[NTLUG:Discuss] Alphaserver 2100 -- Digital Archive

[comcast]

Wow, thanks for the in-depth info :)
Collin

-----Original Message-----
From: discuss-bounces at ntlug.org [mailto:discuss-bounces at ntlug.org] On Behalf
Of Bryan J. Smith
Sent: Friday, May 21, 2004 2:03 PM
To: discuss at ntlug.org
Subject: [NTLUG:Discuss] Alphaserver 2100 -- Digital Archive

Comcast wrote:  
> I just purchased a Compaq Alphaserver 2100 and was wondering if
> anybody has any experience with these, or basically... any advice as
> to what to do with it.  From what I've gathered it's a 4 processor DEC
> Alpha based system with 275MHz CPU's.  I've some some various reports
> of running Redhat 7.1/2 on these machines, but the majority of OS's
> that are reported as running are Digital Unix, Tru64 Unix. FreeBSD and
> OpenVMS.  I've never owned a server before and any advice or knowledge
> that anybody has would be greatly appreciated.

It's an original EV4 (?) 21064 system, although I believe it will take
the entry-level EV5 21164 300MHz chip as well.  21164 is somewhat faster
than the 21064 MHz for MHz (and 21264 even more so beyond the 21164).

Here's the best link I found:  
  http://h18002.www1.hp.com/AlphaServer/archive/2100/2100_tech.html  

Digital used to maintain the _absolute_best_ technical documentation of
_any_ OEM.  They kept _everything_ on-line.

Chris Cox:  
> I have a cheaper Alphaserver (1 processor) running Debian. That
> worked fine for me (but it's pretty old Debian and needs an update).
> I had to pull the non-functional video card and replace it with
> a Mach 64 based card.. that worked fine.  Had to boot initially with
> MILO on a floppy and then convert things to boot off the hard drive.
> Unit is designed (mine anyway) for Windows NT ONLY.

There are two firmwares available.  Some Digital systems _only_ support
one.  Others, largely OEM/whitebox 21064/21164/21264 mainboards support
_either_ for maximum OS compatibility.

 ARC:  Windows NT (and 2000, never relased publicly) and Linux
 SRM:  Digital UNIX (aka Tru64), OpenVMS and Linux

Linux supports either firmware type, although you must use the correct
bootloader to support each.

[ Correct any improper assumptions of mine below ]

- Alpha:  A Testament

Alpha was the _best_ ever engineered instruction set architecture
(ISA).  Digital didn't invent RISC, they just made it anal.   Digital
also showed that it was far better to use binary emulation/translation
than to put backward compatibility in hardware.  The only catch with
Digital's binary approach was that you had to run the same OS.  E.g.,
NT/Alpha could only run NT/x86, VMS/Alpha could only run VMS/VAX,
Linux/Alpha could only run Linux/x86, etc...

Unfortunately for Intel, they didn't realize this until their IA-64
design _failed_ to deliver on its promises of explicitly parallel
instruction computation (EPIC) and branch predication.  IA-64 "Madison"
(Itanium III) is basically IA-64 retro-fitted with Alpha 21164
technology.  This includes Digital's binary technology which causes the
IA-64 to run x86 _faster_ than via its hardware instructions.

- Once there was Standford MIPS and Berkeley RISC (soon-to-be SPARC)

In the late '80s, there was a standardization effort to move to MIPS,
one of the two original RISC approaches (actually came after Berekley
RISC).  MIPS was particularly sweet in that it was bi-endian (Berkeley
RISC was big endian only), and could support just about any OS.  It was
a true 32-bit architecture, and many UNIX vendors were standardizing on
the R2000 and, later, R3000 for their workstations/servers.

Unfortunately, special interest caused the MIPS standardization efforts
of the late '80s to break down.  Sun continued with its SPARC, IBM and
Motorola started to work on the PowerPC variant of the Power and Digital
was already working on their VAX replacement chip.  Digital wanted a
design with a 25-year lifespan.  They designed it to be 64-bit from the
get go, and have 0 support for _any_ 8 or 16-bit instructions (32/64-bit
only).  They invited Intel at one point, but Intel pulled out, taking a
_lot_ of trade secrets with them.  This resulted in a lawsuit that came
later in the mid-'90s.

There are several different Alpha processor revisions:  
 21064
 21164
 21164A
 21164PC
 21264
 21264PC
 21364

The original 21064 came out in the early '90s.  All Alpha processors
_lack_ 8 or 16-bit instructions.  They are a "pure/anal" 32/64-bit
microprocessor.  Digital left out even 8 or 16-bit load/store opcodes
until the 21164 after enough people complained always loading 32/64-bit
was inefficient.

- EV56:  The 500MHz Bomb

The 21164A EV56 was probably the most well known.  Released in 1996, it
was the first microprocessor to reach 500MHz at air cooled.  And that
was at a fab level of 0.35um micron, whereas Intel couldn't not get its
Pentium Pro to get past 200MHz at 0.25um micron.

The 21164PC was an attempt to cut costs and gather OEM support.  The L2
cache was dropped from the 21164A, and five (5) multimedia instructions
were added.  Yes, _only_ 5.  Digital's attitude on the Alpha was that
_no_ instruction should _ever_ hold up execution in a pipe, and anything
complex enough could.  A very anal, but powerful RISC approach IMHO.

- Intel's fatal error:  Had the chance to grab Alpha in 1997

Intel had the chance to go Alpha in 1997, and fix everything.  But no,
they were committed to their engineering EPIC/Predication ideal in
Itanic.  A brilliant and innovative design, IA-64 _failed_miserably_ in
silicon.  More on that below.

Digital failed to update its fabs -- easily 2-3 generations behind IBM
and Intel.  It quickly fell behind and Palmer started to sell off units,
including Digital's innovative Semiconductor division which basically
invented  _everything_ since the mid-'80s.  This included all major PCI
controller logic and bridges, as well as other I/O and system
interconnect "glue logic."

Palmer offered Intel everything -- again, Intel _could_ have gone Alpha
for 64-bit and we wouldn't be talking about Athlon64/Opteron today. 
Digital's FX!32 would handle all the major NT/x86 compatibility issues,
and NT 5.0 ran fully 64-bit with FX!32 integrated on the Alpha
_well_before_ it became Windows 2000 and 32-bit only (who do you think
wrote most of 64-bit NT?  That's a long story in itself too!  ;-).

Intel bought Digital's fabs as part of the 1997 lawsuit settlement, and
eventually Alpha itself not much after that.  AMD, Samsung and one other
licensee was required for the deal to go through by the FTC.  It didn't
matter.  Most Digital engineers either went to AMD, or started Alpha
Processor, Inc. (API), which eventually became API Networks.  API
Networks was basically absorbed by AMD recently.

- EV6 and Beyond:  Bad for Alpha, Good for AMD

The Alpha 21264 EV6 was way late.  But its Slot-1/2 interconnect, more
commonly found in the Socket-462 Athlon32 today, became commonplace. 
It's a very I/O throughput switching interconnect.  It was the precursor
to API Network's Lightspeed Data Transport (LDT), now more popularly
known as AMD's HyperTransport which is used in Socket-754/940
Athlon64/Opteron, as well as Apple G5/IBM PowerPC 970 and other
chips/chipsets (even for Intel mainboards).

I believe the Alpha 21364 is in sampling now.  It's largely just a
compatibility chip as HP-Compaq-Intel works to get people over to IA-64
for OpenVMS.  OpenVMS will also use IA-64 (Itanium) now, using binary
emulation/translation to run on the Itanium.**


-- 
Bryan J. Smith, E.I. -- b.j.smith at ieee.org



_______________________________________________
https://ntlug.org/mailman/listinfo/discuss