AMD [IBM] press release

Mark Hahn hahn at
Wed Nov 20 11:23:46 PST 2002

>    Judging by the remarks of Dr. Mark Hahn, I was not successful

thanks for the honorary doctorate ;)

>   > it has nothing to do with "complete".  this category of hardware is 
>   > quite different from PCs, not just in price, but in reliability
>   > and design.
>    This is one point in which we clearly have a difference of
> opinion concerning the computer market in general.  During several
> years of asking for price quotes, it has been my impression that a
> high-performance "server" from Compaq or HP (when they were separate
> companies) had a price that was about twice what I would need to pay
> to get an equivalent computer assembled locally, at the same level
> of performance and sometimes with the same motherboard.

hey, I prefer building my own machines too, but there are rational 
reasons that people pay the premium, for brand name boxes.
for instance, for grant-related reasons, my desktop is a Compaq
P4/1.7G.  while I was pissed that someone choose this for me without
asking, it is a very nice machine, which would not be doable as a 
whitebox PC.  not exactly.  for instance, it's quiet, and contains
some interesting airflow engineering.  the board is mundane, as are
all the components.  but I expect the PS and CPU fan will also last longer 
than a noname part from the corner computer store.  and it comes with 
a three-year warranty, which some people consider to be critical.

> commercial software that does not run efficiently on a cluster, so I
> was wondering if a 4-CPU shared memory PC would be a cost-effective
> choice for the group that ran this commercial software.  The Dell

depends on what the software is doing.  for instance, if it's very
dram intensive, it'll run PAINFULLY badly.  if it's cache-friendly,
it'll fly.

> Poweredge 6600 has a cost of $10,288 for a two-processor version with
> a 2-GHz Xeon CPU, one GigaByte of memory, and one 18 GB Ultra 160
> SCSI harddisk.  But wouldn't I be able to get a two-processor PC with
> this level of performance for about half the price?  (Perhaps even
> with two-way interleaved memory.)

and a PE 2600 with 2x2.4/1G/18G would be $4k.  people often don't realize
how amazingly expensive chassis are.  for a cluster I just ordered,
we're paying something like 30% of the per-node price for the chassis!

using rock-solid white-box parts, the aforementioned cluster will have 
2x2.4G xeon nodes for about $1800 apiece.

>    My impression is that the price of a Dell server is at the low end
> of the Xeon-based servers from various companies.

I don't think so.  Dell's pricing structure is more up-front than
the traditional vendors like HP/Sun/IBM.  if you wave a Dell quote 
at your HP rep, he'll probably come close to matching the price.

>    Overall, the question that I intended to put forward in my previous
> email message was primarily based on the Itanium / Hammer level of CPU
> performance.  It was my impression that by this date Itanium was supposed
> to arrive at the desktop.  In fact it has, since one group here has
> a slick-looking Itanium workstation on loan from HP.  But at what cost?

I don't believe Itanium is close to an appropriate price-performance point
for anything except very unusual applications.  current IT2 machines are 
not terribly impressive CPU-wise, but definitely deliver a lot of bandwidth
(6.4 GB/s dram bandwdith as I recall, with extremely wide onchip buses.)
if that floats your boat, great.  they're too hot and expensive for me,
and I fully expect Opteron systems (with 6.4 GB/s dram *PER*PROCESSOR*)
to be a LOT cheaper.  if I can get a dual-xeon e7500 board for <$400 today,
I expect a much nicer dual-Opteron board to be <$800 in 6 months.

I think everyone is still guessing about prices for Opteron chips themselves;
the core is supposed to be approximately a thoroughbred, and the HT and dram
interfaces are not terribly complex.  if they ship with 1M cache, that'll
definitely drive up the price, though not all that much.  I guess Opteron
will cost $600 or so shortly after intro, which isn't bad.

> the Beowulf group if they knew with the AMD CPUs would have a price
> and availability similar to Itanium 2 initially, for example, in the
> first half of 2003.  Would AMD be thinking, if Intel can sell their
> Itanium 2 at the stratospheric price of a Power4, maybe we should try
> starting at that price level? (???)  

they can't afford to, IMO.  AMD needs to quickly start shipping lots of K8's.
yes, they also need a higher ASP than current K7's, but that's a given,
and not that hard.

AMD has had a whole year to ramp up K8 production; if they don't ship,
say, 4M K8's in 1h2003, the company is gone.

>    I wrote:
>   > Tyan and most of the motherboards on the Supermicro WWW pages.  But on
>   > those pages I do not see any Itanium 2 mother boards.  For low-cost
> and Dr. Hahn replied:
>   > of course: only HP and Intel make It2 boards.
>    I regret that my intended meaning was not clear.  If Intel at 64-bit
> is not competing at the level of having mother boards from various
> manufactures, why would the situation be different for AMD?  (Despite
> the announcements.)

afaikt, IT2 is big and hot and expensive to produce.  that drastically
limits both its appeal in the market, and its appeal to Intel as a way
to use fab capacity.  suppose (WAG) that Intel can churn out 4x as many
P4/2.6's as they can It2's on the same wafer/fab/linewidth.  do they 
have excess capacity to dedicate to It2's?  are they worried about AMD
once again taking market share in the bread-and-butter desktop market?

I think of the It2 as a placeholder - not especially attractive to buy,
not wonderful to produce, but since it does well on spec, it's useful
for Intel marketing.  the fact that hardly anyone buys them hasn't 
hurt Ferrari, has it?

>    This is very useful information.  I have the impression that there
> is substantial uncertainty with regard to the timing of the availability
> of the AMD processors and motherboards.

that's a VERY nice understatement.  IMO, AMD has completely botched
their job of evangelizing Hammer.  Intel, at least, has managed to 
not completely lose ia64 credibility, in spite of being many years late.

> Would anyone else on the
> beowulf mailing list have similar information about how concrete
> are the announcements?

AMD seems to confuse teasing with marketing.  afaik, there are no useful
non-NDA facts available.  AMD's official statements are not trustworthy,
since they so blithely stretch their roadmap.

regards, mark hahn.

More information about the Beowulf mailing list