Sign in with
Sign up | Sign in
Your question

Should I wait for hypertransport2

Last response: in Overclocking
Share
Anonymous
a b K Overclocking
October 23, 2004 8:37:25 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

X-No-Archive: Yes

I am looking for a Socket 939 motherboard
-- i was thinking about the
MSI K8T Neo2-FIR or MSI K8N Neo2 Platinum

but then i heard about hypertransport 2

http://www.tomshardware.com/hardnews/20041020_103235.ht...
http://www.via.com.tw/en/products/chipsets/k8-series/k8...

should i wait,? can i get this today ?
or is this just hype - again - ?
i plan on using a Athlon 64 3000+


thanks for any reply

More about : wait hypertransport2

Anonymous
a b K Overclocking
October 23, 2004 11:38:08 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

> I am looking for a Socket 939 motherboard
> -- i was thinking about the
> MSI K8T Neo2-FIR or MSI K8N Neo2 Platinum
>
> but then i heard about hypertransport 2
>
> http://www.tomshardware.com/hardnews/20041020_103235.ht...
> http://www.via.com.tw/en/products/chipsets/k8-series/k8...
>
> should i wait,? can i get this today ?
> or is this just hype - again - ?
> i plan on using a Athlon 64 3000+

In a single-chip system, I don't believe that it's going to make any
measurable difference. Because the memory controller is embedded in the
CPU, the only big bandwidth users of the HT link will be the AGP port and
PCI bus(ses). With AGP 8x only using 2.1 GB/s, and a 32-bit, 33-MHz PCI bus
only using about 110 MB/sec, even HyperTransport 1.0 (12.8 GB/s) is barely
even warming up.

The real workout for HT is in a multi-processing system, where one CPU can
access memory attached to another CPU. With each Opteron's memory
controller having a theoretical max of about 6 GB/sec, once you get 4 (or
more!) CPUs in a system, it's easy to see how HT 2.0 is necessary.

steve
Anonymous
a b K Overclocking
October 24, 2004 2:31:09 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

The new MSI Nforce 4 motherboards should be available very soon,they offer
PCI-E support and the SLI version offers dual Video card support. The price
should be in the same range as whats available now. You would have to buy a
PCI-E video card. If your not in a big hurry for a new suytem i would wait.
DOUG
Related resources
Anonymous
a b K Overclocking
October 24, 2004 9:55:46 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

"Courseyauto" <courseyauto@aol.com> wrote in message
news:20041024063109.22107.00005659@mb-m07.aol.com...
> The new MSI Nforce 4 motherboards should be available very soon,they offer
> PCI-E support and the SLI version offers dual Video card support. The
price
> should be in the same range as whats available now. You would have to buy
a
> PCI-E video card. If your not in a big hurry for a new suytem i would
wait.
> DOUG


Yea.. I been saying this since the a64's came out.. Wait for pci-express
support and THEN go for it.. waiting time is over really :) 
Anonymous
a b K Overclocking
October 25, 2004 6:42:38 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

"rstlne" <.@text.news.virgin.net> wrote in message
news:C0Sed.118$6U.2@newsfe1-gui.ntli.net...
>
> "Courseyauto" <courseyauto@aol.com> wrote in message
> news:20041024063109.22107.00005659@mb-m07.aol.com...
> > The new MSI Nforce 4 motherboards should be available very soon,they
offer
> > PCI-E support and the SLI version offers dual Video card support. The
> price
> > should be in the same range as whats available now. You would have to
buy
> a
> > PCI-E video card. If your not in a big hurry for a new suytem i would
> wait.
> > DOUG
>
>
> Yea.. I been saying this since the a64's came out.. Wait for pci-express
> support and THEN go for it.. waiting time is over really :) 
>
>

What is annoying to me though, is that when DDR2 comes out a few more months
down the road, you know AMD will come to you and say "hey, we are going to
change the on die memory controller to cater for DDR 2, or HT2, or Dual
core, or whatever, so you guys will have to change sockets again"
Anonymous
a b K Overclocking
October 25, 2004 7:37:46 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

>What is annoying to me though, is that when DDR2 comes out a few more months
>down the road, you know AMD will come to you and say "hey, we are going to
>change the on die memory controller to cater for DDR 2, or HT2, or Dual
core, or whatever, so you guys will have to change sockets >again"


You make it sound like they will drop the 939 and you will have to buy whatever
the make . DDR2 is much slower than DDR1,you can buy DDR1 that will keep up
with just about whatever FSB you can run at the moment.
Anonymous
a b K Overclocking
October 25, 2004 9:24:05 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

ok thanks for all the replies the weath of info is eye opening - :) 
i will get a "NEO2" motherboard for now
which one is the best for AMD64 3000+
MSI K8T Neo2-FIR or MSI K8N Neo2 Platinum

i will wait for the bigger upgrade may be next summer or winter--

I really need that SLI option-- but the app i need to run is not ready yet :
http://www.bionicfx.com/

i hope to have a machine that will replace all my current audio gear
- i am close now but- i use alot! of DSP --vst and dxi softsynths and fx---
:)  and want to use more -
there are dsp cards out but the GPU option is better in the long run
because i also am getting into 3d


thanks again guys - keep overclocking :) 




"Courseyauto" <courseyauto@aol.com> wrote in message
news:20041025113746.16740.00002459@mb-m11.aol.com...
> >What is annoying to me though, is that when DDR2 comes out a few more
months
> >down the road, you know AMD will come to you and say "hey, we are going
to
> >change the on die memory controller to cater for DDR 2, or HT2, or Dual
> core, or whatever, so you guys will have to change sockets >again"
>
>
> You make it sound like they will drop the 939 and you will have to buy
whatever
> the make . DDR2 is much slower than DDR1,you can buy DDR1 that will keep
up
> with just about whatever FSB you can run at the moment.
Anonymous
a b K Overclocking
October 26, 2004 1:51:02 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

"Steve Wolfe" <unt@codon.com> wrote in message
news:2u0eu8F2417kcU1@uni-berlin.de...
>> I am looking for a Socket 939 motherboard
>> -- i was thinking about the
>> MSI K8T Neo2-FIR or MSI K8N Neo2 Platinum
>>
>> but then i heard about hypertransport 2
>>
>> http://www.tomshardware.com/hardnews/20041020_103235.ht...
>> http://www.via.com.tw/en/products/chipsets/k8-series/k8...
>>
>> should i wait,? can i get this today ?
>> or is this just hype - again - ?
>> i plan on using a Athlon 64 3000+
>
> In a single-chip system, I don't believe that it's going to make any
> measurable difference. Because the memory controller is embedded in the
> CPU, the only big bandwidth users of the HT link will be the AGP port and
> PCI bus(ses). With AGP 8x only using 2.1 GB/s, and a 32-bit, 33-MHz PCI
> bus

thats all changing with PCI express

> only using about 110 MB/sec, even HyperTransport 1.0 (12.8 GB/s) is barely
> even warming up.
>
> The real workout for HT is in a multi-processing system, where one CPU
> can
> access memory attached to another CPU. With each Opteron's memory
> controller having a theoretical max of about 6 GB/sec, once you get 4 (or
> more!) CPUs in a system, it's easy to see how HT 2.0 is necessary.
>
> steve
>
>
Anonymous
a b K Overclocking
October 26, 2004 1:51:03 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

> > In a single-chip system, I don't believe that it's going to make any
> > measurable difference. Because the memory controller is embedded in the
> > CPU, the only big bandwidth users of the HT link will be the AGP port
and
> > PCI bus(ses). With AGP 8x only using 2.1 GB/s, and a 32-bit, 33-MHz PCI
> > bus
>
> thats all changing with PCI express

Not really. x16 PCI-E still only takes 8.1 GB/sec. Maxing out your
legacy PCI bus as well, you're still only at about 8.2 GB/sec, you're still
not filling up the HT link. Alternatively, on a strictly PCI-E board, all
20 lanes running full-out still can't quite fill a HyperTransport link, and
you're going to be EXTREMELY hard-pressed to get all 20 lanes running
full-out.

steve
Anonymous
a b K Overclocking
October 30, 2004 1:43:31 AM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

Steve Wolfe wrote:
>>> In a single-chip system, I don't believe that it's going to make
>>> any measurable difference. Because the memory controller is
>>> embedded in the CPU, the only big bandwidth users of the HT link
>>> will be the AGP port and PCI bus(ses). With AGP 8x only using 2.1
>>> GB/s, and a 32-bit, 33-MHz PCI bus
>>
>> thats all changing with PCI express
>
> Not really. x16 PCI-E still only takes 8.1 GB/sec. Maxing out your
> legacy PCI bus as well, you're still only at about 8.2 GB/sec, you're
> still not filling up the HT link.

Check your maths :)  A 1GHz hypertrnasport link moves 1000*16/8*2 (1GHz * 16
bits / 8 bits per byte * 2 transfers per cycle) = 4GBytes/sec each
direction. So a x16 PCI-E link that moves 250*16=4GBytes/sec (HDD
manufacturer GBytes that is :)  ) each direction will exactly fill a 1GHz
hypertransport link. I've got a niggling little factoid stored away that the
current PCI-E protocol only allows a maximum of 32 lanes (it's probably in
the specs somewhere, but I lack the time to go hunt it down at the moment
:)  ), which means that the current version of PCI-express will happily eat
two full hypertransport links.

With an Athlon64, you've only got one link to use, so there's not too much
that you can do. However, even on a 1xx Opteron system you've got three I/O
links to burn, so you can potentially have 48 PCI-E lanes going full tilt. I
wouldn't be overly surprised if someone develops a dual x16 PCI-E Opteron
system, especially seeing some of the funky designs that Sun has come up
with using existing chipsets.

[...]

--
Michael Brown
www.emboss.co.nz : OOS/RSI software and more :) 
Add michael@ to emboss.co.nz - My inbox is always open
Anonymous
a b K Overclocking
October 30, 2004 1:43:32 AM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

> > Not really. x16 PCI-E still only takes 8.1 GB/sec. Maxing out your
> > legacy PCI bus as well, you're still only at about 8.2 GB/sec, you're
> > still not filling up the HT link.
>
> Check your maths :)  A 1GHz hypertrnasport link moves 1000*16/8*2 (1GHz *
16
> bits / 8 bits per byte * 2 transfers per cycle) = 4GBytes/sec each
> direction.

I've seen quite a lot of references to 12.8 GB/sec peak bidirectional, or
6.4 unidirectional. But, if your numbers are right, you have a good point.

steve
Anonymous
a b K Overclocking
October 30, 2004 2:46:08 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

Steve Wolfe wrote:
>>> Not really. x16 PCI-E still only takes 8.1 GB/sec. Maxing out
>>> your legacy PCI bus as well, you're still only at about 8.2 GB/sec,
>>> you're still not filling up the HT link.
>>
>> Check your maths :)  A 1GHz hypertrnasport link moves 1000*16/8*2
>> (1GHz * 16 bits / 8 bits per byte * 2 transfers per cycle) =
>> 4GBytes/sec each direction.
>
> I've seen quite a lot of references to 12.8 GB/sec peak
> bidirectional, or
> 6.4 unidirectional.

This is the absolute max you can move and still be within the hypertransport
specification. The hypertransport spec allows speeds of up to 800MHz
(despite curent CPUs operating at 1GHz :)  ) and 32 bit widths. The
A64/Opteron cannot be configured to run in 32-bit mode (ie: canot have a
unidirectional link, must be able to receive data over the link as well) so
is limited to having 16-bit links. So each direction is limited to
3.2GBytes/sec if they want to stay within the spec (unless they change
sockets to allow for 32-bit hypertransport). Ramping the frequency up to
1GHz allows for 4GBytes/sec, and since AMD essentially define the spec, they
can almost do what they please :) 

The 12.8 number is a little deceptive as it counts data going both ways, so
it's impossible to move 12.8GB from A to B in one second. Most busses
nowadays are full-duplex so it makes more sense to only refer to the amount
of data that can be moved in one direction. Or perhaps it's just me :) 

[...]

--
Michael Brown
www.emboss.co.nz : OOS/RSI software and more :) 
Add michael@ to emboss.co.nz - My inbox is always open
Anonymous
a b K Overclocking
October 30, 2004 2:46:09 PM

Archived from groups: alt.comp.hardware.overclocking.amd (More info?)

> This is the absolute max you can move and still be within the
hypertransport
> specification. The hypertransport spec allows speeds of up to 800MHz
> (despite curent CPUs operating at 1GHz :)  ) and 32 bit widths. The
> A64/Opteron cannot be configured to run in 32-bit mode (ie: canot have a
> unidirectional link, must be able to receive data over the link as well)
so
> is limited to having 16-bit links. So each direction is limited to
> 3.2GBytes/sec if they want to stay within the spec (unless they change
> sockets to allow for 32-bit hypertransport). Ramping the frequency up to
> 1GHz allows for 4GBytes/sec, and since AMD essentially define the spec,
they
> can almost do what they please :) 
>
> The 12.8 number is a little deceptive as it counts data going both ways,
so
> it's impossible to move 12.8GB from A to B in one second. Most busses
> nowadays are full-duplex so it makes more sense to only refer to the
amount
> of data that can be moved in one direction. Or perhaps it's just me :) 

You're absolutely right. Looking at an article on Anandtech today, they
mentioned that even going down to an x4 PCI-E connection shows only a
negligible decrease in video performance in games, so even if a
HyperTransport 1.0 doesn't have quite enough bandwidth to fill an x16 PCI-E
and a regular PCI bus at the same time, I still don't think it's going to
hurt anything. I'm not in the market for an A64 board at the moment, but if
I were, I wouldn't care whether it was HT 1 or 2. On the other hand, next
year I'll be buying at least another 4-way Opteron, probably two of them -
and having HT 2.0 on those would be *very* nice. Unfortunately, I have a
feeling that those machines (where it will make the most difference) will be
considerably slower to see HT 2 units in production.

steve
!