GTX 950, GTX 960 segmented ram systems as well ?

Hello,

Here is a truely computer architectural question for you:

Do NVIDIA's Graphics Cards/Chips models GTX 950 and GTX 960 have segmented
memory models as well ? Just like the GTX 970 ?

I post on thread on this on geforce forum but it mysteriously disappeared...
other delete or again a forum bug.

Still like to get to the bottom of this since I may be interested in a new
graphics card now or in the future.

I ask people to run my benchmark at this site:

http://www.skybuck.org/CUDA/Bandwidt...200.16/Packed/

Downoad and extract it to a folder on file system, then run the B32 or B64
executable on respecitively 32 bit or 64 bit operating system.

Push start button to start program. (If it takes a bit long try reducing
estimated bandwidth to half).

Save the chart and log to file and send it to
[s][k][y][b][u][c][k][2][0][0][0][@][h][o][t][m][a][i][l][.][c][o][m]

without the []

Anyway so far one person responded, a person interested in benchmarking
these graphics cards and had many youtube videos up of game benchmarks.

The graph/chart did not look good for the GTX 960, 27 memory blocks of 128
MB were allocated.

This means only 3456 MB was available to CUDA for some reason. This reminds
me of GTX 970 which also had this strange segmented memory system.

First some basic information about the memory architecture.

Each memory controller is 64 bit wide, which has little to do with it, but
ok.

The GTX 970 apperently has 4 memory controllers. The GTX 960 has 2 memory
controllers.

The reads/writes can be coalesced/grouped together.. so that's why this
might not be that important.

Each memory controller seems to be connected to a caching system L2 cache
and the L2 cache connected to GPC groups.

GPC groups contain SMM units.

Currently each GPC group contains 8 SSM units.

Apperently these can be considered the cores of the machine/chip.

Now let's consider the GTX 970/980:

It looks like it has 32 of these SSM units.

However looking a bit more closely... there seem to be 2 vertically below
each other but these seem to be counted as only 1.

So 16 SSM units.

Let's suppose there is a 4 GB memory system.

It's supposed to be a parallel system, so these 16 units have to read the
memory in parallel ?! How to do it ?

Also keep in mind the caching system.

Because of the caching system and perhaps the way the memory
controllers/memory works it's beneficial to for example do it as follows:
0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15, repeat.

This means every SSM unit can read from it's own offset.

What this means it the total addressing space of 4 GB needs to be divided by
16 to get some idea of the maximum ammount of blocks that be present in the
system.

4096 / 16 = 256

So this means 256 MB memory blocks.

16x256 = 4096.

So there should be 16 memory chips.

Is this is indeed the case remains doubtfull but is also beside the point.

Apperently only 4 memory controllers which mimic this 16 memory chip design.
Perhaps it offers a bit more flexibility for larger data
structure/requirements...

Anyway...

Now ask yourself question what happens if a couple of these SSM units fall
out/are broken/defective ?

Apperently NVIDIA has found a way to still keep these chips functioning and
wents to sell them... problem is they lie about these broken/defective chips
which is not good.

So let's examine a broken GTX 980=GTX 970

3 are kaputt. So this leave 13 functioning SSM units and also perhaps a
broken L2 cache section.

Those ssm units that are broken can no longer participate in calculations so
memory access has to be re-ordered as follows:

0,1,2,3,4,5,6,7,8,9,10,11,12, repeat.

So only 13 functional.

But since the whole thing was designed for 256 memory blocks this gives the
following:

13x256 = 3328 MB.

This leaves the upper 3x256 = 768 in a broken state.

Not sure if this theory is correct but could be.

Now onto the GTX 960:

This thing has 8 of these SSM units... like a GTX 980 chopped in half.

Though it also has 4 GB of ram... so somehow that must be divided as well it
seems:

4096 / 8 = 512 MB blocks.

Now suppose one goes kaputt:

7x512 = 3584 MB.

Again a segmented memory system.

Perhaps nvidia is selling of defective GTX 960... or perhaps it's just GTX
970 cut in half.....

That kinda funny... if you think about it.. they could maybe double their
profits lol. Though if it's actually possibly to cut a GTX 970 in half
literally and still use it... I don't know

Though this posting tries to explain why a segmented memory system might
still be possible and that it might be unrelated to memory controlers and
their bit width.

It has more to do with the way the SSM units are addressing memory, the
logic behind it.

Perhaps they are hard-wired... it must be... otherwise it might have been
less of a problem.

So they seem to be basically hardwire to address the memory system in a
certain parallel way.. and it's inflexible.

Perhaps future models will have some flexibility build in.. such that these
segmented memory systems might go away.

I am not necessarily against re-using defective/broken chips if some parts
of it still function... but it should be made clear that it's so... and not
lied about.

I don't like it if I would buy one of these graphics cards and then later
find out that they cause all kinds of stutter when games go over the 3.5 GB
limit in memory usage.

That is why I may have to pass up on those cards and wait for something
better to come along or perhaps try a card of the competitor AMD which does
not have CUDA support, so also there AMD shoots themselfes in the foot
somewhat.

First I'd like to know the thruth about these cards, much more and many more
people need to investigate and run these kinds of benchmarks to get a better
idea of what is going on.

Though I can also understand if people don't care but then later don't come
bitching your system is lagging/stuttering !

Less demanding games might run fine... or they might not... depending on
where the memory is allocated in this ram system.

Also concerning GTX 960 weird 3.5 GB limitation of 4 GB ram might have to do
with drivers ? very maybe ? though there is a distinct pattern here !

Bye,
Skybuck.

One question which needs explanation:

If the cores/SSM are truely hardwired to their own offsets, then how is it
possible for a defective chip to still function with 3.5 GB ram ?

If SSM units fall out/are defective... and this can happen anywhere.. then
this would create gaps, example:

0,1,2,3,X,5,6,X,8,9,10,11,12,X,14,15

I randomly disabled some SSM units indicated by X.

The solution seems simple enough:

Disable memory chips which are associated/hard-wired to X.

And allocate memory in such a way that these gaps are avoided ?

Perhaps this is what is going on internally in the driver or so.

Bye,
Skybuck.

On Wed, 30 Sep 2015 14:33:30 +0200, "Skybuck Flying"
Gave us:

I post on thread on this on geforce forum but it mysteriously disappeared...
other delete or again a forum bug.

You're an idiot. The NVidia forum is the ONLY forum you should have
posted your inane question to.

On Wed, 30 Sep 2015 14:33:30 +0200, "Skybuck Flying"
Gave us:
[i]
Save the chart and log to file and send it to
[s][k][y][b][u][c][k][2][0][0][0][@][h][o][t][m][a][l][.][c][o][m]

without the []

Do you mean ???

On 2015-09-30 08:33, Skybuck Flying wrote:

Hello,

Here is a truely computer architectural question for you:

Do NVIDIA's Graphics Cards/Chips models GTX 950 and GTX 960 have
segmented memory models as well ? Just like the GTX 970 ?

Good question; I haven't looked into it, but maybe they learned from
their "mistake" on the 970 and changed the design a little? I'd google a
little for you, but since the 980Ti came out, making the plain 980 more
affordable, I haven't looked at anything else.

Still like to get to the bottom of this since I may be interested in a
new graphics card now or in the future.

I ask people to run my benchmark at this site:
http://www.skybuck.org/CUDA/Bandwidt...200.16/Packed/
Downoad and extract it to a folder on file system, then run the B32 or
B64 executable on respecitively 32 bit or 64 bit operating system.

I don't think a lot of people will run it, what with security and all;
could be a virus, lol ;-)

[snip]

--
! _\|/_ Sylvain /
! (o o) Member-+-David-Suzuki-Fdn/EFF/Red+Cross/Planetary-Society-+-
oO-( )-Oo DOS Tip #17: Add DEVICE=FNGRCROS.SYS to CONFIG.SYS.

"B00ze" wrote in message ...

On 2015-09-30 08:33, Skybuck Flying wrote:

Hello,

Here is a truely computer architectural question for you:

Do NVIDIA's Graphics Cards/Chips models GTX 950 and GTX 960 have
segmented memory models as well ? Just like the GTX 970 ?

"
Good question; I haven't looked into it, but maybe they learned from
their "mistake" on the 970 and changed the design a little? I'd google a
little for you, but since the 980Ti came out, making the plain 980 more
affordable, I haven't looked at anything else.
"

Perhaps GTX 980 is trustable.

The rest is doubtfull... even the Titan X can cause troubles with not enough
main system ram.

NVIDIA will have to step up and be more forth coming with these kinds of
problems.

Also seeing CUDA not being able to allocate all RAM is highly suspicious !


Gonna wait this one out until things become more clear or new products...

As far as GTX brand is concerned it's damage beyond repair.

Still like to get to the bottom of this since I may be interested in a
new graphics card now or in the future.

I ask people to run my benchmark at this site:
http://www.skybuck.org/CUDA/Bandwidt...200.16/Packed/
Downoad and extract it to a folder on file system, then run the B32 or
B64 executable on respecitively 32 bit or 64 bit operating system.

"
I don't think a lot of people will run it, what with security and all;
could be a virus, lol ;-)
"

Yeah that's a bit of a problem for independant software developers.

Perhaps there should be some chain of trust for independant software.. a
little bit like the www.thepiratebay.org

Anyway for the fun of it I ran it through virus scan total... funny enough
it was already scanned a couple of months ago or so... had it re-analyzed.

It comes up clean ! Pfew !

https://www.virustotal.com/en/file/8...is/1443736205/

Bye,
Skybuck.