Why is it AMD's destiny to battle Intel?

Jerry R -The essence of my arguement is that AMD should have mastered a single, high-margin market like mobile before tackling even more formidable architectures.

Jerry your arguments are dead on and exactly what many of us have been saying for years. But realize that Hector was handed the helm with the course already set. Jerry pushed Hector out of the plane with only an unfinished Opteron parachute. There's no going back and no way to change things. Either it works before they hit the ground or it doesn't.

Joe - What is the Opteron yield?

It's too early to estimate.

Joe - There has to be tons of idle capacity, fabs that are not fully ramped.

Why is that Joe? Please show us how Intel can be the world's largest supplier of uPs, Chipsets and Flash plus a major player in uCs, network controllers, PDAs, Ethernet, automotive, wireless etc etc yet somehow you know that there is excess capacity. How do you come about this information? Please share with us just how much capacity it actually takes to produce all the products they make.

HP Integrity Superdome Performance Results (as of 7/30/03)


http://www.hp.com/products1/servers/integrity/performance.html

HP Integrity Superdome Performance Results (as of 7/30/03)


http://www.hp.com/products1/servers/integrity/performance.html

HP and Oracle Set Transaction Processing World Record - Break 800K Barrier

http://www.hp.com/hpinfo/newsroom/press/2003/030730a.html

HP and Oracle Set Transaction Processing World Record - Break 800K Barrier

HP Integrity Servers and Oracle Database Establish Record-breaking Benchmarks on Multiple Operating Systems Across Diverse Workloads

PALO ALTO, CALIF., JULY 30, 2003
--------------------------------------------------------------------------------

HP (NYSE:HPQ) and Oracle Corp. (Nasdaq:ORCL) today announced the world-record benchmark result of 824,164 transactions per minute (tpmC) on the Transaction Processing Council's TPC-C benchmark(1) achieved by the first ever system, clustered or non-clustered, to eclipse the 800K tpmC barrier.

An HP Integrity Superdome server with 64 Intel® Itanium® 2 processors 6M running the 64-bit HP-UX 11i v2 operating system, HP StorageWorks Virtual Array 7110 and Oracle Database technology outperformed all competitive offerings for transaction processing, while also delivering a top 10 TPC-C single-system price to performance ratio for high-end system of $8.28/tpmC.(1)

"HP's strong alliance and strategic development with Oracle enables us to deliver performance that firmly establishes HP Integrity servers as the standard for 64-bit architectures," said Rich Marcello, senior vice president and general manager, HP Business Critical Servers. "With world-record performance on an industry-standard platform, HP Integrity servers offer customers the flexibility and value required for the demanding business challenges of managing an adaptive enterprise."

HP and Oracle also announced the leading TPC-C benchmark result for a four-processor system. The HP Integrity rx5670 server running HP-UX 11i v2 and Oracle Database technology posted a result of 131,639 tpmC with a price to performance ratio of $7.25/tpmC,(2) beating all previously published UNIX based server results on transaction processing cost.

"Oracle Database technology and HP Integrity servers provide a strong platform for running both memory and processing intensive solutions," said Andrew Mendelsohn, senior vice president, Database Server Technologies at Oracle. "Both companies are committed to providing a scalable, reliable and flexible platform to run demanding mission-critical solutions, while meeting customer requirements of performance and better business value."

Record Result for Oracle Applications Standard Benchmark

HP and Oracle also established a new world record for the Oracle Applications Standard Benchmark (OASB) on four-processor systems. The HP Integrity rx5670 server running HP-UX 11i v2 and Oracle Database achieved 6,440 concurrent users, with an average response time of 0.6 seconds.(3)

The companies also produced the best ever four-processor OASB result for a Linux-based server, achieving 5,992 concurrent users, with an average response time of 0.6 seconds,(4) running Red Hat Enterprise Linux AS 2.1 and Oracle Database technology. Both the four-processor UNIX and Linux-based operating system results exceed all other four-processor system results across all other architectures on the market.

HP Integrity Servers Deliver Record-setting Performance

"Customers are beginning to ask more of the standards-based server products they are installing in their enterprise data centers," said Jean S. Bozman, research vice president, IDC. "The migration to industry-standard hardware components is driven mainly by cost advantages conferred by volume shipments. But the ability to produce high-performance benchmarks and to support scalability of 64 processors, or more, on a server system that is based on industry-standard hardware components is noteworthy, and it should contribute to accelerated adoption of HP's Itanium 2-based Integrity servers for commercial application deployments."

In addition to these benchmark records, HP Integrity servers, ranging from one- and two-processor to 64-processor systems, have established dozens of record benchmark results on multiple operating systems and workload categories. Record-breaking results range from the first 64-way SPECjbb2000 Java Server benchmark of more than 1 million operations per second to the fastest 64-way microprocessor Linpack performance and record-setting SPECweb99_SSL performance.(5)

A complete list of record benchmark results produced on HP's Integrity servers is available at www.hp.com/products1/servers/integrity/performance.html.

wbmw - can you explain how that works?

Sure. You said you'd like to buy more AMD if it drops below $7.00, right? Let's assume you wanted to buy 1000 shares at or below $7.00. You sell 10 contracts of Sept $7.00 Puts and put about $500 in your pocket now. You will be obligated to buy 1000 shares at any time up to expiration on the 3rd Friday of Sept. Potential downside is that AMD could temporarily drop to $6.50 then return to $7+ but your hands are tied and you can't buy at $6.50. You may never get assigned at all but you still keep that $500.

INTEL CORPORATION: Intel joins Eclipse consortium

SANTA CLARA, Calif., Jul 30, 2003 (M2 PRESSWIRE via COMTEX) --

Intel Corporation today announced that it has become a supporting member of Eclipse, a consortium of software development tool vendors working together to create an open platform for tool integration.

Intel will be represented on the Eclipse Board of Stewards by Jonathan Khazam, general manager of Intel's Software Products Division and will offer plug-ins for the Eclipse development environment to help developers optimize applications for Intel Pentium 4, Intel Xeon(TM), and Intel Itanium processor-based systems.

"Intel is pleased to participate as a member of Eclipse and will play an active role in fostering tool interoperability and integration within the Eclipse framework," said Jonathan Khazam. "Our commitment to open standards and deep knowledge of processor architecture will help us enable software developers to harness the full power of Intel processors for applications."

Intel plans to integrate its development tools, such as Intel Compilers and Intel VTune(TM) Performance Analyzers, into the Eclipse development environment. More information on Intel's tools is available at www.intel.com/software/products.

"Eclipse applauds Intel's significant decision to embrace Eclipse technology and become a supporting member of the consortium," said Skip McGaughey, chairperson of the Eclipse Board of Stewards. "We expect Intel's leadership and experience in processor and tools implementation to benefit all developers that utilize Eclipse based tools and offerings."

Eclipse is an open platform for tool integration built by a community of tool providers interested in creating better development environments and fostering product integration.

The Eclipse framework offers tool vendors a flexible plug-in architecture for easy integration that promotes interoperability between tools. Eclipse provides developers with a rich, easy-to-use development environment that can be customized with the tools needed to meet their specific challenge. Eclipse can be found on the Web at http://www.eclipse.org.

wbmw - I'm comfortable with my current shares, but I'll buy more if they fall below $7 again.

So write some $7.00 Puts.

Petz - So Intel makes its chipsets using older fabs unsuitable for CPUs. This contradicts your argument that chipsets take a lot of Intel's fab capacity.

Huh?

There's no way you can make a 150mm chip plus a 20mm chip any cheaper than you can make a 165mm chip, especially when you include the additional costs for two lines, integrating and testing the interface, 2x packaging, additional motherboard costs, etc.

We're talking a 130mm2 chip and another 50-60mm2 chip. That brings it pretty close to a 193mm2 Opteron but Opteron is on a very expensive low yielding SOI process and the Intel versions aren't, assuming Intel has good yields.

July 28, 2003
HP, IBM Ready New 8-Way Servers


By Jeffrey Burt


Dell Inc.'s decision to exit the eight-processor server arena adds fuel to the growing scale-out versus scale-up data center architecture debate.

Dell has long been a vocal proponent of scaled-out data center architecture—one in which multiple smaller servers are linked to give enterprises computing power equal to that of larger symmetric multiprocessing systems.


Meanwhile, IBM and Hewlett-Packard Co. are each preparing upgrades that address the scale-up concept, which calls for adding bigger systems with more chips when greater processing power is required.

Dell, of Round Rock, Texas, last week announced it has stopped development of its eight-processor PowerEdge 8450, deciding instead to focus research and development on two- and four-way systems that could be used in scaled-out architectures.

Analysts said that within five years, the performance of two- and four-way systems will have improved enough to handle the larger back-end applications—such as databases—that the larger servers typically run now.

HP and IBM officials agreed that scale-out is a good option for some. But they still see eight-way servers as an important part of their data center strategies. HP, of Palo Alto, Calif., sees its eight-way ProLiant DL740 and DL760 G2 systems as bridges between its 32-bit Xeon-based systems and 64-bit Itanium-based servers, for example.

HP will refresh those eight-way servers early next year with the next Xeon MP chip from Intel Corp., said Colin Lacey, group manager for HP's MP Servers.

In addition, eight-processor systems are part of an overall plan by HP to streamline the hardware components of its server lines—IA-32, Itanium and PA-RISC—by sometime in 2005 and to create a common software stack for all three by the end of this year, Lacey said.

Meanwhile, IBM will continue to upgrade its eServer x440 line of Intel-based servers that can support up to 32 processors, said Deepak Advani, vice president of the Armonk, N.Y., company's xSeries servers. By the end of the year, IBM will introduce a system powered by anywhere from four to 16 of Intel's new Itanium 2 6M chips, giving enterprises high-end multiprocessor capability in 32-bit and 64-bit computing, Advani said.

"We do not see in two- and four-ways the raw horsepower we need with an eight-way box," said Jason Robohm, director of technical services at Crossmark Holdings Inc., in Plano, Texas.

While multiple smaller servers linked together may give users good failover protection, the larger systems offer such features as hot-swappable RAID and the ability to add processors as demand grows, said Robohm, who uses HP eight-way servers.

wbmw - Porting to AMD64 FAQ

So it's not quite so "fully compatible" as we've been lead to believe...