General Schvantzkoph - future best CPU

[wizzywiz]

Hi General, I've been reading your posts on different groups about the
performance of current CPU chips, and you seem to be pretty
knowledgeable.

My current dilemma is shoud I buy now or wait? Currrently a Conroe is
probably the best bet (which your posts have helped convince me of),
even though I have large RAM requirements. I need fast handling of
single-threaded custom number crunching programs, with multi-gigabyte
datasets. Multiple CPUs aren't my solution since the amount or RAM
needed to make it possible to run multiple instances of our cruncher is
prohibitive.

I can hold off on buying as late as next May. What do you see coming
down the pike by then that may offer better performance in my
application?

Thanks...

 

[rohit.nadig]

I would hold off until May.

Intel's latest P6 architecture based chips are much faster than their
P4 architectures.

Conroe is going to be follwed by Penryn.

 

[David Kanter]

Hi General, I've been reading your posts on different groups about the
performance of current CPU chips, and you seem to be pretty
knowledgeable.

My current dilemma is shoud I buy now or wait? Currrently a Conroe is
probably the best bet (which your posts have helped convince me of),
even though I have large RAM requirements. I need fast handling of
single-threaded custom number crunching programs, with multi-gigabyte
datasets. Multiple CPUs aren't my solution since the amount or RAM
needed to make it possible to run multiple instances of our cruncher is
prohibitive.

I can hold off on buying as late as next May. What do you see coming
down the pike by then that may offer better performance in my
application?

How many GB of memory do you need?

DK

 

[General Schvantzkoph]

Hi General, I've been reading your posts on different groups about the
performance of current CPU chips, and you seem to be pretty
knowledgeable.

My current dilemma is shoud I buy now or wait? Currrently a Conroe is
probably the best bet (which your posts have helped convince me of),
even though I have large RAM requirements. I need fast handling of
single-threaded custom number crunching programs, with multi-gigabyte
datasets. Multiple CPUs aren't my solution since the amount or RAM
needed to make it possible to run multiple instances of our cruncher is
prohibitive.

I can hold off on buying as late as next May. What do you see coming
down the pike by then that may offer better performance in my
application?

Thanks...

The Conroe core is going to be the heart of Intel's CPUs for the next
couple of years, there will be tweaks and clock speed improvements in that
time but nothing that will make a big difference to single threaded
applications. Intel has a four core package coming in the next couple of
months but all it is, is two Conroe chips in the same package sharing the
same front side bus. The FSB is the weak point of Intel architectures.
It's adequate for a single dual core Conroe but I think it's likely to be
a bottleneck for four cores. AMD has announced 4x4 which is just a
rebranding of Opteron 2xx. It uses two packages connected via
hypertransport, each with their own memory controller. The AMD solution
has twice the memory bandwidth of the Intel solution and supports twice as
many DIMMs. If you need four cores then AMD is still the way to go. Each
core is slower but the aggregate system performance should be
significantly higher in a four core system. In a two core system Intel is
going to hold the lead for at least the next year. The AMD K8L might give
them the lead again but there is no way of knowing that right now. Also I
think that CPUs are pretty close to their limits for single thread
performance, at least for general purpose performance. For vectorizable
applications there is probably still some room for improvement.

In terms of memory sizes, it's a good bet that 8G systems will become
affordable in the next 12 months. The 2G DIMMs are available already but
they are still very expensive, by May they should be more affordable.

My personal plans are to wait until December and then get a Conroe system.
My reasons for waiting until December is a combination of technical and
tax reasons. The tax reason is that I want to deduct the system this year.
I also think that by December there will be enough people who are running
Linux on Conroe that any issues, if any, will have been worked out. I'd
like 8G of RAM but it doesn't seem like that will be very cost effective
this year, however you can always upgrade the RAM in a system so there is
no reason to wait until 2G DIMMs get cheap.

 

[wizzywiz]

The Conroe core is going to be the heart of Intel's CPUs for the next
couple of years, there will be tweaks and clock speed improvements in that
time but nothing that will make a big difference to single threaded
applications. Intel has a four core package coming in the next couple of
months but all it is, is two Conroe chips in the same package sharing the
same front side bus. The FSB is the weak point of Intel architectures.
It's adequate for a single dual core Conroe but I think it's likely to be
a bottleneck for four cores. AMD has announced 4x4 which is just a
rebranding of Opteron 2xx. It uses two packages connected via
hypertransport, each with their own memory controller. The AMD solution
has twice the memory bandwidth of the Intel solution and supports twice as
many DIMMs. If you need four cores then AMD is still the way to go. Each
core is slower but the aggregate system performance should be
significantly higher in a four core system. In a two core system Intel is
going to hold the lead for at least the next year. The AMD K8L might give
them the lead again but there is no way of knowing that right now. Also I
think that CPUs are pretty close to their limits for single thread
performance, at least for general purpose performance. For vectorizable
applications there is probably still some room for improvement.

In terms of memory sizes, it's a good bet that 8G systems will become
affordable in the next 12 months. The 2G DIMMs are available already but
they are still very expensive, by May they should be more affordable.

My personal plans are to wait until December and then get a Conroe system.
My reasons for waiting until December is a combination of technical and
tax reasons. The tax reason is that I want to deduct the system this year.
I also think that by December there will be enough people who are running
Linux on Conroe that any issues, if any, will have been worked out. I'd
like 8G of RAM but it doesn't seem like that will be very cost effective
this year, however you can always upgrade the RAM in a system so there is
no reason to wait until 2G DIMMs get cheap.

Thanks General. I had hoped you might mention some new chip coming from
AMD. I think there is supposed to be a new architecture out...

Anyway, maybe this will be useful:

Core 2 Duo System
Intel Boxed C2D E6600 CPU with Intel CPU heatsink and fan
Supermicro PDSMI+ 1PCI/1PCIe slots, two Gbit NICs
8192mb DDR2-533 SDRAM, 4 DIMMs
Seagate 80GB SATA 3Gb/s drive, 7200RPM
Sony DWQ30A DVD-R/RW/+R/RW/CD-R/RW drive, dual layer DVD+R
Onboard graphics card
Antec Solution Series SLK2650BQE ATX case, 8 drive bays, 350W power
Logitech 3 button mouse
IBM 104-key PS/2 keyboard
Built-in PCI sound card
Built-in network card
Fedora Core 5 x86_64
Two year replacement warranty

Total price: $2199.00 (plus shipping of around $35.00)

 

[NoNoBadDog!]

Thanks General. I had hoped you might mention some new chip coming from
AMD. I think there is supposed to be a new architecture out...

Anyway, maybe this will be useful:

Core 2 Duo System
Intel Boxed C2D E6600 CPU with Intel CPU heatsink and fan
Supermicro PDSMI+ 1PCI/1PCIe slots, two Gbit NICs
8192mb DDR2-533 SDRAM, 4 DIMMs
Seagate 80GB SATA 3Gb/s drive, 7200RPM
Sony DWQ30A DVD-R/RW/+R/RW/CD-R/RW drive, dual layer DVD+R
Onboard graphics card
Antec Solution Series SLK2650BQE ATX case, 8 drive bays, 350W power
Logitech 3 button mouse
IBM 104-key PS/2 keyboard
Built-in PCI sound card
Built-in network card
Fedora Core 5 x86_64
Two year replacement warranty

Total price: $2199.00 (plus shipping of around $35.00)

AMD does have new chips coming out, and they are not stuck with shared L2
cache.

They will be available in Q2 2007 (Feb).

Word is they will be faster than the Conroe..will still be AM2, and will
upgrade to future quad core chips on the same motherboard...a feature Intel
refuses to see the benefit of.

Bobby

 

[Bill Todd]

NoNoBadDog! wrote:

....

AMD does have new chips coming out, and they are not stuck with shared L2
cache.

On the whole, shared L2 cache (if implemented intelligently) is a *good*
thing to be 'stuck' with.

They will be available in Q2 2007 (Feb).

Word is they will be faster than the Conroe

Could you provide a reference for this 'word'? I haven't heard of any
architectural changes that would suggest this, nor that they'll be
clocked sufficiently faster than current parts to accomplish it.

So while quad-socket (and eventually 8-socket) AMD configurations may
retain (or if necessary regain) their lead due to their *system*
architecture, I know of no reason to expect very much relative
improvement at the individual processor level.

- bill

 

[Merrill P. L. Worthington]

NoNoBadDog! wrote:

...



On the whole, shared L2 cache (if implemented intelligently) is a *good*
thing to be 'stuck' with.



Could you provide a reference for this 'word'? I haven't heard of any
architectural changes that would suggest this, nor that they'll be
clocked sufficiently faster than current parts to accomplish it.

So while quad-socket (and eventually 8-socket) AMD configurations may
retain (or if necessary regain) their lead due to their *system*
architecture, I know of no reason to expect very much relative
improvement at the individual processor level.

- bill

Its interesting to track the development of faster processors.

But IMO, the issue isn't CPU horsepower. That's not the typical
bottleneck in a system. Its data access/hard drive speed.

The real performnance increases will come when hard drives can be
replaced by somce kind of inexpensive flash memory array capable of
persistent data storage and high-speed access. It needs to be of a
capacity large enough to replace a basic system disk. Maybe some
storage can be on a hard drive (maybe terabyte-size for long-term
storage), but o/s, program loading, and basic i/o needs to be on
something much faster than even the fastest of todays hard drives.

 

[General Schvantzkoph]

Thanks General. I had hoped you might mention some new chip coming from
AMD. I think there is supposed to be a new architecture out...

Anyway, maybe this will be useful:

Core 2 Duo System
Intel Boxed C2D E6600 CPU with Intel CPU heatsink and fan
Supermicro PDSMI+ 1PCI/1PCIe slots, two Gbit NICs
8192mb DDR2-533 SDRAM, 4 DIMMs
Seagate 80GB SATA 3Gb/s drive, 7200RPM
Sony DWQ30A DVD-R/RW/+R/RW/CD-R/RW drive, dual layer DVD+R
Onboard graphics card
Antec Solution Series SLK2650BQE ATX case, 8 drive bays, 350W power
Logitech 3 button mouse
IBM 104-key PS/2 keyboard
Built-in PCI sound card
Built-in network card
Fedora Core 5 x86_64
Two year replacement warranty

Total price: $2199.00 (plus shipping of around $35.00)

The one thing wrong with that system is the power supply, I'd want at
least a 500W supply. 500W will give you room to add more drives in the
future. Power supplies should be overkilled, you don't want to waste
time trying to figure out why your system is unstable just to save $25
on the supply. Also I'd put in a bigger disk, a 250G Seagate is priced at
$75 so there isn't any reason to get something as small as 80G.
Supermicro makes great motherboards, that's a good choice. The onboard
graphics is also a good choice for a server, you'll only need the
graphics when you install the OS so you don't want to waste money (and
more importantly power) on something that you won't be using.

 

[Bill Todd]

Its interesting to track the development of faster processors.

But IMO, the issue isn't CPU horsepower. That's not the typical
bottleneck in a system.

It does, however, with the possible addition of RAM performance appear
to be the bottleneck in the system under discussion here: see the
original post.

- bill

 

[Del Cecchi]

The one thing wrong with that system is the power supply, I'd want at
least a 500W supply. 500W will give you room to add more drives in the
future. Power supplies should be overkilled, you don't want to waste
time trying to figure out why your system is unstable just to save $25
on the supply. Also I'd put in a bigger disk, a 250G Seagate is priced
at
$75 so there isn't any reason to get something as small as 80G.
Supermicro makes great motherboards, that's a good choice. The onboard
graphics is also a good choice for a server, you'll only need the
graphics when you install the OS so you don't want to waste money (and
more importantly power) on something that you won't be using.

And I wasn't aware that IBM still made keyboards, much less "ps/2"
keyboards.

del cecchi

 

[Benjamin Gawert]

* General Schvantzkoph:

In terms of memory sizes, it's a good bet that 8G systems will become
affordable in the next 12 months.

Depends on what you mean with "affordable". It's extremely likely that
8GB won't be standard in two years, forget about 12 month. Besides that,
8GB requires an 64bit OS (if you don't want to use the crippled and slow
PAE of Windows Server or 32bit Linux), and with no sign of coming masses
of new 64bit applications this simply is useless, especially since 32bit
Windows programs are still limited to 2GB...

Benjamin

 

[Benjamin Gawert]

* General Schvantzkoph:

The one thing wrong with that system is the power supply, I'd want at
least a 500W supply. 500W will give you room to add more drives in the
future. Power supplies should be overkilled

Only if energy is free for you. Partially loaded standard PSUs usually
are _very_ ineffective...

Benjamin

 

[Merrill P. L. Worthington]

It does, however, with the possible addition of RAM performance appear
to be the bottleneck in the system under discussion here: see the
original post.

- bill

I did consider that as I have been following this thread. But the
thread seemed to take a tangent.

 

[Wes Newell]

* General Schvantzkoph:


Only if energy is free for you. Partially loaded standard PSUs usually
are _very_ ineffective...

Well, it would sure be more effective than a PSU that wouldn't even boot
the system.

 

[NoNoBadDog!]

NoNoBadDog! wrote:

...


On the whole, shared L2 cache (if implemented intelligently) is a *good*
thing to be 'stuck' with.

So please enlighten me on how sharing an L2 cache can be a bgood thing? The
proc-proc comms consume clock cycles. The cores cannot independently pwoer
down...the latency because of the shared L2 architecture is frightening..

See the recent press on the AMD website, and the speech that Hector Ruiz
gave regarding Apples decision to use AMD. Can't believe you haven't read
anything about K8L.

Could you provide a reference for this 'word'? I haven't heard of any
architectural changes that would suggest this, nor that they'll be clocked
sufficiently faster than current parts to accomplish it.

So while quad-socket (and eventually 8-socket) AMD configurations may
retain (or if necessary regain) their lead due to their *system*
architecture, I know of no reason to expect very much relative improvement
at the individual processor level.

- bill

Because direct connect and Hypertransport is better than what Intel
insipidly continues to cling to. Because an on-die memory controller is
better than shared L2 cache for proc-proc comms.


I used to be a supported of Intel. But as long an Intel refuses to see that
they are not headed in the right direction, I'll continue to buy AMD.

Now, if Intel were to adopt on-die memory controller, Direct Connect
Architecture, and Hypertransport, I'd give them due consideration.

Bobby

 

[Bill Todd]

So please enlighten me on how sharing an L2 cache can be a bgood thing?

It allows asymmetrical cache use, such that the processor that can
benefit more from a larger cache than it would have if each had 1/Nth
the shared size as private L2 cache can get it. In a single-thread
environment such as the one of specific interest to this discussion (see
original post), that's a *major* win.

The

proc-proc comms consume clock cycles.

Wrong: the principal negative impact is the circuit overhead of having
an additional port to the L2, which might add noticeable latency to an
L1 but is pretty well down in the noise compared to existing L2 latency.

The cores cannot independently pwoer

down...

Wrong - at least to any more degree than would be true with the shared
L3 that AMD is planning. There is no reason why processors cannot idle
at low power while sharing an L2 or L3 cache, and with appropriate
circuitry they could even power down completely if that were worthwhile.

the latency because of the shared L2 architecture is frightening..

Wrong, strike three: see quantitative comment above.

See the recent press on the AMD website, and the speech that Hector Ruiz
gave regarding Apples decision to use AMD.

Are you also ignorant of how to post a URL? Vague wavings of the hand
in some general direction do not qualify as 'references'.

While I like AMD and the healthy competition which it provides Intel,
they're as prone to marketing bullshit as anyone else - though I'd still
be a bit surprised if they'd come right out and state that K8L would
beat Conroe on single-thread benchmarks (the specific context of this
discussion, in case you forgot - and the assertion for which I requested
a reference from you).

Can't believe you haven't read

anything about K8L.

I've read quite a bit about K8L - by all appearances in considerably
more detail than you have.

Because direct connect and Hypertransport is better than what Intel
insipidly continues to cling to.

Apparently you'd like to equate multi-socket system performance to the
single-thread performance that was the subject of this thread before you
butted in - despite my rather specific differentiation between them
above. But that level of confusion from you is not really surprising.

Because an on-die memory controller is

better than shared L2 cache for proc-proc comms.

Dear me - now you've descended into *complete* drivel. I've been
writing kernel code for 3 decades and SMP/NUMA kernel code dependent
upon inter-processor communication for 1/3 of that time, so don't really
feel the need to listen further to someone who so clearly has no clue
whatsoever what he's talking about. But I did think it appropriate to
take a moment to point out your confusion to anyone who might have been
misled by it.

- bill

 

[already5chosen]

Thanks General. I had hoped you might mention some new chip coming from
AMD. I think there is supposed to be a new architecture out...

Anyway, maybe this will be useful:

Core 2 Duo System
Intel Boxed C2D E6600 CPU with Intel CPU heatsink and fan
Supermicro PDSMI+ 1PCI/1PCIe slots, two Gbit NICs
8192mb DDR2-533 SDRAM, 4 DIMMs
Seagate 80GB SATA 3Gb/s drive, 7200RPM
Sony DWQ30A DVD-R/RW/+R/RW/CD-R/RW drive, dual layer DVD+R
Onboard graphics card
Antec Solution Series SLK2650BQE ATX case, 8 drive bays, 350W power
Logitech 3 button mouse
IBM 104-key PS/2 keyboard
Built-in PCI sound card
Built-in network card
Fedora Core 5 x86_64
Two year replacement warranty

Total price: $2199.00 (plus shipping of around $35.00)

Officially, you can't plug CPU marked E6600 into Mikilteo-2 motherboard
like PDSMI+. You need Xeon 3060. Of course, physically E6600 and Xeon
3060 is the same thing but you are likely to lose warranty.

Do you realize that Mikilteo-2 is server chipset and its on board
graphic (ATI ES1000) is not designed even for 2D workstation?
Here the data page of ATI ES1000. IMHO, 1280x1024, 85Hz is NOT enough.
http://www.ati.com/products/embedded/es1000/features.html

Even onboard Intel G965 (like in Supermicro PDSBA ) is much better. Of
course, external graphics is the best but you could consider it an
overkill. Anyway, PDSBA supports external graphics. PDSMI+ doesn't

BTW, memory price is bargain. Normally, 2GB DIMMs for single-socket
Intel systems are very expensive. Your price is so good that it is
almost too good to be true. Could you drop a link to all of us?

 

[Stephen Sprunk]

My current dilemma is shoud I buy now or wait? ...
I can hold off on buying as late as next May. What do you see coming
down the pike by then that may offer better performance in my
application?

The simple answer when dealing with any tech purchase is to wait as long
as you can stand it, because there's always something better, faster,
and cheaper about to come out.

Of course, you have to weigh the cost of not having the newer machine
while you wait, so it's not so simple in reality, particularly if you
don't have a dollar cost for the time spent waiting. Is having today's
hottest machine for the next 6 mos worth the extra money you'd spend
over what that machine would cost 6 mos from now? Only you can answer
that.

I'd advise you to check out AnandTech, Tom's Hardware, and other similar
enthusiast sites to get the roadmaps from Intel and AMD and make your
own decisions. My personal strategy is to buy one of the first boards
available with a new socket and the cheapest processor that'll fit at
the time, then hold off on upgrading the CPU until the socket is dead
and the fastest processor for it goes on sale for being "obsolete". I
skip the next socket, then repeat the process when the _next_ socket
comes out. I don't stay at the cutting edge, but I'm reasonably close
and spend less than half the money of people that upgrade every chance
they get.

S

 

[ranjit_mathews]

* General Schvantzkoph:


Depends on what you mean with "affordable".

Apparently, he doesn't consider it to be affordable now or to have been
affordable last year, so he must mean that they will become as
affordable as today's 4GB (or 2GB?) systems. I'd be tickled to death if
a pair of 4GB SO-DIMMs in a year for today's price of a pair of 1GB
SO-DIMMs ... or if a set of 4 2GB DIMMs will be available for today's
price of a set of 4 1/2 GB DIMMs (they're currently different in price
by a factor of 10 or so).

It's extremely likely that
8GB won't be standard in two years, forget about 12 month. Besides that,
8GB requires an 64bit OS (if you don't want to use the crippled and slow
PAE of Windows Server or 32bit Linux), and with no sign of coming masses
of new 64bit applications this simply is useless, especially since 32bit
Windows programs are still limited to 2GB...

64 bit Windows 2003 server will probably be quite standard in 12
months.