On 4/16/2024 10:44 PM, Newyana2 wrote:
On 4/16/2024 5:44 PM, Paul wrote:
If the video card plays mpeg2 or H.264, chances are better it will
work and have some quality.
My laptop for example, has a single core, and is little better than
a Pentium 4. But it still plays video, because the chipset has
a decoder. But the laptop isn't good at much else. The CPU sees
to that (it is like a 64-bit version of the AthlonXP ).
Would it be accurate to say the newer the graphics, the more
it's taking over the work, or does it really depend on the hardware?
The 710 seems to handle streaming fine, but actually it's been better
since awhile back when I discovered that I had some old RAM that
fit that machine, jackinh up the total from 4 to 6 GB. Before the
machine was slightly groggy and video would occasionally stutter.
The computer is a Dell XPS625. My father ordered it custom back
in 2010, but then at some point his mental functioning deteriorated
to the point that he could no longer use it. So now it's one of my
streaming boxes.
First you would want to review the BIOS settings.
I haven't committed all these things to memory :-)
There was some deal with AMD at one point, where you
could decouple dual-channel. This allowed cores (like on
your quad core) to make independent requests on a single
channel. Ah, yes, it was called "Unganged". It could depend
on what the machine is doing, as to which option is better.
A lot of software is single core and does not take
advantage of multiple cores, and leaving the RAM in dual
channel mode may help. On a server (multiple uncorrelated processes),
maybe Unganged is better.
At one time, the AMD memory controller was not very sophisticated.
Intel had Flex memory, whereas AMD did not bother with this. It
would have made their test bench too large to test. With Intel,
it didn't matter which holes you put the RAM in, the Intel could
pair them you virtually in a sense. The AMD was too clumsy for
this, and required to put stuff in the correct holes. Nobody
would tell you that you made a mistake.
Ch0 Ch1
| |
1GB 2GB <---- Not matched across, preferred to be matched horizontally
2GB 1GB This is 6GB in single channel mode (a bit slow).
You won't like my drawing -- consult user manual for diagram :-)
Ch0 Ch1
| |
2GB 2GB <---- Now we have a better-performing setup. It can't interleave
1GB 1GB between rows quite as well as it might, but it is still a lot better.
Some things will be more snappy (partly because AMD cache at one time
was a little bit slow, and there was more of a dependency on RAM
performance because of the hit rate. The Intel design seemed to
"hide" issues with RAM better). Now it's dual channel mode. It can
rock back and forth between rows, but the pattern will be a bit
asymmetric.
Ch0 Ch1
| |
2GB 2GB <---- Depending on the stride through memory, when you access a block of
2GB 2GB locations, it can alternate between the two rows. This gives several
percent more performance. You can keep pages open on two ranks
on each DIMM, times the number of DIMMs, times the number of banks inside
the DIMM. Intel had a name for this, that I forget, this mode
(symmetric something-or-other). It's still dual channel, only with less
waiting for precharge or something.
But something might also be going on with VRAM on the video card, versus an allocation of system memory to help the VRAM. Otherwise, I don't know
what would be clogging it up. If my video card has VRAM, perhaps I would configure things so only that VRAM was on offer, and I would not want it playing with my system RAM. Some of the low end AMD cards, could use a little of both. And some cards didn't have local VRAM at all, and using system RAM wasn't helping matters (when it's DDR2 and a bit slow).
With things like Bulldozer and Piledriver, Intel style hyperthreading versus the sort of hybrid approach AMD used (Y shaped major axis, the bottom of the
Y is a shared resource, the top is duplicated stuff and might have beat
Intel given how Intel's worked). You need a "CPU driver" to get the
best from the hardware (might have been Scheduler related). Sometimes they didn't tell people this, and the customers did not know that a little extra performance would come their way, if they tracked down the driver on the AMD.com site.
Maybe if I was an Asus customer, I would not expect to find that driver
on their site, and I would have to spelunk the AMD site instead. Which
is harder than you would think. This is the kind of stuff that
should have been bundled into a Windows Update. And perhaps today, with the "drivers on Catalog server" thing, it would get done that way. The CPU driver may have been a bit late arriving, which makes it harder for a Dell to put
it on the hard drive on the day of release. It could have been a year late. There was some fighting about money, behind the scenes, and this resulted
in issues of this type. And eventually we had Microsoft indicating "no support for new processors on OS nn", as evidence there was no free lunch available.
Your box is tune-able, but you really needed to be reading the forums
of the time, and keeping your notes file up to date, on the latest
discovered "issue".
Yours is a Phenom II, so you don't have the "half my TLB is dead" issue.
https://www.anandtech.com/show/2489
*******
If you go back a long long ways in time, the only thing the video card
had was IDCT (inverse discrete cosine transform). Lots of video format
(except Cinepak), they use "macro blocks", square groups of pixels, and
the video was modified in the frequency domain. This involves doing
some sort of "trig" to convert the content from the frequency domain
to something else. The video card offered IDCT for hardware acceleration.
This might mean the CPU transferred the block to the video card, pushes a button,
coefficients come back, it copies them out.
This is ridiculously CPU intensive. My software "buddy" at work, I'd add
some accelerator to the hardware when I designed something, and she'd bench
the damn thing and tell me it was shit (but in a polite way). And so IDCT probably fell into that area, appreciated that someone tried to do it,
but not enough automation.
Fast forward to today, the video card does everything. It's probably given
a pointer in memory, it uses DMA, loads the video card, some vertex or texture processors do all the math, video loads at a certain offset in the frame buffer,
you're watching a movie. Maybe the CPU finds out, and loads another 64KB into RAM and the CPU goes back to sleep.
if the video is 720x480, the video card can have a scaler to change
the dimensions. It used to take 40% of a Pentium 4 to "do this by hand".
When a hardware scaler was added to the video card, suddenly this was free.
It's the progress of the automation in there, that is the shocker. We've come all the way from "pathetic acceleration", to "series of logic blocks, still
not tied together well", to "runs the show almost completely itself". The
video card likely can't read the disk drive, and so that's where it has
to stop. That is the boundary.
It would take a while, to find out which generation of hardware reached the full automation level. Anandtech stopped evaluating video playback after
a while, because the "engineering tweaking" had stopped. Maybe full automation has been around for a decade or so. A very wild guess. It's sort of the
way that nobody pays any attention to Direct2D performance. Nobody uses
a stopwatch on BitBLT any more :-)
And having MPEG2 and H.264, that subset is enough to keep a person
happy, even if you have to transcode from other formats when building
your library of titles.
Paul
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)
