Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > alt.os.linux.mint > #45565 > unrolled thread

GPU advise?

Back to article view | Back to alt.os.linux.mint

Contents

  GPU advise? Edmund <nomail@hotmail.com> - 2025-10-21 10:46 +0200
    Re: GPU advise? Paul <nospam@needed.invalid> - 2025-10-21 06:27 -0400
      Re: GPU advise? Edmund <nomail@hotmail.com> - 2025-10-21 14:37 +0200
      Re: GPU advise? Peter McD <peter.posts@gmx.net> - 2025-10-23 20:01 +0200
        Re: GPU advise? Edmund <nomail@hotmail.com> - 2025-10-23 23:13 +0200
        Re: GPU advise? Paul <nospam@needed.invalid> - 2025-10-24 00:08 -0400
          Re: GPU advise? Peter McD <peter.posts@gmx.net> - 2025-10-24 10:51 +0200
            Re: GPU advise? Paul <nospam@needed.invalid> - 2025-10-24 09:38 -0400
              Re: GPU advise? Peter McD <peter.posts@gmx.net> - 2025-10-25 13:13 +0200

#45565 — GPU advise?

GPU advise?

Planning to go 4k and interested in Kali.
Don't really care about gaming but Hi(er) frame rate
in 4k on two monitors is a must, any suggestion and why?


ps
I read that although AMD works well nowadays, it doesn't support
something needed for Hashcat (CUDA 5)


-- 
-------------

Godspeed for Assange
Amnesty for Snowden
Rehabilitation for heroes

Edmund

[toc] | [next] | [standalone]

#45566

On Tue, 10/21/2025 4:46 AM, Edmund wrote:
> GPU advise?
> 
> Planning to go 4k and interested in Kali.
> Don't really care about gaming but Hi(er) frame rate
> in 4k on two monitors is a must, any suggestion and why?
> 
> 
> ps
> I read that although AMD works well nowadays, it doesn't support
> something needed for Hashcat (CUDA 5)

2D acceleration support, has not advanced in years.

3D support could be in the form of OpenGL or Vulcan,
and that can uses shaders on a card. The main acceleration
you might see then, is when gaming. For the games that use
accelerated 3D.

CUDA is part of that 3D support. and also runs shader CUDA
programs, where "computation is a side effect of pumping
other constructs through the shader engine". You can crack
passwords, just as easily as playing a game.

The lowest end current generation card is an RTX5050,
which would be gutless, but a bit stronger than a GTX1650.
The RTX5050 might be $250 USD. The RTX5090 might be $2000.
In all cases, the cards don't feel like value for money,
when compared to the price range of some previous generations.

AMD and Intel are OpenCL for computation rather than CUDA.

But just for the record, I've had considerable trouble
getting adventurous programs to use CUDA when it is offered.
A lot of programs will annoy you, by using the CPU version
of OpenCL. And after you've spent a lot of money,
receiving "resistance" from the application software,
is more than a bit unnerving.

I don't think my CUDA ever cracked a password, because
of the problems getting it to engage. Yes, 2GB of libraries
were loaded. I ticked all the boxes for the SDK on both programs.

I did manage to compile the "Smoke Demo" on Windows in Visual Studio
Community Edition, and that used CUDA OK. But some other codes,
like some stuff in KALI, I was getting refusal to proceed with
the CUDA and it just used the CPU (and promised a cracked
password in about 13 years or so).

Here is a table I keep for this. The cards use closed loop feedback
when running the silicon, and the power limit is policed by the
hardware. The driver adjusts the card clock, so the power is not exceeded.
An RTX5060 Ti 16GB for $430, would give 4608 CUDA cores and
be similar to a GTX1080 from 8 to 10 years ago. At least the three
top cards in the list, will require a power supply with a different
connector on it. You'll have to check the prices to re-calibrate my
numbers.

RTX 5090          GB202-300   21,760  32 GB, 512-bit  575 W   $2,000
RTX 5080 Super    GB203-450   10,752  24 GB, 256-bit ~415 W   TBD
RTX 5080          GB203-400   10,752  16 GB, 256-bit  360 W   $1,000
RTX 5070 Ti Super GB203-350    8,960  24 GB, 256-bit  350 W   TBD
RTX 5070 Ti       GB203-200    8,960  16 GB, 256-bit  300 W     $750
RTX 5070 Super    GB205-400    6,400  18 GB, 192-bit  275 W   TBD
RTX 5070          GB205-300-A1 6,144  12 GB, 192-bit  250 W     $550
RTX 5060 Ti       GB205-300    4,608  8/16G, 128-bit  180 W   $380 / $430
RTX 5060          GB206-250    3,840   8 GB, 128-bit  145 W   $300
RTX 5050          GB207        2,560   8 GB, 128-bit  130 W   $250

When you look at a "bazaar seller" listing, like BestBuy, you might
see this.

   RTX5090     $4000
   RTX5090     $2000  <=== retail store price on the BestBuy shelf
   RTX5090     $4000  <=== scalper price, from outside the store
                           They bought all the cards up, to make them scarce.

When you click a button for "show me retail items only" the list
changes to show just the $2000 item.

   RTX5090     $2000

One day I was looking through the list, and I was seeing a sea of
$4000 items, and I was saying to myself "this can't be right", I
checked and I had forgotten to tick the box to cut out the
scalpers. If at first the prices don't seem right, make sure
you've eliminated the dodgy private sellers from the list.

I don't consider cards like this, to be an especially great value.
If you want real horsepower, the prices have gone through the roof.

The cards come in different lengths. The "$250 USD" card is
$400 in local currency here, it has three fans (which is patently
ridiculous). There are also cards with two fans, and they are
more likely to fit in the PC. Not all PCs can accept long cards,
the cards can bump into your Southbridge heatsink or bump
into the SATA right-angle connectors.

The cards are at a minimum, dual slot. Make sure the metalwork
on your case, does not "conflict" with the flat two-slot-wide metal
plate on the faceplate. I have one PC case, where a two slot card
binds when you try to insert it. If I empty the case and take
it into the shop and bash it with a hammer... I can fix that.

A couple of my power supplies, have multiple 2x4 connectors and that
might be sufficient up to 300W. Check the card picture for
evidence of the connector type there. If shopping that table
there, and not going too high up into the table, my 750W PSU would
likely be sufficient. For the top card in the list, I'd have to
go shop for a new 1200W supply with the 16 pin connector (2x6+4, the
4 pins are status/declaration pins). The bottom card in the list
might have a 2x3 or a 2x4 on it, and the PSU that have 2x4
connectors, the end part unsnaps to convert the connector into
a 2x3. All the supplies in the room, have at least one
connector for the bottom card in the list.

The BestBuy would likely have the 5070 and the 5070 Ti, due to
the prices being "mid-range" and that's the range they want
to sell into. My real computer store has a wider stock,
but their stocking levels are not good. For example, when I
wanted an 8TB hard drive for backups, the chain of stores
had exactly one of those drives in stock and that was it.
While they list a lot of products, some listings aren't
all that real.

You can use the listing here, to compare horsepower on the cards.
The amount of VRAM is important, but this benchmark does not
take that into account. For games, you want a decent amount,
which is why I picked out the 16GB card to cover gaming.
The $430 card then, the 16GB one, it's a way to get a 16GB
card, with some number of CUDA on it.

  https://www.videocardbenchmark.net/high_end_gpus.html

The utility here, does not depict all the hardware resources
in a card. But this is just to show you what you might see
after buying one.

  # Review (not all cards get a review)

  https://www.techpowerup.com/review/msi-geforce-rtx-5060-ti-gaming-16-gb/44.html

  # Most cards have a spec summary on this site.

  https://www.techpowerup.com/gpu-specs/geforce-rtx-5060-ti-8-gb.c4246

     Display outputs include: 1x HDMI 2.1b, 3x DisplayPort 2.1b

The Wiki gives some info on resolution support.

  https://en.wikipedia.org/wiki/HDMI

     HDMI 2.1 ... including 4K 120 Hz and 8K 60 Hz

There was one case, where a card promised the 2.1 spec
on NVidia, and the owner could not get the card to switch
to 4K 120Hz. If you're shooting for a very high FPS at
high resolution, you'll have to Google and see if anyone
managed to run the rate you wanted to run. But for a vanilla
4K monitor like mine, cards now will run 4K 60FPS no problem.
It was some older Intel iGPU that might have salted in
some 4K 30FPS numbers.

    Paul

[toc] | [prev] | [next] | [standalone]

#45567

On 10/21/25 12:27, Paul wrote:
> On Tue, 10/21/2025 4:46 AM, Edmund wrote:
>> GPU advise?
>>
>> Planning to go 4k and interested in Kali.
>> Don't really care about gaming but Hi(er) frame rate
>> in 4k on two monitors is a must, any suggestion and why?
>>
>>
>> ps
>> I read that although AMD works well nowadays, it doesn't support
>> something needed for Hashcat (CUDA 5)
> 
> 2D acceleration support, has not advanced in years.
> 
> 3D support could be in the form of OpenGL or Vulcan,
> and that can uses shaders on a card. The main acceleration
> you might see then, is when gaming. For the games that use
> accelerated 3D.
> 
> CUDA is part of that 3D support. and also runs shader CUDA
> programs, where "computation is a side effect of pumping
> other constructs through the shader engine". You can crack
> passwords, just as easily as playing a game.
> 
> The lowest end current generation card is an RTX5050,
> which would be gutless, but a bit stronger than a GTX1650.
> The RTX5050 might be $250 USD. The RTX5090 might be $2000.
> In all cases, the cards don't feel like value for money,
> when compared to the price range of some previous generations.
> 
> AMD and Intel are OpenCL for computation rather than CUDA.
> 
> But just for the record, I've had considerable trouble
> getting adventurous programs to use CUDA when it is offered.
> A lot of programs will annoy you, by using the CPU version
> of OpenCL. And after you've spent a lot of money,
> receiving "resistance" from the application software,
> is more than a bit unnerving.
> 
> I don't think my CUDA ever cracked a password, because
> of the problems getting it to engage. Yes, 2GB of libraries
> were loaded. I ticked all the boxes for the SDK on both programs.
> 
> I did manage to compile the "Smoke Demo" on Windows in Visual Studio
> Community Edition, and that used CUDA OK. But some other codes,
> like some stuff in KALI, I was getting refusal to proceed with
> the CUDA and it just used the CPU (and promised a cracked
> password in about 13 years or so).
> 
> Here is a table I keep for this. The cards use closed loop feedback
> when running the silicon, and the power limit is policed by the
> hardware. The driver adjusts the card clock, so the power is not exceeded.
> An RTX5060 Ti 16GB for $430, would give 4608 CUDA cores and
> be similar to a GTX1080 from 8 to 10 years ago. At least the three
> top cards in the list, will require a power supply with a different
> connector on it. You'll have to check the prices to re-calibrate my
> numbers.
> 
> RTX 5090          GB202-300   21,760  32 GB, 512-bit  575 W   $2,000
> RTX 5080 Super    GB203-450   10,752  24 GB, 256-bit ~415 W   TBD
> RTX 5080          GB203-400   10,752  16 GB, 256-bit  360 W   $1,000
> RTX 5070 Ti Super GB203-350    8,960  24 GB, 256-bit  350 W   TBD
> RTX 5070 Ti       GB203-200    8,960  16 GB, 256-bit  300 W     $750
> RTX 5070 Super    GB205-400    6,400  18 GB, 192-bit  275 W   TBD
> RTX 5070          GB205-300-A1 6,144  12 GB, 192-bit  250 W     $550
> RTX 5060 Ti       GB205-300    4,608  8/16G, 128-bit  180 W   $380 / $430
> RTX 5060          GB206-250    3,840   8 GB, 128-bit  145 W   $300
> RTX 5050          GB207        2,560   8 GB, 128-bit  130 W   $250
> 
> When you look at a "bazaar seller" listing, like BestBuy, you might
> see this.
> 
>     RTX5090     $4000
>     RTX5090     $2000  <=== retail store price on the BestBuy shelf
>     RTX5090     $4000  <=== scalper price, from outside the store
>                             They bought all the cards up, to make them scarce.
> 
> When you click a button for "show me retail items only" the list
> changes to show just the $2000 item.
> 
>     RTX5090     $2000
> 
> One day I was looking through the list, and I was seeing a sea of
> $4000 items, and I was saying to myself "this can't be right", I
> checked and I had forgotten to tick the box to cut out the
> scalpers. If at first the prices don't seem right, make sure
> you've eliminated the dodgy private sellers from the list.
> 
> I don't consider cards like this, to be an especially great value.
> If you want real horsepower, the prices have gone through the roof.
> 
> The cards come in different lengths. The "$250 USD" card is
> $400 in local currency here, it has three fans (which is patently
> ridiculous). There are also cards with two fans, and they are
> more likely to fit in the PC. Not all PCs can accept long cards,
> the cards can bump into your Southbridge heatsink or bump
> into the SATA right-angle connectors.
> 
> The cards are at a minimum, dual slot. Make sure the metalwork
> on your case, does not "conflict" with the flat two-slot-wide metal
> plate on the faceplate. I have one PC case, where a two slot card
> binds when you try to insert it. If I empty the case and take
> it into the shop and bash it with a hammer... I can fix that.
> 
> A couple of my power supplies, have multiple 2x4 connectors and that
> might be sufficient up to 300W. Check the card picture for
> evidence of the connector type there. If shopping that table
> there, and not going too high up into the table, my 750W PSU would
> likely be sufficient. For the top card in the list, I'd have to
> go shop for a new 1200W supply with the 16 pin connector (2x6+4, the
> 4 pins are status/declaration pins). The bottom card in the list
> might have a 2x3 or a 2x4 on it, and the PSU that have 2x4
> connectors, the end part unsnaps to convert the connector into
> a 2x3. All the supplies in the room, have at least one
> connector for the bottom card in the list.
> 
> The BestBuy would likely have the 5070 and the 5070 Ti, due to
> the prices being "mid-range" and that's the range they want
> to sell into. My real computer store has a wider stock,
> but their stocking levels are not good. For example, when I
> wanted an 8TB hard drive for backups, the chain of stores
> had exactly one of those drives in stock and that was it.
> While they list a lot of products, some listings aren't
> all that real.
> 
> You can use the listing here, to compare horsepower on the cards.
> The amount of VRAM is important, but this benchmark does not
> take that into account. For games, you want a decent amount,
> which is why I picked out the 16GB card to cover gaming.
> The $430 card then, the 16GB one, it's a way to get a 16GB
> card, with some number of CUDA on it.
> 
>    https://www.videocardbenchmark.net/high_end_gpus.html
> 
> The utility here, does not depict all the hardware resources
> in a card. But this is just to show you what you might see
> after buying one.
> 
>    # Review (not all cards get a review)
> 
>    https://www.techpowerup.com/review/msi-geforce-rtx-5060-ti-gaming-16-gb/44.html
> 
>    # Most cards have a spec summary on this site.
> 
>    https://www.techpowerup.com/gpu-specs/geforce-rtx-5060-ti-8-gb.c4246
> 
>       Display outputs include: 1x HDMI 2.1b, 3x DisplayPort 2.1b
> 
> The Wiki gives some info on resolution support.
> 
>    https://en.wikipedia.org/wiki/HDMI
> 
>       HDMI 2.1 ... including 4K 120 Hz and 8K 60 Hz
> 
> There was one case, where a card promised the 2.1 spec
> on NVidia, and the owner could not get the card to switch
> to 4K 120Hz. If you're shooting for a very high FPS at
> high resolution, you'll have to Google and see if anyone
> managed to run the rate you wanted to run. But for a vanilla
> 4K monitor like mine, cards now will run 4K 60FPS no problem.
> It was some older Intel iGPU that might have salted in
> some 4K 30FPS numbers.
> 
>      Paul

WOW, Thanks a lot Paul, you are the best.


-- 
-------------

Godspeed for Assange
Amnesty for Snowden
Rehabilitation for heroes

Edmund

[toc] | [prev] | [next] | [standalone]

#45596

Am 21.10.25 um 12:27 schrieb Paul:
> On Tue, 10/21/2025 4:46 AM, Edmund wrote:
>> GPU advise?
...
> A couple of my power supplies, have multiple 2x4 connectors and that
> might be sufficient up to 300W. Check the card picture for
> evidence of the connector type there. If shopping that table
> there, and not going too high up into the table, my 750W PSU would
> likely be sufficient. For the top card in the list, I'd have to
> go shop for a new 1200W supply with the 16 pin connector (2x6+4, the
> 4 pins are status/declaration pins). The bottom card in the list
> might have a 2x3 or a 2x4 on it, and the PSU that have 2x4
> connectors, the end part unsnaps to convert the connector into
> a 2x3. All the supplies in the room, have at least one
> connector for the bottom card in the list.
> 
...
Check power requirements of the card before you buy.
Two years ago I bought a Radeon RX 6600, good enough for my Computer 
(500W power supply) but didn't check the card's power requirements which 
is > 500W, so I had to replaced the power supply as well.

Peter

[toc] | [prev] | [next] | [standalone]

#45597

On 10/23/25 20:01, Peter McD wrote:
> Am 21.10.25 um 12:27 schrieb Paul:
>> On Tue, 10/21/2025 4:46 AM, Edmund wrote:
>>> GPU advise?
> ...
>> A couple of my power supplies, have multiple 2x4 connectors and that
>> might be sufficient up to 300W. Check the card picture for
>> evidence of the connector type there. If shopping that table
>> there, and not going too high up into the table, my 750W PSU would
>> likely be sufficient. For the top card in the list, I'd have to
>> go shop for a new 1200W supply with the 16 pin connector (2x6+4, the
>> 4 pins are status/declaration pins). The bottom card in the list
>> might have a 2x3 or a 2x4 on it, and the PSU that have 2x4
>> connectors, the end part unsnaps to convert the connector into
>> a 2x3. All the supplies in the room, have at least one
>> connector for the bottom card in the list.
>>
> ...
> Check power requirements of the card before you buy.
> Two years ago I bought a Radeon RX 6600, good enough for my Computer 
> (500W power supply) but didn't check the card's power requirements which 
> is > 500W, so I had to replaced the power supply as well.
> 
> Peter
> 
I have enough power and considering  this one

https://www.notebooksbilliger.de/inno3d+geforce+rtx+5060+ti+16gb+twin+x2+886634?nbbct=4004_idealo
-- 
-------------

Godspeed for Assange
Amnesty for Snowden
Rehabilitation for heroes

Edmund

[toc] | [prev] | [next] | [standalone]

#45598

On Thu, 10/23/2025 2:01 PM, Peter McD wrote:
> Am 21.10.25 um 12:27 schrieb Paul:
>> On Tue, 10/21/2025 4:46 AM, Edmund wrote:
>>> GPU advise?
> ...
>> A couple of my power supplies, have multiple 2x4 connectors and that
>> might be sufficient up to 300W. Check the card picture for
>> evidence of the connector type there. If shopping that table
>> there, and not going too high up into the table, my 750W PSU would
>> likely be sufficient. For the top card in the list, I'd have to
>> go shop for a new 1200W supply with the 16 pin connector (2x6+4, the
>> 4 pins are status/declaration pins). The bottom card in the list
>> might have a 2x3 or a 2x4 on it, and the PSU that have 2x4
>> connectors, the end part unsnaps to convert the connector into
>> a 2x3. All the supplies in the room, have at least one
>> connector for the bottom card in the list.
>>
> ...
> Check power requirements of the card before you buy.
> Two years ago I bought a Radeon RX 6600, good enough for my Computer 
> (500W power supply) but didn't check the card's power requirements 
> which is > 500W, so I had to replaced the power supply as well.
> 
> Peter

The math used in such estimates is bad, and you can do better than
that as a user, to estimate the power. Let us give this a try.

Modern video cards use closed loop feedback to control
power consumption. The video card clock will drop, rather
than the card exceeding the capabilities of its onboard
power converter.

https://www.techpowerup.com/gpu-specs/radeon-rx-6600.c3696

   RX 6600 draws power from 1x 8-pin power connector,
   with power draw rated at 132 W maximum              <=== max is policed by the card itself

Processors are harder to estimate.

https://www.intel.com/content/www/us/en/products/sku/236773/intel-core-i9-processor-14900k-36m-cache-up-to-6-00-ghz/specifications.html

   Processor Base Power  125 W
   Maximum Turbo Power   253 W     <=== 28 seconds, 56 seconds, or in some cases "continuous"

50W for motherboard. Add 50% to overall number for safety on PSU sizing (Pauls Margin).
The fast transitions on the mobo/gpu power converters tends to trash
marginal supplies, which is why we bump up the PSU a tiny bit. This helps
provide a cover story for inferior supplies (so if you "buy cheap" it still works).

   (132 + 253 + 50) * 1.5 = 435 W (zero margin) or 652.5 W (Pauls Margin)

If you have a "lesser processor" or one that does not use
a crazy turbo level like that one does, the number will be lower.
My CPU is 65W. It's an 8C 16T.

https://www.techpowerup.com/cpu-specs/ryzen-7-5700g.c2472

   TDP:   65 W
   PPT:   61-88 W    # Includes SOC power

I'm getting unreasonable behavior in Ryzen Master right now, and
it looks (via Kill-A-Watt meter on this computers plug), that
the 88W estimate on turbo will have to do for the maths.

   (132 + 88 + 50) * 1.5 = 270 W (zero margin) or 405 W (Pauls Margin)

The nearest size there would be a cheesy 450W. (The system ran
off a 450W at first, it was just in the box.) Or for the
lower value of "we won't margin it" a quality 350W would do.
Except, quality is part of the shopping experience, and we
end up buying larger supplies just because they are overall
better as a product. The PSU market is not a fully populated
matrix, and a lot of crap exists below 500W. HP and Dell
manage to get quality supplies, yet with very low power limits.

So while you have been given an "estimate" by some manufacturer,
this is not a religion and we seek to refine the number.

I'd give you more refined numbers, except once they moved to sleeved power,
ribbon power cables and black conductor markings, I can't slip
my clamp-on DC ammeter onto my board and give realistic power
for various things. The 50W number for the motherboard is
only approximate and assumes some DIMM power, and
assumes the Southbridge is 5W to 8W, and a Northbridge on
an older system with an iGPU could be 15W or so. Some of the
power on the Northbridge, comes from the high speed buses on it.
On the 440BX, Northbridges were running at around 2W back then.
Some brands didn't have heatsinks on top. Modern motherboards,
the Northbridge is inside the CPU now.

Once you work the details, the number is anywhere from 270W to 650W.

The number depended a lot, on the CPU chosen to run with it.

You compute a "custom number" to suit the size of the processor.
On some systems, the GPU number is bigger than the CPU.
On other systems, the CPU number is bigger than the GPU.

Quality supplies would be 500W and members of the 80+ program
(supplies that are more efficient than the old days where a
PSU was 65% efficient and threw out heat like a pig). The reason we
margin supplies, is to account for "wear behavior". I had a perfectly
fine Enermax, power calc was in-bounds and margined, but the fast
transitions on the new hardware, I was seeing the Enermax
get weaker by the day as if the Enermax was saying "Please, make it stop".
The 12V measured value with my multimeter, wasn't as perky any more.
I had to change out the Enermax and put in a small 80+ supply (620W).
And that's been running for at least two years now. You can't feel
any heat coming out of it right now.

*******

There are some web sites, that will do the kind of simple-minded
power calc I did above. One thing to watch for though, is
the "reasonable-ness test". One site years ago, used to
estimate *25W per DIMM* which is utter fantasy. You could cook
an egg on the DIMMs, if that were actually true.

You can chase down the numbers from first principles.

  https://www.micron.com/sales-support/design-tools/dram-power-calculator

  # A little too much detail but still interesting
  https://www.mouser.com/pdfDocs/tn4007_ddr4_power_calculation.pdf

Or, ask an AI. The answer given is simplistic. The extra DDR5 power is the PMIC.

###  Estimated Power Consumption by DDR Generation

| Type      | Voltage (Typical)        | Power per DIMM (Idle–Active) | Notes                            |
|-----------|--------------------------|------------------------------|----------------------------------|
| **DDR2**  | 1.8 V                    | 3–4.5 W                      | Older systems; higher idle power |
| **DDR3**  | 1.5 V (1.35 V for DDR3L) | 2–3 W                        | More efficient than DDR2         |
| **DDR4**  | 1.2 V                    | 2–3 W                        | Improved power gating and lower voltage |
| **DDR5**  | 1.1 V                    | 2.5–4 W                      | Includes onboard PMIC; higher bandwidth and density |

The numbers do not add. If two DIMMs are on the same channel, the power
would be 2+3W or 5W for the two DIMMs. If four DIMMs were present 2+3+2+3
(that uses two different channels, and two DIMMs are "active" at the same time).
If you feel the DIMMs, the power is 2.5W time-averaged on each DIMM.
So that's part of your motherboard 50W hand-wavey estimate, throws in some DIMMs.

The physical evidence is, those idle numbers are on the high side. They
are numbers for "Idle Windows Desktop", not S3 sleep. At S3 sleep, refresh
draws 1 watt per DIMM (measured). The DIMM is in auto-refresh mode then.
And in S3 sleep, all the DIMMs use the same power, as none of them can be
in the active state then.

Summary - I don't like that the video card companies don't just give a TDP
          for the card, and have the customer do the math. At least then
          we could justify our PSU purchases and "how much is enough".

   Paul

[toc] | [prev] | [next] | [standalone]

#45599

Am 24.10.25 um 06:08 schrieb Paul:
> On Thu, 10/23/2025 2:01 PM, Peter McD wrote:
>> Am 21.10.25 um 12:27 schrieb Paul:
>>> On Tue, 10/21/2025 4:46 AM, Edmund wrote:
>>>> GPU advise?
>> ...
>>> A couple of my power supplies, have multiple 2x4 connectors and that
>>> might be sufficient up to 300W. Check the card picture for
>>> evidence of the connector type there. If shopping that table
>>> there, and not going too high up into the table, my 750W PSU would
>>> likely be sufficient. For the top card in the list, I'd have to
>>> go shop for a new 1200W supply with the 16 pin connector (2x6+4, the
>>> 4 pins are status/declaration pins). The bottom card in the list
>>> might have a 2x3 or a 2x4 on it, and the PSU that have 2x4
>>> connectors, the end part unsnaps to convert the connector into
>>> a 2x3. All the supplies in the room, have at least one
>>> connector for the bottom card in the list.
>>>
>> ...
>> Check power requirements of the card before you buy.
>> Two years ago I bought a Radeon RX 6600, good enough for my Computer
>> (500W power supply) but didn't check the card's power requirements
>> which is > 500W, so I had to replaced the power supply as well.
> 
> The math used in such estimates is bad, and you can do better than
> that as a user, to estimate the power. Let us give this a try.
> 
> Modern video cards use closed loop feedback to control
> power consumption. The video card clock will drop, rather
> than the card exceeding the capabilities of its onboard
> power converter.
> 
> https://www.techpowerup.com/gpu-specs/radeon-rx-6600.c3696

...

All in all I agree that the average power requirements are much lower 
than that what my power supply could deliver. My computer uses a 65W 
processer (so to say) and it temperature is currently around 38°C, so I 
estimate the whole system may normally use around 50 W.

But I also read that the RX6600 power consumption may spike to more than 
500W under certain conditions and I want that covered.

Peter

[toc] | [prev] | [next] | [standalone]

#45604

On Fri, 10/24/2025 4:51 AM, Peter McD wrote:

> All in all I agree that the average power requirements are much lower than that
> what my power supply could deliver. My computer uses a 65W processer (so to say)
> and it temperature is currently around 38°C, so I estimate the whole system may 
> normally use around 50 W.
> 
> But I also read that the RX6600 power consumption may spike to more than 500W 
> under certain conditions and I want that covered.
> 
> Peter

That suggests a rather large regulator block on the RX6600.
(The RTX5090 can hit 600W on its regulator, but more square
inches would be put aside for it.

It doesn't look that big here.

https://www.techspot.com/articles-info/2343/images/2021-10-12-image-j.webp

   ( https://www.techspot.com/review/2343-amd-radeon-rx-6600/ )

This one has 25 phases for 600 watts.

   "RTX5090 takeapart"

   https://www.youtube.com/watch?v=mClh3SDF2SM

Your card has a 150W connector on the upper right. There
are six phases showing. If we pulled a number out of the air,
of 35 watts per phase, maybe it could supply 210 watts.

It's supposed to limit at 132W.

GPU-Z shows the power-state, so that you can figure out
what limit the card has hit. There is a text label identifying
the name of the state. Detection of overcurrent should be
automatic, unless the user has a software that disables a limiter.

   https://www.techpowerup.com/download/techpowerup-gpu-z/

   # Check under the Sensors tab, play a game and see what it says

The 5+1 phase regulator, looks like it would "squawk" if the
GPU attempted to draw 500W.

Before modern cards got these features, when Furmark first
came out, you could damage a video card with Furmark. Then
the drivers got "recognition" of Furmark, and they would
turn down the clock on purpose, to prevent Furmark from
breaking something. This is open loop control, implemented
with software. And it's pretty sloppy.

But eventually the hardware people said enough is enough,
and they put all sorts of controls in. There should also be
one for temperature. NVidia released a couple drivers,
where a mistake to one of the source files resulted in
the fans not spinning. Which makes cards hot. When that
happens on the CPU, the "THERMTRIP" feature causes the
entire PC power to go off. But the GPU back then
was not protected. The introduction of closed loop
control, having what was supposed to be an automatic
fan response curve, these were things invented to put
a stop to "low quality product behaviors".

To hit 500W, I think someone would have to hack the
controls and turn off the power limiter. Then, you need
a stimulus capable of driving the card that hard.

Also, if the card was doing that, the loom feeding the
2x4, it would have an overcurrent detection on it.
The PSU is going to go off if that transient lasts
for enough milliseconds. When a PSU has a large amperage
on +12V, the wiring is split into separate looms, and
the looms can have limits around 20 amps or so.

   Paul

[toc] | [prev] | [next] | [standalone]

#45618

Am 24.10.25 um 15:38 schrieb Paul:
> On Fri, 10/24/2025 4:51 AM, Peter McD wrote:
> 
>> All in all I agree that the average power requirements are much lower than that
>> what my power supply could deliver. My computer uses a 65W processer (so to say)
>> and it temperature is currently around 38°C, so I estimate the whole system may
>> normally use around 50 W.
>>
>> But I also read that the RX6600 power consumption may spike to more than 500W
>> under certain conditions and I want that covered.
>>
> 
> That suggests a rather large regulator block on the RX6600.
> (The RTX5090 can hit 600W on its regulator, but more square
> inches would be put aside for it.
> 
> It doesn't look that big here.
> 
> https://www.techspot.com/articles-info/2343/images/2021-10-12-image-j.webp
> 
>     ( https://www.techspot.com/review/2343-amd-radeon-rx-6600/ )
> 
> This one has 25 phases for 600 watts.
> 
>     "RTX5090 takeapart"
> 
>     https://www.youtube.com/watch?v=mClh3SDF2SM
> 
> Your card has a 150W connector on the upper right. There
> are six phases showing. If we pulled a number out of the air,
> of 35 watts per phase, maybe it could supply 210 watts.
> 
> It's supposed to limit at 132W.
> ...

I already upgraded the power supply to 800W when I changed the card 
about two years ago. So the money is spent.
Next time I will lock closer to what is required when I change parts.

Thank you for your answer anyway, I know now a lot more  about graphics 
card power requirements which might come in handy when I need an other one.

Peter





> GPU-Z shows the power-state, so that you can figure out
> what limit the card has hit. There is a text label identifying
> the name of the state. Detection of overcurrent should be
> automatic, unless the user has a software that disables a limiter.
> 
>     https://www.techpowerup.com/download/techpowerup-gpu-z/
> 
>     # Check under the Sensors tab, play a game and see what it says
> 
> The 5+1 phase regulator, looks like it would "squawk" if the
> GPU attempted to draw 500W.
> 
> Before modern cards got these features, when Furmark first
> came out, you could damage a video card with Furmark. Then
> the drivers got "recognition" of Furmark, and they would
> turn down the clock on purpose, to prevent Furmark from
> breaking something. This is open loop control, implemented
> with software. And it's pretty sloppy.
> 
> But eventually the hardware people said enough is enough,
> and they put all sorts of controls in. There should also be
> one for temperature. NVidia released a couple drivers,
> where a mistake to one of the source files resulted in
> the fans not spinning. Which makes cards hot. When that
> happens on the CPU, the "THERMTRIP" feature causes the
> entire PC power to go off. But the GPU back then
> was not protected. The introduction of closed loop
> control, having what was supposed to be an automatic
> fan response curve, these were things invented to put
> a stop to "low quality product behaviors".
> 
> To hit 500W, I think someone would have to hack the
> controls and turn off the power limiter. Then, you need
> a stimulus capable of driving the card that hard.
> 
> Also, if the card was doing that, the loom feeding the
> 2x4, it would have an overcurrent detection on it.
> The PSU is going to go off if that transient lasts
> for enough milliseconds. When a PSU has a large amperage
> on +12V, the wiring is split into separate looms, and
> the looms can have limits around 20 amps or so.
> 
>     Paul
> 
> 
> 

[toc] | [prev] | [standalone]

Back to top | Article view | alt.os.linux.mint

csiph-web