However, there was a desktop "engineering prototype" that was kind of quiet.
No picture was shown. It sounds like they
made a multichip module from 2 die to test things.
If they had a dual-core Xeon, they would have
said so clearly and not just mentioned after
a demo that the machine was using an
engineering prototype dual-core.
But a dual-core Itanium with 24 MB of on
chip cache is much harder to make than
an Opteron with 2 MB of cache.
AMD will pass through 10,000 chips/month well
ahead of Intel.
That graph is out of date. For more current info
check out the SPECmine top 20.
Opteron has seen clock speed increases since then
and compiler improvements so now is faster than
Itanium on SpecInt and not too far behind on
floating point.
The chip only has one crossbar, one memory controller,
one set of 3 hypertransports, and one set of
pins/package. Only the caches and the heart
of the CPU are duplicated. So it is much
less than double the size.
The other big win is now one motherboard
with 4 sockets handles 8 cores. So you
buy half the motherboards, half the racks,
and pay half the co-location fee.
Here are some
interesting numbers on Itanic and Opteron sales.
Total Itanic system revenue was $319 mil in Q2,
up from $287 in Q1 for 11% increase.
Total servers at 5,665 in Q2 compared to 2,717
in 2003-Q2.
Opteron servers total
60,000 and generated $191m in revenue, according to Gartner. Previously reported that Opteron
increased by 81% from Q1 to Q2. This 60,000 compares to just 2,735 shipments and $8m in revenue in the same quarter last year.
At this rate the dollars should be about
equal in Q3, though many more Opterons since
they are cheaper. After that Opteron should be
way ahead.
The server can be equipped with up to 64GB of DDR (Double Data Rate) memory, two CD-ROM drives and eight Serial ATA hard-disk drives, according to Iwill.
"This is a very cost-effective way of building a high-performance server," Lin said, adding that Iwill expects to see the eight-way Opteron server be competitive against more expensive servers based on Intel Corp.'s 64-bit Itanium 2 processor.
Potential customers can obtain samples of the eight-way server now, and the server will be in volume production at the end of the year, Lin said.
But Opteron doesn't really compete with Itanium on the high end. Itanium2 has far more advanced RAS and enterprise features, and is a faster chip.
At the moment you can buy larger Itanium
systems than current Opteron systems, so
Itanium has retreated to the one place it
currently beats Opteron. It might take
a year or even two, but Opteron will get to
large systems too.
As for reliability, I don't agree that Itanium
has an advantage. It is a huge chip, with
many transistors, that runs hot, on a large
module with many extra connections. This is not
a recipe for reliability. It could get
better at 90 nm or 65 nm.
Also, you can buy twice the number of Opterons
and do a Tandem style system with
more reliability for the same money.
Opteron competes squarely with the Xeon. And while AMD is definitely gaining there, Xeon still outsells it by huge amounts.
Opterons CAN NOT ADDRESS more than 1TB. A 2-way Opteron system CAN NOT ADDRESS more than 1TB, a thousand-way Opteron system CAN NOT ADDRESS more than 1TB. Got it?
I have never said that the current Opteron or
the current Hypertransport-1 does more than
40-bits of live external address. Internally
the registers and code are 64 bit.
And if you think hypertransport2 having 64-bit addresses means "Opterons will clearly be going to 64-bit addresses" then...
If they are using 64-bit addresses in the
overall system, and could address a full
2^64 address space worth of memory with
enough CPUs, what would you call it?
Once they have Hypertransport-2 the
large Opteron systems will be able to
use as much memory as any SGI system.
Lastly, it isn't just a matter of "putting power on more address pins". The entire MMU and memory controller will have to be redesigned.
It depends on the initial design, but maybe
it does take some touch up.
I don't think there is any need
for more than 1 TB of memory attached to
each CPU for the next 10 to 15 years. So
they should have plenty of time to work on it.
10) Compatibility mode is so slow people say it is non-compatible
Yeah it is pretty slow.
At least we can agree on something.
8) Alpha development halted for years before Itanic caught up
And the Alpha development team is now working for Intel. Let's see what they've done with IA64 next year.
The magic of the Alpha team was in coming up with
a great Instruction Set Architecture. Given
how amazingly well the Xeon does against RISC
chips it is clear that Intel has the ability
to make good chips even with a not-so-good
architecture. The AMD guys had it right when
they got some of that team to help with the
initial X86-64 architecture. Now that Itanium's
architecture is set in stone I don't see
a few guys from Alpha being able to do
anything Intel could not already do (again Xeon
shows Intel has talent).
7) Microsoft calls AMD64 by the name x64
You smart computers good.
I can not make sense of that comment.
6) Code-size is twice x64, so needs 2x cache and 2x memory bandwidth
You do realise that code bandwidth and cache usage is usually orders of magnitude less than data bandwidth and cache usage?
No, I don't think that is true (URLs please).
Also, the same designers that thought code size
did not matter designed their data formats.
For example, they can not sign extend a 16-bit
value to 64-bit, so you always have to have
64-bits. So like the code segments, Itanic's
data segments are also often double the
size of those for an AMD64.
5) Fewer software tool vendors support it now than a year ago
One notable tools vendor - Parametric Technology Corp. - pulled back its support for Itanium this year, and the Northrop Grumman customer said at least three other tools vendors, including Synopsys, have reneged on their support as well.
This is a big clue.
4) Even before Win-x64 is released, MS supports more products on x64
The intersection of people who care about MS, and people who care about Itanium is unsurprisingly, probably quite small.
The set of people who care about Itanium
is quite small, so any intersection of this
set with any other will be quite small.
3) Don't know anyone buying one, just someone who got one for free
When did you last leave your parents' house?
I own my own place on a tropical island in
the caribbean, thank you very much.
With a population of 12,000 here I know of only
one person who owns an Itanic. He
more or less got it for free from HP since
he has a product that HP would like to see
ported. With 100,000 sales in 2003, is seems
like less than 1 Itanic per 2,000 people
there, so I suspect others are in the same
situation I am in. Do you know anyone who
paid for one?
2) Marketing keeps telling people to buy the not-yet-released version
Wha?
Intel keeps promoting the next chip. You
have even done it in this thread where you
say that the coming 1.7 Ghz 9MB chip will
be fast. As long as this goes on people
keep waiting for the coming stuff, and not
buying the current stuff. It is a classic
way high tech companies that get behind in
the tech fail.
Also note that 1.7 Ghz is
not really much more than 1.5 that has been
out for a year. In fact, AMD will probably
increase clocks from 2.4 to 2.6
in the next couple months.
10) Compatibility mode is so slow people say it is non-compatible
9) Sales can generously be described as "flat"
8) Alpha development halted for years before Itanic caught up
7) Microsoft calls AMD64 by the name x64
6) Code-size is twice x64, so needs 2x cache and 2x memory bandwidth
5) Fewer software tool vendors support it now than a year ago
4) Even before Win-x64 is released, MS supports more products on x64
3) Don't know anyone buying one, just someone who got one for free
2) Marketing keeps telling people to buy the not-yet-released version
1) People call it Itanic
In addition, 64-bit-capable x86-64 servers continued to ramp in volume with server units based on AMD Opteron processors growing 81.1 percent sequentially in Q2.
Itanic shipments are not going up like this.
Opterons units/month are probably more than
Itanic units/year at this point and Opteron
is growing far faster. So if SGI wants
to make money, switching to a CPU that people
really like could help.
The
new hypertransport-2
has 64 bit addresses. So Opterons will clearly
be going to 64-bit addresses at the system level, even if they each
only have 40-bits of live address pins for their
local memory. But if people get close to
putting 1 TB per CPU, you can be sure they
will put power on more address pins in future chips.
A couple months after AMD said they had taped out their
dual core Opteron, Cray and others said they would be
upgrading to that in 2005. Putting two Opterons next
to each other with their hypertransports talking to
each other is so easy that I suspect AMD's first silicon worked.
I suspect dual-core Opterons will be in production way ahead of dual-core Itanics,
since they are so much smaller. And we will see 8-core Opterons before
we see 4-core Itanics. A number of
people have
working 8-way Opteron motherboards and I have
not heard of anyone getting more than 2 Itanics
on a motherboard.
All the while AMD production volumes will be far higher
than Itanic volumes.
If SGI wants to stop loosing money, they should come out with an Opteron CPU-brick fast.
"Won't be safe for long" means
Itanic is sinking. Really.
The Opteron can't even physically address 1/8th the amount of memory that SGI has in their biggest systems.
1 TB is more than most people need this year,
and next years Opterons could extend this.
Also it isn't very impressive that the Opteron just barely beats a chip released years ago in some areas and gets trounced by it in others.
At spec.org Opteron is up with recent Itaniums
at costing 3 times as much on specInt and not far
behind on FP. Again, a 4-way Opteron
motherboard is probably cheaper than a 2-way
Itanium. If you compare equal investments
in money, Opteron is very impressive.
The Itanium can scale up to 8, 16, 32 or even 64 processor modules...
A new
chipset for 32-way opterons
should be out by the end of the year.
So yes, at the moment Itanium has retreated
to multiprocessors where the only comperable Opteron
system is the
Cray XD1.
But it won't be safe for long.
Nonetheless, Itanium2 still has a FP performance superior to Opteron.
Only in FP-Performance/CPU, and not by that
much. By most
other measures Opteron wins. If you look at
FPP/$, FPP/watt, FP/motherboard, FP/cubic-foot
Opteron wins. Opteron has 4-way on a motherboard,
and moving to 8 and 16-way on a motherboard
while Itanic is looking at 2 to 4-way.
In the end price/performance is
far more important that absolute performance
per CPU.
Intel admitted the iAPX 432 was a failure and killed it. However, they have not yet admitted the ia64 is, so the blood will keep draining for awhile yet. Sometimes a clear failure hurts less than a not so clear failure.
Another thing is they seem to have reduced
development of the x86 since they thought
Itanic was working. They put their best
guys on the Itanic and just sort of tweeked
the P4 for years. This let AMD pass them.
Intel could loose many billions because of
this. Intel's stock is already down tens
of billions this year.
The iAPX 432 was a complete and utter failure; by comparison, ia64 has been a success. Ia64 *is* seeing design wins, and does have market niches where it's a good choice. The iAPX 432 was just a dog.
Intel is said to have spent $5 billion on
Itanic. They sold 100,000 chips in 2003.
If you say $3,000/chip that is $300 mil.
I think it is safe to say that Intel has
lost more money on ia64 than they did on
iAPX 432. And the percentage of x86 that
moved to ia64 is also about 0.
Intel admitted the iAPX 432 was a failure
and killed it. However, they have not yet
admitted the ia64 is, so the blood
will keep draining for awhile yet. Sometimes
a clear failure hurts less than a not so
clear failure.
Those offerings from Rocketcalc are dual Operon boards with orginary clustering. The only difference is, they put them in one case.
Ouch. I did not realize that. Sun's 4-way has
different cards that are connected together with
hypertransport links and I thought these guys
were doing the same thing in the backplane.
I have seen a 4-way Opteron motherboard
(I think it was Tyan) for about $1,000. Getting
this, CPUs, etc would not be all that expensive.
So the $6,000 price seemed plausible.
The AMD multiprocessor plan seems like an Itanic killer.
History will probably record
Itanium
as Intel's VLIW attempt to break away from x86.
Very similar to their failed
IAPX_432 object oriented attempt and their
i860 RISC attempt to move customers away from x86.
Each time Intel seemed to use the architectural fad of the
day to try to move people off x86 but without
good backward compatibility. Each time seems
to have failed badly.
Slashdot Mirror
That board does look nice. There is a new 366 Mhz Geode as seen on this 5 watt board.
Ok, try this picture of yohan.
However, there was a desktop "engineering prototype" that was kind of quiet. No picture was shown. It sounds like they made a multichip module from 2 die to test things. If they had a dual-core Xeon, they would have said so clearly and not just mentioned after a demo that the machine was using an engineering prototype dual-core.
But a dual-core Itanium with 24 MB of on chip cache is much harder to make than an Opteron with 2 MB of cache. AMD will pass through 10,000 chips/month well ahead of Intel.
The other big win is now one motherboard with 4 sockets handles 8 cores. So you buy half the motherboards, half the racks, and pay half the co-location fee.
Somehow a Slashdot thread on Itanium and Opteron did not get into the Intel section.
Note that Intel has shown a wafer of dual-core Itanics but has not yet shown a working one. Surely they would rather have shown a working one if they could.
Here are some interesting numbers on Itanic and Opteron sales. Total Itanic system revenue was $319 mil in Q2, up from $287 in Q1 for 11% increase. Total servers at 5,665 in Q2 compared to 2,717 in 2003-Q2. Opteron servers total 60,000 and generated $191m in revenue, according to Gartner. Previously reported that Opteron increased by 81% from Q1 to Q2. This 60,000 compares to just 2,735 shipments and $8m in revenue in the same quarter last year. At this rate the dollars should be about equal in Q3, though many more Opterons since they are cheaper. After that Opteron should be way ahead.
As for reliability, I don't agree that Itanium has an advantage. It is a huge chip, with many transistors, that runs hot, on a large module with many extra connections. This is not a recipe for reliability. It could get better at 90 nm or 65 nm.
Also, you can buy twice the number of Opterons and do a Tandem style system with more reliability for the same money.
I agree with this.But we are talking about server things like Exchange Server, Commerce Server, Host Integration Server. Again, a big clue.
I own my own place on a tropical island in the caribbean, thank you very much. With a population of 12,000 here I know of only one person who owns an Itanic. He more or less got it for free from HP since he has a product that HP would like to see ported. With 100,000 sales in 2003, is seems like less than 1 Itanic per 2,000 people there, so I suspect others are in the same situation I am in. Do you know anyone who paid for one? Intel keeps promoting the next chip. You have even done it in this thread where you say that the coming 1.7 Ghz 9MB chip will be fast. As long as this goes on people keep waiting for the coming stuff, and not buying the current stuff. It is a classic way high tech companies that get behind in the tech fail.Also note that 1.7 Ghz is not really much more than 1.5 that has been out for a year. In fact, AMD will probably increase clocks from 2.4 to 2.6 in the next couple months.
And sometimes the truth is funny too.10) Compatibility mode is so slow people say it is non-compatible
9) Sales can generously be described as "flat"
8) Alpha development halted for years before Itanic caught up
7) Microsoft calls AMD64 by the name x64
6) Code-size is twice x64, so needs 2x cache and 2x memory bandwidth
5) Fewer software tool vendors support it now than a year ago
4) Even before Win-x64 is released, MS supports more products on x64
3) Don't know anyone buying one, just someone who got one for free
2) Marketing keeps telling people to buy the not-yet-released version
1) People call it Itanic
The new hypertransport-2 has 64 bit addresses. So Opterons will clearly be going to 64-bit addresses at the system level, even if they each only have 40-bits of live address pins for their local memory. But if people get close to putting 1 TB per CPU, you can be sure they will put power on more address pins in future chips.
A couple months after AMD said they had taped out their dual core Opteron, Cray and others said they would be upgrading to that in 2005. Putting two Opterons next to each other with their hypertransports talking to each other is so easy that I suspect AMD's first silicon worked. I suspect dual-core Opterons will be in production way ahead of dual-core Itanics, since they are so much smaller. And we will see 8-core Opterons before we see 4-core Itanics. A number of people have working 8-way Opteron motherboards and I have not heard of anyone getting more than 2 Itanics on a motherboard. All the while AMD production volumes will be far higher than Itanic volumes.
If SGI wants to stop loosing money, they should come out with an Opteron CPU-brick fast.
"Won't be safe for long" means Itanic is sinking. Really.
Intel admitted the iAPX 432 was a failure and killed it. However, they have not yet admitted the ia64 is, so the blood will keep draining for awhile yet. Sometimes a clear failure hurts less than a not so clear failure.
I have seen a 4-way Opteron motherboard (I think it was Tyan) for about $1,000. Getting this, CPUs, etc would not be all that expensive. So the $6,000 price seemed plausible.