|
Specifications,
cooling & monitoring, memory
Test
System Specification:
|
Iwill
DK8N Rev 0.92 BIOS 21July04
|
|
CPUs
|
2x Opteron 250 2.4GHz CG-stepping
|
|
DDR
|
4x
256MB OCZ Registered/ECC 2-3-2-6 PC3200
|
|
Waterblocks
|
2x Zalman ZM-WB2
|
|
Waterpump
|
Eheim
1048
|
|
Radiator/Fans
|
Swiftech
676/2x Panaflo 120x38mm LIA
|
|
Northbridge
heatsink
|
Zalman
NB47J
|
|
Video-card
|
Gainward
FX5200
|
|
Monitor
|
Eizo
21"
|
|
Mouse/Keyboard
|
MS
USB Intellimouse Optical/MS USB Internet Pro
|
|
Sound
|
Onboard
Realtek ALC655
|
|
Speakers
|
Videologic
5+1
|
|
LAN
|
Onboard
Marvell/NFORCE3 Gigabit
|
|
33MHz
32-bit
PCI
|
Symbios
22801 [LSIlogic U40SE] 2-ch ultra wide SE
|
|
66MHz
64-bit PCI
|
Compaq/HP
SmartArray 5304 4-ch U160 1.92
|
|
133MHz
64-bit PCI-X
|
Compaq/HP
SmartArray 6402 2-ch U320 3.54
|
|
SCSI
Hard
Disks - 5 arrays: single 15K rpm drive RAID0; 4-drive RAID1+0;
3-drive RAID5; 9-drive [shared] 8+hotspare RAID5 & 8-drive RAID0
|
13x
fujitsu MAN318xx (9 external); 3x fujitsu MAN337xx (external); 1x
fujitsu MAS337xx
|
|
SCSI
CDRW/CDROM/DVDROM - all ultra narrow; 2 to each channel SYM 22801
|
Yamaha
F1S & Plextor 12/10/32S (external); Plextor
TS40i; Pioneer 305S
|
|
PATA
Hard Disk
|
Quantum
Fireball [boot - all OS' & data on SCSI arrays]
|
|
PATA
DVD +/- RW
|
BenQ
822A
|
|
System
enclosure
|
Macase
ATX fileserver (customised)
|
|
External
mass storage enclosure
|
CIDesign
tower (customised); 2x Intel SC5xxx 5-drive hotswap cages; 1x
3-drive hotswap
cage; redundant PSUs
|
|
External
USB2 HD/DVD
|
(in)
'Sohotank' U7 - B2 - 1
|
|
UPS'
|
APC
SmartUPS 900; APC BackUPS Pro 650
|
|
EPS12V
SSI PSU
|
PC
Power & Cooling TurboCool 510AG
|
|
32-bit
OS'
|
MS
Server 2003 Enterprise; XP Pro SP1 & SP2
|
|
32-bit
drivers
|
NFORCE
2.34~8 (SMP); Realtek 1.61; 8131 1.80; NVIDIA 61.77
|
CPU
& chipset cooling
AMD's
retail Opteron package's airsink has been the copper base Ajigo MF043-44
- this follows their
design guide *.pdf specification for a 'High Performance Heatsink'
extremely closely, & has a respectable C/W of 0.31 or better. An issue
for some or many users may be that this fansink has a thin (15mm]
variable-speed 70mm Delta fan - some feel this has an unpleasant tone to the noise
at full speed - & a quieter 70mm fan with superior static pressure
characteristics is difficult to source.
These
fansink & frame assemblies, as secured to the motherboard, are over
100mm long, 75mm wide, & just under 60mm high. The
DK8N comes with a pair of insulated metal backplates -
one shown above
fastened to the underside of its frame - which go under the motherboard,
screwed to & locating plastic retention frames above
(also provided).
The fansinks
are easy to install into the frames, which provide a
very secure mount with the correct clamping pressure via a cam &
spring. If you think you might later upgrade from this default AMD
cooling solution
(some alternatives, such as the one below, use different
backplates) we suggest you do not peel off & use the self-adhesive
pads at the top of the DK8N backplates - once stuck on they're very tough
to remove.
It is
also worth noting when selecting an enclosure that removing the fansinks
from their frames - to, say, replace the heatsink compound - is impossible
without clear access from above to both ends. We suggest more than the
usual care to
remove any unneeded motherboard standoffs or other protrusions on the
motherboard tray which might contact the board or backplates.
|
We
chose to use the Zalman WB2 waterblocks - nicely made, though note
their inlet/exhaust fittings, for all tube sizes, have the same internal diameter of
7mm (just over 1/4") - this will
cause inefficient flow in systems based around larger ID tubing than
6.5~8mm. American users, where common tubing IDs are
3/8" (9.5mm) or greater, may prefer alternative waterblocks
less likely to starve the intake side of high pressure pumps.
|

|
|

|
After
careful measurement of the gross length of our video & 32-bit
PCI cards (190mm or less), we chose to replace the default heatsink fitted to the
NFORCE3 Pro/250 chipset with a Zalman NB47J - the default heatsink
appeared adequate, so this was very likely an unnecessary fuss-pot
precaution - but we suspect anything Iwill put a heatsink on probably
needs it . . . . .
|
CPU
& voltage monitoring
The
Iwill MPX2 we
tested has a feature called 'Iwill Processor
Shelter' - a hardware-level unattended shutdown, set at a danger-level temperature
selected by the user - at 95C, say, to shut the whole shooting-match off
if a CPU-cooling fan died. At present, the DK8N lacks this valuable feature - relying
perhaps on CPU-level thermal tripping when meltdown looms.
We are
most disappointed to see the similar 'Hardware Health Configuration'
in-BIOS feature quoted in the DK8N manual is not enabled in the current
21st July 04 BIOS. It would be a comfort to control unattended shutdown
events, especially when Opterons cost so much . . . .
.
. . an explanation for this may lie in the apparent inaccuracy of in-BIOS
monitoring derived from the onboard Winbond 83627THF monitoring IC via the
NFORCE3 SMB.
Since
no functional in-OS monitoring utility is provided by Iwill or NVIDIA, it is not possible to get a clear idea of
the degree of error. However, temperatures as reported in-BIOS appear
suspiciously unchanged in different (measured) ambient temperatures - all
reports to date appear to show that all DK8N's report 50~ish C, no matter which
model Opterons are
installed, nor how they are cooled, nor which country the report comes from. In addition, the +3.3
voltage line is reported as 3.0x - unlikely when using a PC Power &
Cooling 510W PSU
We
feel these failures to control unattended shutdown events
& provide accurate in-OS monitoring are unacceptable in an expensive workstation
motherboard: NVIDIA have a 'System Utility' which functions in
other NFORCE-based uniprocessor systems - this utility reports nothing of
value in the DK8N, with an irritating message that function relies upon
vendor-specific optimisations.
Since
the Winbond 83627 family of monitoring ICs are well-proven, we feel the
problem is likely to lie with the BIOS &/or the NFORCE3 Pro/250's SMBus: it will be interesting to see just how
rapidly NVIDIA will help Iwill produce an effective solution to this
credibility-level issue.
|
The
usually excellent third-party freeware monitoring utilities such as MotherBoardMonitor &
Speedfan do
not (cannot, due to false reporting from the DK8N) report correct information with the DK8N: Speedfan sets itself up automatically,
while MBM may be set up to use 'Winbond 1' (which does vary in an
approximate relation to changing CPU load) - the other 'Winbond
2 Diode' & 'Winbond 3 Diode' sensors do not vary in a believable
relation to changing CPU load. Here, 'Sensor 3' is using 'Winbond 1'
Both utilities
show improbably low temperatures - these CPUs & the intake to
their radiator are in an ambient temperature of around 29C - &
some weird voltages. Changing the NVIDIA SMBus driver from the
included v4.04 to v4.4 did nothing to correct matters. |

|
|
DK8N
cooling & monitoring - conclusions:
We
chose to use watercooling with our DK8N due to noise & space
considerations - you need a very large enclosure to have enough room for
an adequately quiet air-cooling solution. Many shudder at the thought of
water meeting silicon; if so we recommend reading Joe Citarella's AMD64
& Socket 939/940 fansink reviews at
www.overclockers.com - checking the measured noise rating &
gross dimensions of the cooling solutions.
The
DK8N has been released to sale without user control over heat-related shutdown events or
a working Win32 monitoring utility. We have been unable to find a working
third-party solution, despite contact with the helpful author of the
Speedfan monitoring
utility. Many professional users will see these
issues as a disincentive to specify continuous duty workstations
based on the DK8N.
|
DK8N
memory & memory settings
We
would suggest the first task of any new DK8N owner is to test their
memory using a Memtest
bootable CD or floppy: in order to get correct reporting from this
excellent free utility
you first need to enter the DK8N BIOS & disable 'USB Legacy Support'
- so best to use a PS/2 keyboard for this test.
Despite
suspecting the DK8N might be picky about memory, for performance reasons
we chose OCZ
'PC3200' low-latency 2-3-2-6 DDR - this to our fury had one faulty stick
among the four we purchased; though for once a 'lifetime guarantee' meant
something - OCZ replaced the memory within 48 hours. We
estimate that using 2-3-2-6 'low latency' memory is worth around 5~7%
extra bandwidth over ordinary 2.5-3-3 or 3-3-3 stuff.
We would
anyway suggest
caution when choosing aggressively timed memory: some of it is specified
at 2.7v or more, while the DK8N has no DDR voltage adjustment & is rated
for the Jedec norm of 2.5v. Our OCZ stuff is rated at the outer limit of
Jedec specs at 2.6v; & it is possible that a full load of eight sticks
of this or similar 'low-latency' memory, especially if larger sticks than
we used, might run into voltage-drop issues.
The
only 'low latency' 400MHz Registered/ECC DDR we know of specified to run
at 2.5v is Mushkin stuff:
we do not specifically recommend this brand or model, having received poor
service from them after being sold expensive faulty memory.
Iwill
told us they have so far tested full eight-stick quantities of the following brands
& models of Registered/ECC DDR
in the DK8N:
|
|
Apacer |
Micron |
Kingston |
Transend |
Samsung |
|
DDR400/1GB |
Infineon
HYB25D256400BC-5 78.01058.112 |
|
|
|
|
|
DDR400/512MB |
SAMSUNG
K4H560838E-TCCC AG32L72T8SQC4S |
MT 46V64M4-5B C
& MT 46V16MB-75A |
|
SAMSUNG K4H560838E-TCCC
& MOSEL V5BC2256804SAT5B |
|
|
DDR400/256MB |
SAMSUNG
K4H560838E-TCCC AG32L72T8SQC4S |
|
|
|
|
|
DDR333
512MB |
|
|
SAMSUNG K4H560438E-TCB3 KVR333X72RC25/512 |
|
|
|
DDR333/256MB |
|
|
SAMSUNG K4H560838F-TCCC |
|
|
|
DDR266/1GB |
|
|
|
|
SAMSUNG K4H510638D-TCA2 |
|
DDR266/512MB |
|
|
|
SAMSUNG K4H560438D-TCB0 |
|
In-BIOS
memory settings allow tweaking of timings, otherwise these are read off
the memory's SPD; more importantly, there are various optional ECC
settings available. We chose to enable 'DRAM BG Scrub' (BG=background) at
640ns - this does not have much if any performance cost & is a slight
extra security-blanket.
We
also experimented with enabling 'chip kill' - this is an aggressive scrub
option & adds an extra latency cycle: you will see what if anything
this option 'costs' in the benchmarking section of this review - the
results surprised us.
For
32-bit Windows [non-NUMA] you can enable node interleaving for maximum
performance; for 32-bit & 64-bit NUMA-aware operating systems you need
to
disable node interleaving.
|
DK8N
memory - conclusions
If
you want to use high-performance memory, keep a close eye on
first-hand reports from the usual
forums. We found OCZ 2-3-2-6 stuff to work OK, in a four-stick
dual-channel configuration. The Iwill-tested
sticks are all low or medium performance memory. |
|