Slashdot Mirror
Los Alamos to Use AMD's Opteron in Linux Clusters
nuke-alwin writes "eWeek is reporting that Los Alamos National Laboratory announced it will use more than 3,300 Opteron chips in two of its Linux clusters. According to the article 'The key to Opteron, as it tries to gain traction not only against Intel Corp.'s 64-bit Itanium chip but also its 32-bit Xeon offerings, is its ability to run both 32-bit and 64-bit applications equally well.'"
Intel can't compete with the Opteron on merits alone. It will be interesting to see what they try next.
IBM did that.
We were all warned a long time ago that MS products sucked, remember the Magic 8 Ball said, "Outlook not so good"
Still plenty of floor space in the new building.
Los Alamos to get Lightning computer system from Linux Networx
LANL
For more specific technical detail on the supercomputer Linux Networx is building for Los Alamos, go to
LNXI Newsroom
Quidquid latine dictum sit altum viditur
Microsoft pre-beta Windows for AMD64 available(to resellers)
Opteron 246 Overclocking Benchmarks
Whenever the offence inspires less horror than the punishment, the rigour of penal law is obliged to give way...
Alas that most websites don't set their servers up to survive a slashdotting. Perhaps the web would be a better place if they did.
Yeah, and you also need to reboot your Pentiums to run 16-bit code.
It is just the state of a flag in a control register. In particular, see page 68 of
AMD's Opteron System Programming Guide.
64-bit mode is enabled with the flip of bit 8 of the EFER Model-Specific Register. Otherwise it defaults to 32-bit mode. OS designers should test/set this bit just before running a thread in the scheduler, or jumping into system code as it can only be modified by code running in ring 0. This is the same way people treated the Virtual-8086 (16-bit) mode bit in CR0. In fact, you can combine the protected-mode, virtual-8086 mode, and "long mode" bits to have a variety of register-size and memory addressing modes per thread.
Fuck Beta. Fuck Dice
This is already possible gcc -m32|-m64; granted both the 32-bit and 64-bit libraries need to be available. 32/64-bit compatibility is supported in SuSE Linux Enterprise Server 8 (SLES8); and will be available in RHEL 3.0.
Debian developers are also working on a port to Opteron that does/will support 64 bit Kernel with 32-bit and 64-bit compatibility.
" It is well-known that the Opteron chips and mobos are manufactured in mainland China"
Not according to AMD: AMD Opteron(TM) Processor
All Opterons are made in Germany.
The only two fabs they they have are Dresden, Germany(fab 30) and Austin, Texas(fab 25).
Whenever the offence inspires less horror than the punishment, the rigour of penal law is obliged to give way...
You need the 64 bit OS, as you might expect, because allowing an application to address a 64 bit memory space while the OS is only capable of managing a 32 bit one, well, wouldn't work out very well. Hint: It's really tough returning from system calls made from the upper few terabytes of memory.
So, a 32 bit OS starts out "seeing" your standard 32 bit X86 CPU. This is the legacy mode, and looks like any 32 bit CPU out there today.
On Opteron, a 64 bit OS will also know how to set the new "Long Mode". Bzzt...A bunch of new instructions and handy registers magically appear.
From there, the OS can mark each code segment (your program) as being in "Compatibilty" (32 bit) or "64 bit" mode.
For more than you ever wanted to know...
Opteron Tech Manuals can be found here.
I saw quite a few of their new clusters ready to ship out. I had to constantly wipe the drool off my face while I watched them assemble 2GHZ dual Opteron boxes with 2GB RAM per processor. Their tech is impressive. They have their Ice Box control units and quite a bit of custom control/monitoring hardware that makes building your own cluster seem less advantageous.
They boast #3 on the worlds fastest super computers, so questions about Linux on the "Enterprise" should be easily resolved.
This Tom's hardware review of Opteron vs. Xeon is quite interesting to give a better feel for comparison to todays speeds.
It will be nice when we have some numbers to compare Itanium II direct 64 bit to Opteron, although it doesn't seem much can save the Itanic IMHO.
My Linux Command of the Day site : LCOD
FUD and FUD alone. I bought a dual Opteron and am using it right now.
:) It has a horribly slow software emulation layer to run 32 bit apps.
The Opteron's lower 32 bit registers look just like an Athlon's or P4 registers. When code wants to run in 32 bit mode, it just uses those registers. When the second process wants to run in 64 bit, it uses the lower 32 bit registers as well as the upper 32 new registers.
It's the Itanium chips (64 bit from Intel) that you are thinging of.
AMDs 32 bit implentation is in the silicon. Can't beat that.
Linux Networx is building a second 256 node (512 Opteron chips) cluster (LANL Orange) that will use the Infiniband interconnect from Mellanox technologies. LANL Orange will be the largest Infiniband cluster to date.
therefore: 2816 + 512 = 3328 Opteron chips
The CPU packaging is done in Malaysia or Singapore, the actual dies themselves are all fabbed in Dresden, Germany.
Excellent point; it is 100% true that AMD cooperated with the LinuxBIOS developers but one might also say that the real key is that Eric Biederman (an LNXI developer) was the work horse behind making LinuxBIOS on Opteron a reality.
BTW, most of the LinuxBIOS clusters that were mentioned in Ron Minnich's LinuxBIOS ML post are Linux Networx systems... emm gooood
DARPA started the internet, and DARPA is funded as congress damn well pleases, and funds can be earmarked for certain projects. So yes, Al gore did have a role in creating the internet.
Repeal the DMCA!
Sorry to be bringing up a MS product, but their new 64bit Windows will be able to run 32bit programs with in 64bit OS mode, but not 64bit programs in 32bit mode (at least from my current understanding of the new product line).
In Opteron, 64-bit apps cannot be run under a 32-bit OS. Opteron doesn't recognize code as being 64-bit code unless long-mode is enabled, and once long mode is enabled the OS must be 64-bit (because all switches to more privileged code also switch to 64-bit mode).
Being able to run 32-bit apps under a 64-bit OS was one of the absolutely required features for Opteron, however.
(yeah, I work at amd)
I am not a sig.
It's spelt out in full, page 225:
(...)
(emphasis mine)
So, in other words:
- SSE is twice as fast as single-precision x87
- their initial implementation of SSE2 is actually just an implementation of the instruction set, the actual execution units are borrowed from the x87 and have the same throughput limitations
- if you are writing really large code, you might get some benefits from SSE2 because of L1 savings (but your case has to be pretty borderline for that to make a significant difference, or you have to be able to interleave a significant amount of integer code with your FP code that using SSE2 frees enough decoder bandwith).
- In the future, they might bother with spending more space for more double-precision FP units (I'd think that if they can sort out the manufacturing problems, and the HPC segment seems like the area they're selling to, then it may make sense to spend a few mm to double SSE2 performance. Depends on who actually buys Opterons once the software/OS front lines are settled)
- Today's AMD64 SSE2 performance in double-precision is not going to be leaps away from their x87 stuff (unlike the P4), but you have to remember that AMD's x87 is really good.
- Don't forget that in long mode, you have 16 128-bit registers, which may help your compiler make more out of SSE2/64 code than x87 code; some benchmarks with SSE/SSE2 enabled compilers would help clarifiy whether that's the case or not.
So in short, yeah, they claim a somewhat faster SSE, but their SSE2 is mostly a "yep we can run that code too" item rather than "yep, we can run that code too, and faster to boot".Software failure: Slashdot's system posted my comment in the wrong position. Hopefully this will be posted as an answer to Jaysyn, under comment #6704329
My comment: That's right. In several email conversations, I questioned Vint Cerf about this, and he said that Al Gore was extremely important in making DarpaNet, a research tool at a U.S. government organization, into the Internet, a public utility available to all.
DARPA is the U.S. government's violence research department. DARPA is devoted to finding more efficient ways to kill people. In the beginning of networking computers together, there was no intention of benefiting anyone.
According to Mr. Cerf, Al Gore recognized the importance of a public computer network long before other public officials knew anything about computers, and made sure the public network had funding.
Some have called Mr. Cerf, "The Father of the Internet", but, as his biography says, many people were involved. Mr. Gore was the main promoter, "father", of the public utility we now call the Internet.
No, no. You are obviously missing some facts.
Al Gore lead the process of allowing commercial interests to use the Internet... Previously, it had been non-profit use only. In that process, Gore is quite correct that he helped to create the (commercial) internet, and more specifically, the internet we know today. Without that change, there would be no Amazon, no Yahoo, no slashdot, and you would not have discs mailed to you that allow you to quickly connect... You would have to arrange for access through a university, or someplace similar.
Al Gore's statements are quite correct, in context. It's only when taken out of context (which is what is happening constantly) that it becomes a gag.
Slashdot gets worse every day... Pipedot: News for nerds, without the corporate slant
See link here
This one is an IBM made one. Pretty interesting.
XML is like violence. If it doesn't solve the problem, use more.