Athlon 64 FX 51 mediocre performance on unbuffered memory ..

[Guest]

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

http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml

The Athlon 64 FX 51 with dual channel PC3200 doesn't score remarkably well
on the UNbuffered Sandra memory test -- e.g. compared to a P4 with similar
dual-channel PC3200 DDR.

Why?

Is the Athlon 64 memory interface throttled to a lower total bandwidth? I
thought they were both (AMD and P4) ~ 6.2 GB/s max.

 

[Guest]

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

>>>>> "H" == H W Stockman <stockman3@earth-REMOVE_THIS-link.net> writes:

H> http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml The Athlon
H> 64 FX 51 with dual channel PC3200 doesn't score remarkably well on
H> the UNbuffered Sandra memory test -- e.g. compared to a P4 with
H> similar dual-channel PC3200 DDR.

H> Why?

H> Is the Athlon 64 memory interface throttled to a lower total
H> bandwidth? I thought they were both (AMD and P4) ~ 6.2 GB/s max.

The problem is the NVIDIA nForce 3 pro150 - TRY a VIA based
motherboard. VIA has better memory bandwidth. Hopefully the next round
of nvidia chipsets will have better memory bandwidth. It should be
noted that in real world usage users can't really tell the difference.

Actually according to some of the hardware review sites the SIS amd 64
chipsets have good memory bandwidth also.

Good luck

 

[Guest]

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

"Post Replies Here Please" <spamme@edge.net> wrote in message
news:87brlsa8hb.fsf@spamme.edge.net...
> >>>>> "H" == H W Stockman <stockman3@earth-REMOVE_THIS-link.net> writes:
[...]
> The problem is the NVIDIA nForce 3 pro150 - TRY a VIA based
> motherboard. VIA has better memory bandwidth. Hopefully the next round
> of nvidia chipsets will have better memory bandwidth. It should be
> noted that in real world usage users can't really tell the difference.
>
> Actually according to some of the hardware review sites the SIS amd 64
> chipsets have good memory bandwidth also.

Do you know of any tests of UNbuffered memory performance for the other
chipsets (on the web)?

I've looked high and low, and have found zilch.

 

[chip]

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

"Post Replies Here Please" <spamme@edge.net> wrote in message
news:87brlsa8hb.fsf@spamme.edge.net...
> >>>>> "H" == H W Stockman <stockman3@earth-REMOVE_THIS-link.net> writes:
>
> H> http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml The Athlon
> H> 64 FX 51 with dual channel PC3200 doesn't score remarkably well on
> H> the UNbuffered Sandra memory test -- e.g. compared to a P4 with
> H> similar dual-channel PC3200 DDR.
>
> H> Why?
>
> H> Is the Athlon 64 memory interface throttled to a lower total
> H> bandwidth? I thought they were both (AMD and P4) ~ 6.2 GB/s max.
>
> The problem is the NVIDIA nForce 3 pro150 - TRY a VIA based
> motherboard. VIA has better memory bandwidth.

That comment surprises me. Since the memory controller is part of the chip
with the Athlon 64 (i.e. its not part of the northbridge), then how can one
FX51 chipset have better memory performance than another? It doesn't make
sense, its got nothing to do with the chipset, surely?

I thought the different Athlon chipsets only differeed in terms of their AGP
performance, northbridge-to-southbridge performance, and southbridge
performance. But I am struggling to see how they would have different
memory performance, since none of them go anywhere near the memory.

Perhaps I misunderstand something?

Chip

 

[Guest]

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

"Chip" <anneonymouse@virgin.net> wrote in message
news:c5roab$5288l$1@ID-185713.news.uni-berlin.de...
>
> That comment surprises me. Since the memory controller is part of the
chip
> with the Athlon 64 (i.e. its not part of the northbridge), then how can
one
> FX51 chipset have better memory performance than another? It doesn't make
> sense, its got nothing to do with the chipset, surely?

That was my impression too, but I admit not being up on all the gewgaws.

I was hoping that a few people with other FX 51 motherboards would post some
UNbuffered Sandra benchmarks, showing better performance. I've posted
benchmark requests in many groups; so far, no interest in UNbuffered tests.
I am beginning to think that (1) the market is so fragmented, and only a
very small number have the boards... or (2) the FX 51 doesn't do well on
this test, so people are reluctant to highlight a weakness of the
chip/board.

 

[Guest]

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

"Chip" <anneonymouse@virgin.net> wrote in message
news:c5roab$5288l$1@ID-185713.news.uni-berlin.de...
>
[...]
> That comment surprises me. Since the memory controller is part of the
chip
> with the Athlon 64 (i.e. its not part of the northbridge), then how can
one
> FX51 chipset have better memory performance than another? It doesn't make
> sense, its got nothing to do with the chipset, surely?

Here's the first hint of an explanation...
http://www.lostcircuits.com/memory/reg_ddr/5.shtml

 

[Guest]

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

On Sat, 17 Apr 2004 18:08:28 +0100, Chip wrote:

> That comment surprises me. Since the memory controller is part of the chip
> with the Athlon 64 (i.e. its not part of the northbridge), then how can one
> FX51 chipset have better memory performance than another? It doesn't make
> sense, its got nothing to do with the chipset, surely?
>
The Hypertransport link of the NF3-150 chipset will only run at 600MHz.
3/4 speed iow's. VIA, SIS, and Ali run at full speed (800MHz).

--
Abit KT7-Raid (KT133) Tbred B core CPU @2400MHz (24x100FSB)
http://mysite.verizon.net/res0exft/cpu.htm

 

[Guest]

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

"Wes Newell" <w.newell@TAKEOUTverizon.net> wrote in message
newsan.2004.04.17.23.50.38.332793@TAKEOUTverizon.net...
> On Sat, 17 Apr 2004 18:08:28 +0100, Chip wrote:
>
> > That comment surprises me. Since the memory controller is part of the
chip
> > with the Athlon 64 (i.e. its not part of the northbridge), then how can
one
> > FX51 chipset have better memory performance than another? It doesn't
make
> > sense, its got nothing to do with the chipset, surely?
> >
> The Hypertransport link of the NF3-150 chipset will only run at 600MHz.
> 3/4 speed iow's. VIA, SIS, and Ali run at full speed (800MHz).


What does the hypertransport link have to do with the MCT (memory
controller)?

 

[Guest]

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

On Sun, 18 Apr 2004 00:29:56 +0000, H.W. Stockman wrote:

>
> "Wes Newell" <w.newell@TAKEOUTverizon.net> wrote in message
> newsan.2004.04.17.23.50.38.332793@TAKEOUTverizon.net...
>> On Sat, 17 Apr 2004 18:08:28 +0100, Chip wrote:
>>
>> > That comment surprises me. Since the memory controller is part of the
> chip
>> > with the Athlon 64 (i.e. its not part of the northbridge), then how can
> one
>> > FX51 chipset have better memory performance than another? It doesn't
> make
>> > sense, its got nothing to do with the chipset, surely?
>> >
>> The Hypertransport link of the NF3-150 chipset will only run at 600MHz.
>> 3/4 speed iow's. VIA, SIS, and Ali run at full speed (800MHz).
>
> What does the hypertransport link have to do with the MCT (memory
> controller)?

Nothing, other than that's why the system performance is lower with the
NF3-150 chipset.

--
Abit KT7-Raid (KT133) Tbred B core CPU @2400MHz (24x100FSB)
http://mysite.verizon.net/res0exft/cpu.htm

 

[chip]

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

"Wes Newell" <w.newell@TAKEOUTverizon.net> wrote in message
newsan.2004.04.18.05.57.20.161997@TAKEOUTverizon.net...
> On Sun, 18 Apr 2004 00:29:56 +0000, H.W. Stockman wrote:
>
> >
> > "Wes Newell" <w.newell@TAKEOUTverizon.net> wrote in message
> > newsan.2004.04.17.23.50.38.332793@TAKEOUTverizon.net...
> >> On Sat, 17 Apr 2004 18:08:28 +0100, Chip wrote:
> >>
> >> > That comment surprises me. Since the memory controller is part of
the
> > chip
> >> > with the Athlon 64 (i.e. its not part of the northbridge), then how
can
> > one
> >> > FX51 chipset have better memory performance than another? It doesn't
> > make
> >> > sense, its got nothing to do with the chipset, surely?
> >> >
> >> The Hypertransport link of the NF3-150 chipset will only run at 600MHz.
> >> 3/4 speed iow's. VIA, SIS, and Ali run at full speed (800MHz).
> >
> > What does the hypertransport link have to do with the MCT (memory
> > controller)?
>
> Nothing, other than that's why the system performance is lower with the
> NF3-150 chipset.

True. But the thread is discussing memory performance, not system
performance. Hence HW's question.

Chip.

 

[Guest]

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

On Sat, 17 Apr 2004 18:08:28 +0100, "Chip" <anneonymouse@virgin.net> wrote:

>
>"Post Replies Here Please" <spamme@edge.net> wrote in message
>news:87brlsa8hb.fsf@spamme.edge.net...
>> >>>>> "H" == H W Stockman <stockman3@earth-REMOVE_THIS-link.net> writes:
>>
>> H> http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml The Athlon
>> H> 64 FX 51 with dual channel PC3200 doesn't score remarkably well on
>> H> the UNbuffered Sandra memory test -- e.g. compared to a P4 with
>> H> similar dual-channel PC3200 DDR.
>>
>> H> Why?
>>
>> H> Is the Athlon 64 memory interface throttled to a lower total
>> H> bandwidth? I thought they were both (AMD and P4) ~ 6.2 GB/s max.
>>
>> The problem is the NVIDIA nForce 3 pro150 - TRY a VIA based
>> motherboard. VIA has better memory bandwidth.
>
>That comment surprises me. Since the memory controller is part of the chip
>with the Athlon 64 (i.e. its not part of the northbridge), then how can one
>FX51 chipset have better memory performance than another? It doesn't make
>sense, its got nothing to do with the chipset, surely?
>
>I thought the different Athlon chipsets only differeed in terms of their AGP
>performance, northbridge-to-southbridge performance, and southbridge
>performance. But I am struggling to see how they would have different
>memory performance, since none of them go anywhere near the memory.
>
>Perhaps I misunderstand something?

Probably the integrated memory controller revision/stepping matters just like cache performance on
Intel P4 CPUs was affected by bugs on early steps (P4 FSB 400 CPUs for example) that made it run
slower and even microcode update patches couldn't fix that.

 

[Guest]

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

On Sat, 17 Apr 2004 18:08:28 +0100, "Chip" <anneonymouse@virgin.net> wrote:

>
>"Post Replies Here Please" <spamme@edge.net> wrote in message
>news:87brlsa8hb.fsf@spamme.edge.net...
>> >>>>> "H" == H W Stockman <stockman3@earth-REMOVE_THIS-link.net> writes:
>>
>> H> http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml The Athlon
>> H> 64 FX 51 with dual channel PC3200 doesn't score remarkably well on
>> H> the UNbuffered Sandra memory test -- e.g. compared to a P4 with
>> H> similar dual-channel PC3200 DDR.
>>
>> H> Why?
>>
>> H> Is the Athlon 64 memory interface throttled to a lower total
>> H> bandwidth? I thought they were both (AMD and P4) ~ 6.2 GB/s max.
>>
>> The problem is the NVIDIA nForce 3 pro150 - TRY a VIA based
>> motherboard. VIA has better memory bandwidth.
>
>That comment surprises me. Since the memory controller is part of the chip
>with the Athlon 64 (i.e. its not part of the northbridge), then how can one
>FX51 chipset have better memory performance than another? It doesn't make
>sense, its got nothing to do with the chipset, surely?
>
>I thought the different Athlon chipsets only differeed in terms of their AGP
>performance, northbridge-to-southbridge performance, and southbridge
>performance. But I am struggling to see how they would have different
>memory performance, since none of them go anywhere near the memory.
>
>Perhaps I misunderstand something?

>Probably the integrated memory controller revision/stepping >matters just like
cache performance on

>Intel P4 CPUs was affected by bugs on early steps (P4 FSB 400 >CPUs for
example) that made it run
>slower and even microcode update patches couldn't fix that.

The memory itself will have a big impact on the scores. Its kind of like
CPU's overclock differently and memory sticks very in performance. DOUG

 

[Guest]

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

"Courseyauto" <courseyauto@aol.com> wrote in message
news:20040418193433.23507.00000121@mb-m18.aol.com...
>

> The memory itself will have a big impact on the scores. Its kind of
like
> CPU's overclock differently and memory sticks very in performance. DOUG


I repeat the comments from earlier in the thread:
:: Here's the first hint of an explanation...
:: http://www.lostcircuits.com/memory/reg_ddr/5.shtml

 

[Guest]

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

H.W. Stockman wrote:
> http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml
>
> The Athlon 64 FX 51 with dual channel PC3200 doesn't score remarkably
> well on the UNbuffered Sandra memory test -- e.g. compared to a P4
> with similar dual-channel PC3200 DDR.

It doesn't? From the same site:
http://legitreviews.com/reviews/prescottoc/2.html

I think their "stock" is PC3700 running at 200MHz, with 2-3-3-6 timings.
This gives an unbuffered score of 2662/2694 (int/float).

Their Legacy RAM with the FX51 is run at 200MHz with 2.5-7-3-3 timings (ie:
slower), and gives 2468 (presumably about the same for int and float). It's
not exactly the same, but given that it's running at slower timings (which
heavily influence the unbuffered tests) and that it's registered RAM (which
usually means a cycle or two is "lost" somewhere along the line), I'd say
that the FX51 didn't do too bad of a job. Note that they don't specify which
version of Sandra they are using for the P4 so this introduces a further
variance in the numbers.

And, anyhow, unbuffered tests aren't hugely important. Just about every
recent application that is memory-limited uses block prefetching, so the
"buffered" numbers are the important ones. 5500MBytes/sec is about 86%
efficiency, which is about right for DDR IIRC, and is only just behind the
P4 on 5667MBytes/sec (Aida and Sandra return about the same for buffered
numbers). Again, this difference can be attributed to the registered RAM,
slower timings, and (in this case) different programs used.

[...]

--
Michael Brown
www.emboss.co.nz : OOS/RSI software and more
Add michael@ to emboss.co.nz - My inbox is always open

 

[Guest]

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

"Michael Brown" <see@signature.below> wrote in message
news:KLQgc.1569$2%1.28438@news.xtra.co.nz...
> H.W. Stockman wrote:
> > http://www.legitreviews.com/Reviews/reg_hyperx_2.shtml
> >
> > The Athlon 64 FX 51 with dual channel PC3200 doesn't score remarkably
> > well on the UNbuffered Sandra memory test -- e.g. compared to a P4
> > with similar dual-channel PC3200 DDR.
>
> It doesn't? From the same site:
> http://legitreviews.com/reviews/prescottoc/2.html
>
> I think their "stock" is PC3700 running at 200MHz, with 2-3-3-6 timings.
> This gives an unbuffered score of 2662/2694 (int/float).

Look back at the "possible explanation" given earlier in this thread -- the
type of memory seems to make a 20% difference in unbuffered scores. I had a
very hard time finding any sites that mentioned FX 51 unbuffered
performance, but have seen > 3000 on P4s/ Legacy RAM has actually done
better on the unbuffered scores:
: I repeat the comments from earlier in the thread:
:: Here's the first hint of an explanation...
:: http://www.lostcircuits.com/memory/reg_ddr/5.shtml

> And, anyhow, unbuffered tests aren't hugely important. Just about every
> recent application that is memory-limited uses block prefetching, so the
> "buffered" numbers are the important ones.

Can you find a quantitative reference to the above statements?

I have applications that use largely unbuffered memory access, and I'm
trying to select a processor for them. Block prefetching (hand-encoded) has
sped them up by 20% on both Athlons and P4s. "Memory-limited" and
"unbuffered" are rather unspecific terms; if there is little natural data
reuse in an application that handles large arrays, the L2 caches mainly get
in the way.

 

[Guest]

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

"H.W. Stockman" <stockman3@earth-REMOVE_THIS-link.net> wrote in message
news:HBSgc.2361$e4.1065@newsread2.news.pas.earthlink.net...
[...]
> type of memory seems to make a 20% difference in unbuffered scores. I had
a
> very hard time finding any sites that mentioned FX 51 unbuffered
> performance, but have seen > 3000 on P4s.

For example,
(P4):
http://www.anandtech.com/memory/showdoc.html?i=2019&p=7