VOGONS


First post, by 386SX

User metadata
Rank l33t
Rank
l33t

Hi,

I'm using this Intel D2500HN (Atom D2500@dual 1,86ghz) mini-itx board with a single 4GB DDR3@1066 sodimm ram and obviously I wasn't surprised about its quiet performances (as an in-order execution cpu with no HT) with linux not too much bad on 2D/Office tasks, unfortunately its video subsystem seems not too well supported by drivers and it looks like a basic gfx acceleration is slowing also somethings down.
But yesterday I was running the memtest app at boot and looking at the initial memory test before testing the memory it got like a bench value of 500MB/s while I remember a dual channel DDR400 (!) reached like 1900MB/s. And as I was expecting the time taken to test the whole 4GB memory has been really long.
Is this a chipset performance problem or something is wrong with this ram or mainboard? I never realized that part of the whole slowness could be memory related.
Thanks

Reply 1 of 8, by Deunan

User metadata
Rank l33t
Rank
l33t

A single DIMM stick will bottleneck all modern CPUs - Atoms including. Those require much wider bus, and since the DIMMs can only fit so many chips, you need more slots populated. Not only that, the CPU can then do bank interleaving which also lowers the latency.
This comes at the cost of higher power usage so many low-power systems will come with a single memory slot only - if that's your case, there's nothing you can do. It doesn't help, obviously, that the iGPU will hog a lot of this bandwidth for itself.

Reply 2 of 8, by 386SX

User metadata
Rank l33t
Rank
l33t
Deunan wrote:

A single DIMM stick will bottleneck all modern CPUs - Atoms including. Those require much wider bus, and since the DIMMs can only fit so many chips, you need more slots populated. Not only that, the CPU can then do bank interleaving which also lowers the latency.
This comes at the cost of higher power usage so many low-power systems will come with a single memory slot only - if that's your case, there's nothing you can do. It doesn't help, obviously, that the iGPU will hog a lot of this bandwidth for itself.

Thanks. The board does have double slot for sodimm but I put a 4GB one cause these Atoms can't do dual-channel and I thought that in single channel it wouldn't make much difference. I think I'll look for some higher performances CL5 2x2GB sodimm I've seen some Kingston Hyperx around.
Unfortunately it doesn't have alternatives for the iGPU cause it only support a single PCI old type and I don't have any of those latest pci expensive video cards to accerate somehow gfx tasks helping the cpu.

Reply 3 of 8, by Deunan

User metadata
Rank l33t
Rank
l33t

Ah, you're right, somehow I thought the D2500 is dual-channel. I have N2800 set up as a router and it's single-channel as well but I run this system headless so I don't much care for iGPU performance (or lack of thereof). In that case populating 2 slots probably won't do much, by itself the bank interleaving gets you only 5-10% speedup (and usually it's closer to the lower value). What's worse, my system does not even boot with both slots populated with 2GB sticks, even though it should support up to 4GB total. Again not an issue for my use case but you already have a 4GB stick working. I'd stick to that or look for a newer Z series Atom. Or embedded EPYC, as all Intel iGPUs kinda suck anyway.

Reply 4 of 8, by SPBHM

User metadata
Rank Oldbie
Rank
Oldbie

Single Channel shouldn't be that bad

on AIDA64 read my P4 Northwood with single channel DDR333 on 845PE gets 2.5GB/s
from their database I can see that an Atom 230 with single DDR2 533 gets 3.6GB/s
that a first gen single core (with HT) 1.6GHz Atom, yours shouldn't be all that different

but yes, the IGP is bad, if it's that one they licensed from PowerVR and didn't have good drivers, even the downclocked GMA 950 from 1st gen Atoms is better to use I think

Reply 5 of 8, by 386SX

User metadata
Rank l33t
Rank
l33t
SPBHM wrote:
Single Channel shouldn't be that bad […]
Show full quote

Single Channel shouldn't be that bad

on AIDA64 read my P4 Northwood with single channel DDR333 on 845PE gets 2.5GB/s
from their database I can see that an Atom 230 with single DDR2 533 gets 3.6GB/s
that a first gen single core (with HT) 1.6GHz Atom, yours shouldn't be all that different

but yes, the IGP is bad, if it's that one they licensed from PowerVR and didn't have good drivers, even the downclocked GMA 950 from 1st gen Atoms is better to use I think

That's why I thought something may be wrong also the gma3600 powervr based gpu (yes it is that one) is probably part of the slowness problems. A driver exists on lubuntu x64 and it seems to accelerate but the feeling is not as much as it should. Anyway i found a couple sodimm pc8500 2x2gb hynix and later I'll try these.
If i'll go for a discrete pci gpu i'd need a full case and it would not make much sense anymore this mobo.

Reply 6 of 8, by 386SX

User metadata
Rank l33t
Rank
l33t

UPDATE: unfortunately one of the two sodimm I found today has memtest errors 😵 but one thing is interesting with a single 2GB PC8500 @ 7-7-7-6, memtest shows a 950MB/s value basically double of the previous. I tried again using the 4GB single one and this time gave me 250MB/s????
I search for another memory testing program to check. The memory 4GB is a Crucial one 1,36v and it was new. May it be it's someway incompatible? What about that old low-voltages sodimm compatibility issue?

UPDATE2:

I've tried the mbw memory test program in linux with the usual single-4GB@1066 and these are the results:

AVG Method: MEMCPY Elapsed: 0.03038 MiB: 32.00000 Copy: 1053.415 MiB/s
AVG Method: DUMB Elapsed: 0.01858 MiB: 32.00000 Copy: 1721.976 MiB/s
AVG Method: MCBLOCK Elapsed: 0.02582 MiB: 32.00000 Copy: 1239.489 MiB/s

So I'd suspect that is a memtest problem considering these values, am I right?
Just to complete the bench I add also hardinfo multicore benchmark values:

CPU Blowfish: 11.23
CPU Cryptohash: 82.56
CPU Fibonacci: 3.12 (faster than Core2 E6750 !)
CPU N-Queens: 24.72
FPU FFT 15.99
FPU Raytracing: 16.53 (similar to Core2 T5750)
GPU drawing: 3545.59 Himarks (interesting test need to search some reference values to compare)

Reply 7 of 8, by Deunan

User metadata
Rank l33t
Rank
l33t

Yeah, these values look much more sane. But chances are memtest is really that slow - I mean it's running on some microkernel to be able to access all that memory and display stuff. Could be that code has some issues with particular memory configurations and/or systems. But that doesn't mean a full-blown modern OS will have the same issues as you see.

Reply 8 of 8, by 386SX

User metadata
Rank l33t
Rank
l33t
Deunan wrote:

Yeah, these values look much more sane. But chances are memtest is really that slow - I mean it's running on some microkernel to be able to access all that memory and display stuff. Could be that code has some issues with particular memory configurations and/or systems. But that doesn't mean a full-blown modern OS will have the same issues as you see.

Thanks. I've just put the mainboard itself into my other only case, a micro-atx one cause I was bored of the mini-itx case limitation even if in this way a better mainboard would make more sense but anyway..
Until I will find some good very low power pci modern video card I already put a Xonar D1 pci sound card to make that pci connector useful and it seems my picoPSU with the 60W external power supply is still warm hope it will be enough with a discrete graphic card or in alternative I'll user an internal psu.