SIS 630E perf problems?

SIS 630E perf problems?

[Larry McVoy]

I use bookpcs - all in one, really nice form factor - for build machines,
firewalls (with a USB ethernet), etc.  I used the first generation which
had intel i810 graphics (sucked) but had fairly typical performance, 
competitive for kernel builds with other current platforms at the time.

I recently bought a couple of the second generation of these boxes, these
have an SIS 630E based motherboard.  This has a much better graphics interface,
quite reasonable at 1280x1024, and all the other bits work fine under RH 7.1
without tweaking.

Performance sucks, however.  I did an LMbench run to try and figure out why
and it's obvious - the memory latencies are 430 ns - that's 2x more than
what is reasonable.  I tweaked the various bios settings a bit and could
not get it to change much, maybe 20ns but not the 200ns I was looking for.
The fact that this system is running a celeron with a dinky cache makes it
feel really slow.  These boxes with a 633Mhz celeron feel slower than the
old boxes with a 400Mhz celeron.

All I want to know is if this is in fact the real memory latency for these
motherboards.  Anyone know for sure?

Here's what I could dig out of /proc:

PCI bus devices:
    Host bridge: Silicon Integrated Systems [SiS] 630 Host (rev 32).
    IDE interface: Silicon Integrated Systems [SiS] 5513 [IDE] (rev 208).
    ISA bridge: Silicon Integrated Systems [SiS] 85C503/5513 (rev 0).
    Ethernet controller: Silicon Integrated Systems [SiS] SiS900 10/100 Ethernet (rev 129).
    USB Controller: Silicon Integrated Systems [SiS] 7001 (rev 7).
    USB Controller: Silicon Integrated Systems [SiS] 7001 (#2) (rev 7).
    PCI bridge: Silicon Integrated Systems [SiS] 5591/5592 AGP (rev 0).
    Multimedia audio controller: C-Media Electronics Inc CM8738 (rev 16).
    Communication controller: C-Media Electronics Inc CM8738 (rev 16).
    VGA compatible controller: Silicon Integrated Systems [SiS] SiS630 GUI Accelerator+3D (rev 32).

Re: SIS 630E perf problems?

[Paul Flinders]

Larry McVoy wrote:

>I use bookpcs - all in one, really nice form factor - for build machines,
>firewalls (with a USB ethernet), etc.  I used the first generation which
>had intel i810 graphics (sucked) but had fairly typical performance, 
>competitive for kernel builds with other current platforms at the time.
>
>I recently bought a couple of the second generation of these boxes, these
>have an SIS 630E based motherboard.  This has a much better graphics interface,
>quite reasonable at 1280x1024, and all the other bits work fine under RH 7.1
>without tweaking.
>
>Performance sucks, however.  I did an LMbench run to try and figure out why
>and it's obvious - the memory latencies are 430 ns - that's 2x more than
>what is reasonable.  I tweaked the various bios settings a bit and could
>not get it to change much, maybe 20ns but not the 200ns I was looking for.
>The fact that this system is running a celeron with a dinky cache makes it
>feel really slow.  These boxes with a 633Mhz celeron feel slower than the
>old boxes with a 400Mhz celeron.
>
If there is a problem it appears to be OS, and even processor independent.

I have one of these and had made essentially the same observation as Larry
- that memory bandwidth seems extremely poor.

Larry's mail prompted me to have another look so I installed Win 2k and
ran the Sandra benchmarks - which came up with a memory bandwidth of
130MB/s, about 1/3 of it's reference value for a SiS 630S.

Disk throughput was poor as well, about 8MB/s for a 10G 5400 Seagate
running in Ultra DMA 66. Presumably this is secondary to the lousy memory
bandwidth and (?) points to a bottleneck between chipset & RAM

Mine has a 667Mhz Citrix (Samual I core) and normally runs RH 7.1,

Does anybody else have one?

Re: SIS 630E perf problems?

[Rob Landley]

On Monday 06 August 2001 18:06, Paul Flinders wrote:
> Larry McVoy wrote:
> >I use bookpcs - all in one, really nice form factor - for build machines,
> >firewalls (with a USB ethernet), etc.  I used the first generation which
> >had intel i810 graphics (sucked) but had fairly typical performance,
> >competitive for kernel builds with other current platforms at the time.
> >
> >I recently bought a couple of the second generation of these boxes, these
> >have an SIS 630E based motherboard.  This has a much better graphics
> > interface, quite reasonable at 1280x1024, and all the other bits work
> > fine under RH 7.1 without tweaking.
> >
> >Performance sucks, however.  I did an LMbench run to try and figure out
> > why and it's obvious - the memory latencies are 430 ns - that's 2x more
> > than what is reasonable.  I tweaked the various bios settings a bit and
> > could not get it to change much, maybe 20ns but not the 200ns I was
> > looking for. The fact that this system is running a celeron with a dinky
> > cache makes it feel really slow.  These boxes with a 633Mhz celeron feel
> > slower than the old boxes with a 400Mhz celeron.
>
> If there is a problem it appears to be OS, and even processor independent.
>
> I have one of these and had made essentially the same observation as Larry
> - that memory bandwidth seems extremely poor.
>
> Larry's mail prompted me to have another look so I installed Win 2k and
> ran the Sandra benchmarks - which came up with a memory bandwidth of
> 130MB/s, about 1/3 of it's reference value for a SiS 630S.
>
> Disk throughput was poor as well, about 8MB/s for a 10G 5400 Seagate
> running in Ultra DMA 66. Presumably this is secondary to the lousy memory
> bandwidth and (?) points to a bottleneck between chipset & RAM
>
> Mine has a 667Mhz Citrix (Samual I core) and normally runs RH 7.1,
>
> Does anybody else have one?

I've used a few funky SIS chipsets, on and off, for a long time now, and they 
always have one leeeetle problem...

Try benchmarking it with a lower screen resolution (like 640x480x256 colors). 
 If the video is sharing main memory, it's sharing the memory bandwidth as 
well.  So you've basically got a constant ultra-high-priority DMA going to 
the screen, sucking up bandwidth and fighting with everything else.  
(Everything else MUST lose or the display would sparkle.  
1280x1024x32bitsx70hz is HOW much bandwidth we're talking here?)

The resulting leftover main memory bandwidth is going to SUCK.  Badly.  And 
the higher your resolution, the worse it gets.

The only possible excuse for it is that it's really really cheap to do, and I 
suppose saves space on the motherboard.  And if it's a game box where FPS is 
your primary concern than at least updating screen memory isn't constrained 
by PCI or AGP bus bandwidth...

(P.S.  When your i810, which also sucks a video signal out of main memory, 
had performance you considered not-as-sucky, was your screen resolution 
perchance less than it is now?  Pixels (and hence memory bandwidth 
requirements) increase exponentially with resolution, dontcha know.  And a 
greater color depth or more hertz aren't improving matters either...)

Rob

Re: SIS 630E perf problems?

[Larry McVoy]

> I've used a few funky SIS chipsets, on and off, for a long time now, and they 
> always have one leeeetle problem...
> 
> Try benchmarking it with a lower screen resolution (like 640x480x256 colors). 
>  If the video is sharing main memory, it's sharing the memory bandwidth as 
> well.  So you've basically got a constant ultra-high-priority DMA going to 
> the screen, sucking up bandwidth and fighting with everything else.  
> (Everything else MUST lose or the display would sparkle.  
> 1280x1024x32bitsx70hz is HOW much bandwidth we're talking here?)

OK, a copuple of updates on this:

I wasn't running X when I ran the benchmarks.

I played around with the bios settings enough to make the machine not
pass POST anymore so I reset the CMOS.  Doing that, plus telling the
system to autodetect DRAM clocks dropped the latencies down to 260ns
outside of X and 281ns with X running.  Still not fantastic but good
enough I suppose.

In reference to the fans that someone else mentioned: I think it is the
CPU fan making the noise.  Regardless, the MTBF of the power supply fans
is one year.  I have about a dozen of generation 1 of these bookpcs and
the fans all started failing at 1 year and frying the power supplies.
Those dinky power supplies are hard to find so you want to avoid this.
We bought a pile of fans and replaced them all; one benefit is that the
higher quality fans are much less noisy.  Something to look into.
-- 
---
Larry McVoy            	 lm at bitmover.com           http://www.bitmover.com/lm 

Re: SIS 630E perf problems?

[Rob Landley]

On Tuesday 07 August 2001 11:16, Larry McVoy wrote:
> > I've used a few funky SIS chipsets, on and off, for a long time now, and
> > they always have one leeeetle problem...
> >
> > Try benchmarking it with a lower screen resolution (like 640x480x256
> > colors). If the video is sharing main memory, it's sharing the memory
> > bandwidth as well.  So you've basically got a constant
> > ultra-high-priority DMA going to the screen, sucking up bandwidth and
> > fighting with everything else. (Everything else MUST lose or the display
> > would sparkle.
> > 1280x1024x32bitsx70hz is HOW much bandwidth we're talking here?)
>
> OK, a copuple of updates on this:
>
> I wasn't running X when I ran the benchmarks.
>
> I played around with the bios settings enough to make the machine not
> pass POST anymore so I reset the CMOS.  Doing that, plus telling the
> system to autodetect DRAM clocks dropped the latencies down to 260ns
> outside of X and 281ns with X running.  Still not fantastic but good
> enough I suppose.

Text mode still comes out of main memory, and is still a fairly consistent 
DMA out of it.  I suspect that for every pixel displayed, text mode or not, 
it still has to hit main memory.  Text mode goes through a lookup table 
thingy to max the character against the pixel grid for that character, but I 
suspect it still re-reads the value from main memory each time.

Even if it doesn't and instead caches the value (which, this being SIS, is 
probably wishful thinking: if that was the case they'd just devote 16k or so 
to the text mode circuitry and not have text mode share main memory at ALL, 
which I'm almost certain is NOT the case.  Again, optimizing for text mode 
ain't what a "for windows" hardware company thinks about, even when it's 
easy...)

Anyway, it's still doing nasty things with opening/closing memory banks 
(fighting with what would otherwise be linear burst transfers: we interrupt 
this DMA to seek to somewhere else in memory.  Yes, SIMMS have something like 
a hard drive seek, sending new addresses and opening/closing banks.  Ars 
technica had a nice article on this a while back I could dig up a link to if 
anybody's bored.  It's 1000 times faster than a hard drive seek, but it's 
still there.  The old latency vs throughput issue. Locality of reference is a 
good thing no matter what's going on.  And yes, making this go away and doing 
prefetch (and branch predictions, etc) are half of what the L1 and L2 caches 
do.  And the bus interface units did at least a few bytes read ahead all the 
way back to the 8086, actually.  (6 bytes for the 8086, 4 bytes for the 
8088.  Yes, I'm still writing a book on computer history.  Research 
crystalizing into chapters Real Soon Now (tm).  If I could just stop finding 
MORE info leading to fresh tangents to go down for a bit...  Or if the UT 
library didn't close between midnight and 8 am over the summer session...)

Rob