On Friday 20 August 2004 02:41 pm, Paddy wrote:
> Bob Stia wrote:
> > Hello Sluggers,
> >
> > At risk of starting a hardware holy war, I turn to the Sluggers for
> > their advice and expertise. Finally convinced my wife that the old
> > computer is "old". Want to keep it under $1000 if possible.
> >
.....<snip a bunch>..........
>
> Here are two (2) emails on new boards/processors that may be of
> interest to you.
>
> To: pc_support@matrixlist.com
> From: Bryan J. Smith <b.j.smith@ieee.org>
> Subject: [PC_Support] New Socket-754/939 CPUs, Mainboards, nVidia
> NV45, ATI PCI-Express cards
>
> First off, AMD has rolled out some new Socket-754 and the first new
> Socket-939 Athlon64/FX processors. They vary in memory bus
> (Socket-754 = 1 DDR, Socket-939/940 = 2 DDR), L2 cache size (512KB
> and 1MB) and other details.
>
> Connections to memory in AMD64 are _direct_ and do _not_ go over the
> HyperTransport interconnect -- they are _extra_ wires from the CPU
> _itself_ to memory. That's why it's 754 pins for 1 DDR, and the 939
> pins for 2 DDR. You'll note that 939/940 - 184 (64-bit DDR) =
> 755/756. That's the difference in the pin count right there, the
> number of DDR channels (184 pins for each).
>
> Ziff-Davis/ExtremeTech articles:
> http://www.eweek.com/article2/0,1759,1604838,00.asp
> http://www.extremetech.com/article2/0,1558,1604604,00.asp
>
> AnandTech articles:
> http://www.anandtech.com/mb/showdoc.html?i=2063
> http://www.anandtech.com/cpu/showdoc.html?i=2065
>
> Including a new set of Socket-939 mainboard reviews:
> http://www.anandtech.com/mb/showdoc.html?i=2063
>
> As always, the FSB is no more with AMD64. It's fixed at typically
> 1600MT (mega transactions) over a bi-directional 32-bit (16-bit each
> way) bus -- aka "HyperTransport." This equates to a 6.4GBps
> throughput. Some enthusiast sites refer to this as a 800MHz DDR bus
> (1600MHz effective), which is the signaling. Some vendors use a
> 1200MT or 2000MT implementation, for 4.8GBps to 8.0GBps, which
> enthuasiasts then call a 600MHz DDR (1200MHz effective) or a 1000MHz
> DDR (2000MHz effective).
>
> Opteron 200 and 800 have multiple (2-3) 1600MT@32-bit HyperTransport
> links. These Athlon64/FX models only have 1 HyperTransport link (to
> I/O), as they are uniprocessor (single processor). Again, the memory
> bus is _separate_ from the HyperTransport links. The CPU has _both_.
>
> Now more on the video side, AnandTech has a couple of articles on the
> nVidia NV45:
> http://www.anandtech.com/showdoc.html?i=2064
> http://www.anandtech.com/video/showdoc.html?i=2067
> http://www.anandtech.com/showdoc.html?i=2068
>
> [ The last page of the first article has some of the first picoBTX
> mainboards, with a single PCI slot at the top (reversed orientation).
> ]
>
> There is also another article on the ATI PCI-Express cards:
> http://www.anandtech.com/video/showdoc.html?i=2066
>
> I just wanted to pass these links along for now. I'll get to reading
> and digesting them shortly.
>
>
> --
> Bryan J. Smith, E.I. -- b.j.smith@ieee.org
>
>
> To: pc_support@matrixlist.com
> From: Bryan J. Smith <b.j.smith@ieee.org>
> Subject: [PC_Support] Opteron/Xeon-based PCI-X Workstations/Servers
> -- WAS: PCI-Express vs PCI-X?
>
> George Laiacona wrote:
> > Bryan, In your opinion, who makes the best PCI-X mainboard?
>
> Depends.
>
> If it is a workstation and you need AGP, then a dual-Opteron with
> AMD8151 (AGP3.0), AMD8131 (PCI-X) or AMD8132 (PCI-X 2.0) and
> AMD8111 (PCI/LPC). Pick the latest Tyan or similiar mainboard,
> as long as it has the AMD8131 or AMD8132 on-board for PCI-X -- that's
> the _key_ (the "3" in the AMD chip 81x1). There are several versions
> of Tyan boards, so make sure it's the latest model (Thunder K8W
> "S2885" I believe?), or whatever your integrator (e.g., ASLab,
> Monarch, HP, etc...) sells.**
>
> [**NOTE: Because of the cooling/power requirements, as well as
> form-factor considerations -- e.g., EATX, WTX -- I recommend you use
> a system integrator and do _not_ attempt to build your own
> dual-Opteron 200, let alone dual-Opteron 800. ]
>
> If it is a server, and you can afford 4-way ($12K), I'd just go with
> the HP Proliant DL585. It has dual-AMD8131/8132 chips (plus the
> AMD8111) and is certified for Linux (only 32-bit at this time, but
> 64-bit shortly -- Linux/x86-64 _runs_ fine, it's just not certified
> by HP yet). That's the most powerful I/O I've _ever_ seen in a
> server, and many reviews have confirmed this.
>
> You can look to Xeon as well. Don't even consider Xeon if you want
> more than 4GB of RAM though, because 36-bit extensions are used to
> access more than 4GB on the Xeon. In old 32-bit terms, the
> performance is _literally_ like comparing using true 32-bit memory
> access (Opteron doing true 64-bit* addressing) versus 20-bit DOS plus
> EMS 64KB "paging" to access above 1MB upto 32MB (Xeon doing "paging"
> to access above 4GB upto 64GB).
>
> [*NOTE: Although the Hammer architecture is truly 64-bit (upto 18PiB
> (2^64 ~ 10^18) addressing), current Opterons actually do only 48-bit
> (upto 256TiB (2^48 ~ 10^12) adderssing). The 48-bit is for 32-bit
> compatibility with the 48-bit paging of the 486+ TLB. Future
> Opterons will break 32-bit paging compatibility to do greater than
> 256TiB. ]
>
> > Would I be able to find Linux drivers for the chipsets?
>
> You don't need to. Northbridge support is transparent, as all modern
> PC architectures comply with I2C/I2O, MP1.1/1.4, and any bridge
> interfaces look like PCI.
>
> Even though AMD64/HyperTransport offers switching and Non-Uniform
> Memory Architecture (NUMA), it was designed to be transparent. I.e.,
> if memory that one Opteron needs to access is not local, and
> connected to another Opteron, it is all transparent to the OS
>
> >From a performance standpoint, of course, this is not ideal. There
> > are
>
> NUMA "hints" in kernel 2.4 and NUMA "support" in kernel 2.6 to even
> improve the Opteron's performance.
>
> When it comes to Intel, same deal. The ServerWorks ServerSet logic
> complies with I2C/I2O, MP1.1/1.4 and all PCI bridges work
> transparently.
>
> The only "chipset support" necessary is for the "peripherials."
> E.g., NICs, ATA, etc... The NICs should be well-supported (high-end
> PCI-X), although the
> ATA may vary. ServerWorks (Intel) has had a bad history of poor
> chipset ATA design and performance. The AMD has been the exact
> opposite -- the same UltraATA mode 5/6 logic since the AMD75x seems
> to be in the AMD8111.
>
> I _highly_recommend_ the 3Ware Escalade 9000 series for SerialATA
> RAID. It has both 0 wait state SRAM and then lots of SDRAM. It
> requires a new driver, 3w-9xxx, which is now in the stock 2.6.6+
> kernels. 3Ware includes the GPL driver for 2.4 in the box as well.
> 3Ware has _always_ done an _excellent_job_ in making it _simple_ to
> build the 3w-xxxx (5000-8000 series)
> and 3w-9xxx (9000 series) driver with_out_ recompiling the entire
> kernel. Their Makefile is most excellent IMHO.
>
> > This is for my wish list. I've got room in my budget to build a
> > better system. But I need to move fast, I've only got a couple
> > more weeks to get the pieces ordered. I want to assemble this
> > rather than buy an off-the-shelf server somewhere.
>
> I don't know if I recommend that. If you want Quad-Opteron 800, the
> HP Proliant DL585 is worth the dough (very cost effective).
>
> Otherwise, consider ASLab and Monarch. They will sell you a system
> with little premium over parts, but make sure you get adequate
> cooling and power. Those Extended ATX and WTX Server form-factors
> can be tricky to deal with.
>
> But if you go on your own, I'd buy a "rolling" WTX Case (about $300)
> that also
> fits EATX/similar and a _quality_ 600W+ or dual-460W redundant WTX
> power supply
> that has ATX converters/support. That way, you're covered with
> _anything_ you
> buy or upgrade to in the future. [ FYI, BTX is single processor only
> ]
>
> Damien McKenna wrote:
> > http://www.apple.com/xserve/ ? :o)
>
> First off, while I believe the Linpack numbers of the PowerPC 970,
> I'd dispute those Linpack numbers between Xeon and Opteron. This is
> nothing new as Intel's optimizing compiler really helps its
> SPECint/fp and Linpack numbers (although the optimizations help
> Athlon/Opteron performance in general too), beating AMD 2:1 in many
> synthetic benchmarks.
>
> Real-world performance of actual applications always seems to show
> off the Athlon/Opteron. There are massive performance hits with
> branch prediction, the simplistic FPU and other details in P3+
> processors -- things that are noticable when you turn off SSE (and
> require it to do "lossless math").
>
> The Opteron with its larger L2 cache blows Xeon out of the water --
> unless they are using Xeon 4MB L2. L2 cache really makes a
> difference in many Linpack benchmarks, which means it's not so much
> of a CPU test. So that's my first guess.
>
> Other, select Linpack tests will really hit the CPU, and Opteron and
> PowerPC 970 kill. The Opteron has 2 complex (any) + 1 simple
> (ADD/MULT) FPU pipes, the PowerPC has 2 complex (any) CPU pipes and
> the Xeon has to choose between using only 1 complex (any) operation
> or 2 simple (ADD-only).
>
> Especially when you need _high_accuracy_ double-precision and you
> turn off SSE. The the Xeon's Pentium 4 core just withers. The old
> P3 is 2x the P4, MHz for MHz, so a 1.4GHz P3 is like a 2.8GHz at
> "raw" FPU. But even with SSE-2, Opteron holds its own with register
> renaming and out-of-order execution, using its 16 flat FPU/SSE ()
> registers and all 3 FPU pipes.
>
> I can only assume the Xeons were 4MB L2 and a L2-intensive benchmark
> was used. They might have ran 32-bit, Pentium-SSE-optimized Linux
> code as well, which would not give Opteron its full boost (128-bit
> 3DNow! with _direct_ SS)
>
> PowerPC 970, using its raw dual-pipe FPU and Altivec, is even better.
> Because it's optimized for its less legacy-driven core, unlike
> x86/x87. IBM is hoping they can take over the scientific computing
> space. Indeed this may be the case and more and more Apple G5s are
> ordered for such.
>
> Which makes me wonder, is IBM purposely holding off releasing "cheap"
> PowerPC 970 mainboard to OEMs in en-masse? Because these OEM/Linux
> PowerPC 970 solutions may compete with their own pServer and Apple
> G5s?
Paddy, Thanks for your interest. Looks kind of like overkill for me,
but will check it out carefully.
Bob S.
-----------------------------------------------------------------------
This list is provided as an unmoderated internet service by Networked
Knowledge Systems (NKS). Views and opinions expressed in messages
posted are those of the author and do not necessarily reflect the
official policy or position of NKS or any of its employees.
This archive was generated by hypermail 2.1.3 : Fri Aug 01 2014 - 15:05:27 EDT