Slashdot Mirror

Copyright(c) 1999 - 2003 Intel Corporation. All rights reserved.

This program is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by the Free
Software Foundation; either version 2 of the License, or (at your option)
any later version.

This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
more details.

You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc., 59
Temple Place - Suite 330, Boston, MA 02111-1307, USA.

The full GNU General Public License is included in this distribution in the
file called LICENSE.

Contact Information:
Linux NICS
Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497

That driver is a GPL implementation meaning that the OpenBSD developers are more than welcome to port it at their leisure.

Oh, you want real documentation too. Take a look at developer.intel.com:
Intel 82551ER Fast Ethernet Controller Networking Silicon Datasheet
Intel 82559ER Fast Ethernet Controller Networking Silicon Datasheet
Intel 82559ER EEPROM Map and Programming information

True, Those are for their FastEthernet chipsets, not the Gigabit chipsets that, "Intel steadfastly refuses to provide us with documentation"

Well, what about these?
Intel(R) 82541ER Gigabit Ethernet Controller Networking Silicon Datasheet
Intel(R) 82547GI(EI)/82541ER EEPROM Map and Programming Information Guide

Ok, so the IPSEC chip on the NIC isn't supported nor is there any data on that chip forthcoming. However, there are a number of papers that show that IPSEC and TCP offloading (not to be confused with TCP fragmentation/checksumming) are not efficient. Specifically, the "hardware" IPSEC is done by firmware downloaded to a small embedded processor on the NIC. This small, embedded NIC is not very fast, in fact, its rather slow.

Result:
Processor utilization drops marginally (modern processors can encrypt 10 megabytes/s trivially)
Latency shoots up (It takes the embedded processor longer to encrypt a packet than the host processor would.)

There are a number of papers corroborating that latency has a huge effect on maximum bandwidth.

[I think the paper regarding TCP offloading not being worthwhile is by Mudge. The IPSEC offloading not being worthwhile is my hypothesis, untested, but I feel logically founded.]

My point is that IPSEC offloading is not an advantage - it probably was in 200MHz K6 days, but it certainly is not in 2.0 ghz K7 days.

Other notables, for example, the 3Com 3CR990 still doesn't have IPSEC offloading, despite promises from either the openbsd txp driver or the linux typhoon driver.

Frankly, as far as Gigabit

Copyright(c) 1999 - 2003 Intel Corporation. All rights reserved.

This program is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by the Free
Software Foundation; either version 2 of the License, or (at your option)
any later version.

This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
more details.

You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc., 59
Temple Place - Suite 330, Boston, MA 02111-1307, USA.

The full GNU General Public License is included in this distribution in the
file called LICENSE.

Contact Information:
Linux NICS
Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497

That driver is a GPL implementation meaning that the OpenBSD developers are more than welcome to port it at their leisure.

Oh, you want real documentation too. Take a look at developer.intel.com:
Intel 82551ER Fast Ethernet Controller Networking Silicon Datasheet
Intel 82559ER Fast Ethernet Controller Networking Silicon Datasheet
Intel 82559ER EEPROM Map and Programming information

True, Those are for their FastEthernet chipsets, not the Gigabit chipsets that, "Intel steadfastly refuses to provide us with documentation"

Well, what about these?
Intel(R) 82541ER Gigabit Ethernet Controller Networking Silicon Datasheet
Intel(R) 82547GI(EI)/82541ER EEPROM Map and Programming Information Guide

Ok, so the IPSEC chip on the NIC isn't supported nor is there any data on that chip forthcoming. However, there are a number of papers that show that IPSEC and TCP offloading (not to be confused with TCP fragmentation/checksumming) are not efficient. Specifically, the "hardware" IPSEC is done by firmware downloaded to a small embedded processor on the NIC. This small, embedded NIC is not very fast, in fact, its rather slow.

Result:
Processor utilization drops marginally (modern processors can encrypt 10 megabytes/s trivially)
Latency shoots up (It takes the embedded processor longer to encrypt a packet than the host processor would.)

There are a number of papers corroborating that latency has a huge effect on maximum bandwidth.

[I think the paper regarding TCP offloading not being worthwhile is by Mudge. The IPSEC offloading not being worthwhile is my hypothesis, untested, but I feel logically founded.]

My point is that IPSEC offloading is not an advantage - it probably was in 200MHz K6 days, but it certainly is not in 2.0 ghz K7 days.

Other notables, for example, the 3Com 3CR990 still doesn't have IPSEC offloading, despite promises from either the openbsd txp driver or the linux typhoon driver.

Frankly, as far as Gigabit

Copyright(c) 1999 - 2003 Intel Corporation. All rights reserved.

This program is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by the Free
Software Foundation; either version 2 of the License, or (at your option)
any later version.

This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
more details.

You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc., 59
Temple Place - Suite 330, Boston, MA 02111-1307, USA.

The full GNU General Public License is included in this distribution in the
file called LICENSE.

Contact Information:
Linux NICS
Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497

That driver is a GPL implementation meaning that the OpenBSD developers are more than welcome to port it at their leisure.

Oh, you want real documentation too. Take a look at developer.intel.com:
Intel 82551ER Fast Ethernet Controller Networking Silicon Datasheet
Intel 82559ER Fast Ethernet Controller Networking Silicon Datasheet
Intel 82559ER EEPROM Map and Programming information

True, Those are for their FastEthernet chipsets, not the Gigabit chipsets that, "Intel steadfastly refuses to provide us with documentation"

Well, what about these?
Intel(R) 82541ER Gigabit Ethernet Controller Networking Silicon Datasheet
Intel(R) 82547GI(EI)/82541ER EEPROM Map and Programming Information Guide

Ok, so the IPSEC chip on the NIC isn't supported nor is there any data on that chip forthcoming. However, there are a number of papers that show that IPSEC and TCP offloading (not to be confused with TCP fragmentation/checksumming) are not efficient. Specifically, the "hardware" IPSEC is done by firmware downloaded to a small embedded processor on the NIC. This small, embedded NIC is not very fast, in fact, its rather slow.

Result:
Processor utilization drops marginally (modern processors can encrypt 10 megabytes/s trivially)
Latency shoots up (It takes the embedded processor longer to encrypt a packet than the host processor would.)

There are a number of papers corroborating that latency has a huge effect on maximum bandwidth.

[I think the paper regarding TCP offloading not being worthwhile is by Mudge. The IPSEC offloading not being worthwhile is my hypothesis, untested, but I feel logically founded.]

My point is that IPSEC offloading is not an advantage - it probably was in 200MHz K6 days, but it certainly is not in 2.0 ghz K7 days.

Other notables, for example, the 3Com 3CR990 still doesn't have IPSEC offloading, despite promises from either the openbsd txp driver or the linux typhoon driver.

Frankly, as far as Gigabit

-Intel (and ST to a minor factor) are the only companies investing into this, all competitors have different technologies.

Not true - Lockheed Martin is also a developer. I should also note that there are many private developers that won't ever need a license if they ultimately can't create a working/profitable device.

-A demonstration of reliable high density operation has still to follow. Handling local temperature differences of several hundrend kelvins within a sold state device poses many problems. Thermal stress will contribute to wear out, characteristics drift etc etc..

Intel has already demonstrated a cycle life of 10^12 and expects data retention of 10 years at 120dC.

-How about process compatibility.. no mention about this, yet

Google for this. Intel is throwing the weight of their R&D to OUM because it is so process-friendly. They are currently running OUM alongside the Pentium 3.0Ghz line (not commercially viable / internal R&D use only).

A brief here...

Ovonyx non-volatile memory technology offers significantly faster write and erase speeds and higher cycling endurance than conventional Flash memory. It also has the advantage of a simple fabrication process, which allows the design of semiconductor chips with embedded non-volatile memory using only a few additional mask steps.

Interesting comments from the horse's mouth.

-Intel is also investing in competing technologies

Yes - but they are on record (see original EBN link from parent post) stating that OUM shows the most promise. But, as I did state earlier, I see this as extremely speculative. I would not be here if I was not a whore for the technology. I fully expect to do one of two things with my investment:

1) 100 fold return by 2010
2) lose it all

That is what speculation is all about.

Intel is using EFI for all its reference designs for the La_Grande/Palladium technology. So, the motherboard manufactures are pretty much going to be forced into this. And AMD won't save you. They have jumped onto the TCPA bandwagon.

Moreover your "untrusted" applications will not be able to access your "trusted" applications because they'll be on seperate virtual machines (thanks to Vanderpool, Intel's VMM on a chip). We'll just hope the two VMs share the same HD partition and you can edit the "trusted" binary (somehow, but I know that is not the point of TCPA).

Links:
EFi Article
Intel EFI site
Vanderpool Article

download this software from Pentium: It's on this screen

Intel has been working on EFI since 1998. This is just a rip off of that.

Intel has been working on EFI since 1998. This is just a rip off of that.

His name was Gordon Moore, and he worked for Intel... err, more to the point, he founded Intel along with Andy Grove and Robert Noyce. Of course, like pretty much all the important players in the semiconductor industry, he was working at Fairchild at the time that he released his paper. What he said was:

"The complexity for minimum component costs has increased at a rate of roughly a factor of two per year"

He predicted that this would hold true for 10 years, from 1965 to 1975. Of course, this prediction fairly quickly changed to doubling the transistor count every 18 months, since the 12 month prediction didn't hold true for very long. Now we're doubling transistor count roughly every 24 months.

FWIW anyone who's interested can take a quick read of Moore's whitepaper where this "law" came from.

I guess the choice of manufacturing is a bad word. Intel has said however that they have successfully made fully funtctioning chips in their development fab.

Intel in the past just turns their development fab into a manufacturing fab once it's time to start manufacturing the chip and there is no reason to beleive that it will be any different this time.

Intel's initial preference to supersede 200mm was 400mm, Applied Materials wanted 350mm. The industry settled on 300mm in the end, but there's currently a Japanese consortium working on 400mm wafers.

http://www.intel.com/technology/itj/q41998/article s/art_4.htm

http://www.arofe.army.mil/Reports/webelec/ssdm98z. htm

300nm does sound right since that is right were the cutting edge is for modern fabs.

Second, state of the art has been under 200nm for some time now, so 300nm would be a big step backward. For instance, see this article from three years ago touting 130nm, and then this one from over a ago touting 90nm.

300mm definitely does not refer to "wafer (?) slices" whatever the hell that is supposed to mean.

It's the size of the wafer, the round silicon thing with *many* chips on it.
See this link (to Intel, inappropriately) for more info.

There's a reason they ignored Itanium, it's about upcoming processor technologies. Last I checked there wasn't a new, soon to be released, Itanium that Intel was pushing.

In fact, the current Intel processor roadmap shows the same Itanium 2 processor for the first half of 2004 as it did for the second half of 2003.

I think this is actually what most two pass compilers will do. The intel compilers have this option for instance. Basically you compile w/ profiling on and on the second pass it using the runtime profiling data to recompile source to a new object code. (btw the intel compilers are "free" for opensource/nonprofit use). W/ regards to using GA/SA as part of the optimization it would be (to be honest) a bit weird to have a nondeterministic compiler since in different runs you may end up w/ different optimizations.

-bloo

Close! The way to compart the hardware is to use the compiler recommended by the Manufacturer for each product. Intel, for performance, would reccomend their own compiler while IBM, who contributed the G5 code generator to the GCC project, would recommend theirs (in this case, GCC.)

Using "GCC" for both isn't fair because the code generators and optimizers are completely different. The only fair thing is to use what the Manufacturer suggests for optimum performance.

Apple's initial benchmarks were weird, too, because they compared a machine that would not ship for FOUR months (and I'm being generous here) to a 6-month old DELL unit. Fair would be to ask Dell for a sample of a machine to be released next quarter and test against that.

As it is, the P4, even crippled with HT turned off, BEAT the G5 with its faster bus in all the integer tests. By Apple's own admission.

In his original paper, Moore observed an exponential growth in the number of transistors per integrated circuit and predicted that this trend would continue

A jump from 3ghZ to 500+ghZ isn't really spot on with Moore's Law is it? This is far from mere doubling.

The history of Moore's Law.

Or if you are interested in Moore's original paper, you can find it here.

The history of Moore's Law.

Or if you are interested in Moore's original paper, you can find it here.

Intel are working on linux drivers. They estimate them to be out around the first 1/2 of 2004.

According to Intel the real drivers are "in development" for the wireless chipset. Rumored release is in the first half of 2004.

The XScale has an order of magnitude better performance per watt than PPC - we're talking fractions of a watt flat out at 400MHz (see the spec sheet). Also note that this is an Intel/DEC improved ARM core, and isn't x86 based in the slightest.

This particular XScale also has a large wedge of on-chip flash and a bunch of handy peripherals for mobile devices - which is unsurprising seeing as that's what it seems to have been made for.

It is funny that Motorola favour other manufacturer's chips for embedded low power devices, but the reason is simple: none of their stuff fits the bill. Even their own mobile phones seem to use ARM these days.

The intel Open Source Computer Vision Library They have tools to process stereo usb cameras in real time

Gordon Moore made his famous observation in 1965, just four years after the first planar integrated circuit was discovered. The press called it "Moore's Law" and the name has stuck. In his original paper, Moore observed an exponential growth in the number of transistors per integrated circuit and predicted that this trend would continue.

Gordon Moore made his famous observation in 1965, just four years after the first planar integrated circuit was discovered. The press called it "Moore's Law" and the name has stuck. In his original paper, Moore observed an exponential growth in the number of transistors per integrated circuit and predicted that this trend would continue.

Once we see the first "personal server" come out, there will be no need for the phones to be smart.

What I'm not sure is how the interface to the personal server and the phones will shake out. Ideally, the phones could bluetooth to the server to retreive phone books, etc, but something tells me the phone makers may not want to go this route - though who knows...

Sheesh, why don't we just rename "AskSlashdot" to "AskGoogle" and be done with it? These things are common as muck; single port D-Link's can be had from as little as #30 in the UK if you shop around, and they even do wireless models. Intel's Netport range is also very good (I even have one at home), they are more expensive than the D-Links, but have much broader protocol support. Since you need to talk to non-Windows systems, you'll need to ensure that your device supports a standard like LPD that you can actually talk too (unless you want to install Samba).

...why can't I buy one at Dell?

don't forget that DELL sells Intel only!

OTOH, consider other (possibly more immediately) beneficial programs that use distributed computing. Cancer and Alzheimer's genetic research, for example, can be found through Philanthropic Peer-to-Peer. You donate your cycles, pay some extra for your electric bill, and let them have some computing power that they may not be able to afford directly.

FWIW, Centrino is not a processor, its a mobile technology "suite", I guess you could call it. Its a bundling of the MB, the Pentium M processor, wireless networking, etc.

FWIW, Centrino is not a processor, its a mobile technology "suite", I guess you could call it. Its a bundling of the MB, the Pentium M processor, wireless networking, etc.

Most of Intel's fabs are located in USA (with couple outside, notably in Israel and Ireland). So techncially speaking, they are mostly made in US.

They are *packaged* and *tested* in Malaysia, Costa Rica, and other countries, but the actual core of the chips are manufactured in US.

uhh... Intel processors are manufactured in the US (here).

Well, if you're going to be confrontational about it, let's go:

Feel free to quote meaningless news quips all you like as long as they imply whatever you want to imply.

Me? I'll take some facts and numbers out of official specification documents: The G5 runs between 19 watts of heat dissipation and 42 watts, depending on its clock speed (source as HTML, as PDF)

The Pentium 4 desktop version runs between 60 and 80 watts (source). Indeed, the G5 in its desktop version is in fact competitive with Intel's low power Mobile Pentium 4-M processor for notebooks, which ranges from an impressive 7 watts up to 35 watts (source).

Put that in your pipe and smoke it.

P.S.
This is the direction in which intel is headed with its Mobile Pentium 4 (not M) processor. 70 watts in a laptop.

Well, if you're going to be confrontational about it, let's go:

Feel free to quote meaningless news quips all you like as long as they imply whatever you want to imply.

Me? I'll take some facts and numbers out of official specification documents: The G5 runs between 19 watts of heat dissipation and 42 watts, depending on its clock speed (source as HTML, as PDF)

The Pentium 4 desktop version runs between 60 and 80 watts (source). Indeed, the G5 in its desktop version is in fact competitive with Intel's low power Mobile Pentium 4-M processor for notebooks, which ranges from an impressive 7 watts up to 35 watts (source).

Put that in your pipe and smoke it.

P.S.
This is the direction in which intel is headed with its Mobile Pentium 4 (not M) processor. 70 watts in a laptop.

sPh

This technology will become significantly easier to implement once Intel introduces hardware support for virtualization (called VT - Vanderpool technology).

IDF Demo

During the Fall 2003 IDF demo, Paul Otellini ran a demo with a Tivo-like app running the Simpsons on one virtual machine, concurrently with another virtual machine booting up windows-xp. One computer was running 2 operating systems and driving 2 monitors concurrently.

Right, big tech companies obviously avoid Linux -- say big tech companies like:

• IBM
• SAP
• HP
• Oracle
• Dell
• Intel
• AMD
• Sun

Or which big technology companies were you referrring to exactly? Sure, it was a troll, but hey -- who isn't supporting Linux that doesn't make a competing OS? I mean, heck Sun, HP and IBM do make competing OSes and they've all jumped on the Linux bandwaggon.

Intel's Fortran compiler for Linux is FREE for non-commerical use (not open-source). I'm told that it is very good. Its about time that there was a free F90/F95 compiler available. Intel F90 Compiler. I think that this will go a long way in promoting LINUX as a base for advanced scientific computing.

There is also a good deal of pdf manuals, and I'm sure that more hardcopy manuals would be avalible if I actualy cared to look.

You can't call the Pentium or Itanic open standards.

The opcodes are known and well documented. Multiple companies besides Intel has created their own reimplementation of the "x86 standard". Dozens more individual programmers have duplicated it in software.

It seems much like an open standard.

Heh.

Heh-heh.

So, Mr. Barrett, was the last time this sort of thing has been tried in a locale of 1.2 billion people?

Basically, Intel is scared. If this takes off, Intel will suffer dearly in a market that currently generates 3.2 billion dollars of revenue for them. Roughly 12% of their total revenue comes from China alone.

You'll forgive me, Mr. Barrett, if I have trouble keeping a straight face.

Intel has always pushed MS Windows. Don't you recall their alliance?

I bet that this can be replaced with a pentium 1 + ethernet card + Linux/BSD. It doesn't take a whole lot to be a file server.

Actually, that's an exact description of the "Storage Station" product that Intel sold a couple of years ago - embedded Pentium 266MHz processor, about 2MB of RAM, running BSD, with a 10/100 LAN connection and up to two 30GB hard drives. All in a case the size of a hardback dictionary. An absolutely superb product - mine's possibly the most useful piece of kit I have. Full details at this page here

Anyone know of a similarly-sized product anywhere (no, Mirra is too big - AFAICT it's just a normal ITX mini-tower PC)? My ideal spec these days would include gigabit LAN too..

It's kinda buried - you have to click the "list of evaulation editions" link. But it most certainly exists.

http://www.intel.com/software/products/compilers/f lin/noncom.htm

However, it's listed as non-commercial only - which means (extracting from the license agreement) "If you are using the Materials under the control of a Noncommercial-Use license, you as an individual may use the Materials for only personal, noncommercial and research purposes"

But to a commercial entity, the fees aren't really that steep anyway, and it is a Fortran 95 Compiler, with all kinds of high-performance features that will make you drool. (For example, it can use SIMD/SIMD-II/MMX parellelizing instructions for the loops in your code without any explicit parallelization instructions in the code.)

There's an equilivent version of the Intel C++ compiler too.

For one thing, I wonder what its physical external address bus looks like. Can it address more than 4GB of physical memory without paging?

The Athlon64 and Opteron can.

With the Prescott and this new extreme edition P4 with 2.5MB cache (I shudder at the yield hit that much cache has per wafer for them)...

AGP was necessary because the bottleneck was the PCI bus; 133MB/s bandwidth simply isn't enough to support faster graphics cards, especially if there is a network card on the same bus. Also, any PCI accesses to main memory have to go through the southbridge / IO Controller Hub of the chipset, then through the Northbridge / Memory Controller Hub. The main innovation of AGP, as you rightly state, was that it hung directly off the northbridge and therefore offered a direct path to main memory - critical when AGP was designed, because the idea was that NO-ONE could afford to put more than a few MB of memory on a graphics card. It's become less critical since then, as memory prices have dropped (when AGP was conceived, of course, the average PC had about 32MB memory.. now, graphics cards alone commonly have 256MB).
Because it operates using a totally different topology - not unlike a switched network, in fact, with each device being a "switch" - PCI Express will provide much higher bandwidth, and a high-speed link to main memory, meaning that this advantage of AGP (whether needed or not) is preserved.

In brief, PCI Express operates using high-speed serial wire pairs known as "lanes", that each offer 200MB/s of bandwidth. Lanes can be aggregated to boost this figure higher; an "x16" PCI Express slot has 16 of these lanes. PCI-Express also provides for enhancements deemed necessary in a modern PC environment, like QoS for example.

There's an excellent PCI information page on the Intel developer website, at this page.

Good god, is it really that hard to search for pci express at google.com?

From an Intel developer network page.

Desktop Platforms with PCI Express Architecture will be designed to deliver highest performance in video, graphics, multimedia and other sophisticated applications. PCI Express architecture provides a high performance graphics infrastructure for Desktop Platforms doubling the capability of existing AGP8x designs with transfer rates of 4.0 Gigabytes per second over a x16 PCI Express lane for graphics controllers.

There is even a link to a PDF detailing why PCI Express will be the next choice for graphics.

You mention some good points, but there are many companies making a profit off of open source software other than Linux. MySQL, IBM, Trolltech, Intel, and Dell are just a few (yes, I realize some of these companies make money in other ways as well, but they all report their open source-related activities to be profitable).