PCI-e 3.0 x8 may not have enough bandwidth for RX 5500 XT

[cucker tarlson]

Because AMD is letting AIBs clear inventory of older stock (especially RX 590, RX 580, and RX 570). The 5500 cards (all of them) will come down in price after they're done.

so amd still wants to sell 2.5 year old polaris 10 months after 16-series came out and will continue to mark up their new gen cards until the inventory is gone
this sounds like they're not doing consumers in this segment much favor

let me guess,they'll throw rx590s with 10-game bundles at ampere cards

 

[Vya Domus]

You just confirmed it yourself, after VRAM is full, PCIe is the next bottleneck.

Only thing I confirmed is the fact that GPUs don't use system RAM, ever, they only use what's in the VRAM. After the VRAM is full, you can forget about it, PCIe 4.0 or 3.0 or x8 or whatever, it's over, performance will tank.

The situation where you're running out of memory it's a game over scenario in any software of any kind, only thing you can do is ensure it doesn't crash right away. If you can.

 

[GoldenX]

Memory heaps are used to have quicker stream of RAM<>VRAM, on a driver level. So RAM is used, but never directly.

 

[Vya Domus]

So RAM is used, but never directly.

The GPU fetches instructions and data from it's own memory, all the time, never from system memory. It never uses RAM, directly or indirectly via DMA, I don't know how I could make this any simpler, I have to give up I guess.

Nvidia does have something like that which is reserved for Teslas with CUDA I think, I don't know if anything commercial uses it or if it requires something else to work.

 

[(*^^*)]

*cough* 8 GiB

The question is thus: Why do GeForce cards perform so much better with 4 GiB compared to AMD cards? Maybe it is because of x16 versus x8 (faster RAM access).

In the first place, people who play Assassin's Creed in the 1660 and 5500 series. not many. Most people who need 1660 or 5500XT play lighter games. In the first place, there is no problem even if 5500XT4GB is slightly inferior to 16604GB. Rather, the game structure is optimized for NVIDIA. It can be considered that a little difference is good. Do you express this as failure? Anyway, we have 5600XT. For light gamers who are not interested in being fast, the 5500XT and Gen4 phenomena may be a good thing.

 

[cucker tarlson]

In the first place, people who play Assassin's Creed in the 1660 and 5500 series. not many. Most people who need 1660 or 5500XT play lighter games.

absolutely bs

For light gamers who are not interested in being fast, the 5500XT and Gen4 phenomena may be a good thing.

yes,cause why have your budget card run games without a board that costs more than the card.brilliant.

 

[GoldenX]

In the first place, people who play Assassin's Creed in the 1660 and 5500 series. not many. Most people who need 1660 or 5500XT play lighter games. In the first place, there is no problem even if 5500XT4GB is slightly inferior to 16604GB. Rather, the game structure is optimized for NVIDIA. It can be considered that a little difference is good. Do you express this as failure? Anyway, we have 5600XT. For light gamers who are not interested in being fast, the 5500XT and Gen4 phenomena may be a good thing.

What kind of logic is that?

The GPU fetches instructions and data from it's own memory, all the time, never from system memory. It never uses RAM, directly or indirectly via DMA, I don't know how I could make this any simpler, I have to give up I guess.

Nvidia does have something like that which is reserved for Teslas with CUDA I think, I don't know if anything commercial uses it or if it requires something else to work.

AMD Radeon R9 200 Series - Vulkan Hardware Database by Sascha Willems

vulkan.gpuinfo.org

That's my 270x, you can have access to some "fast driver level 256MB", the rest of the 2GB, and some "additional" 8GB of RAM. All that "VRAM" is available to developers.
Check other cards, you will see similar stuff. The RAM heap will vary in size according to what the PC has.

 

[EarthDog]

It's using RAM in the sense that that's where all the buffers are kept, the buffers however are never accessed directly by the GPU, they have to be in it's VRAM before that's possible.

.....which it uses the pcie bus to send from ram to the card.

The 5700XT has more memory and it also has more PCIe lanes. Everything else works in the same way.

Right... just showing that overall performance doesnt matter from 4.0 x16 to 3.0 x16 when vram isnt filled. Here we have vram being filled and the pcie bus is used to swap that data and it's faster. If it isnt the pcie bus... the only variable that changed... what is it?

The games simply swap memory in and out of the card's memory, it never gets directly accessed from across the PCIe lanes.

correct. But that data swaps from system ram back to vram over the pcie bus to be read. If it is frequently accessed and swapped, makes sense what we are seeing...what the tests show.

The question is thus: Why do GeForce cards perform so much better with 4 GiB compared to AMD cards? Maybe it is because of x16 versus x8 (faster RAM access).

..this and it's better memory compression algos too.

 

[GoldenX]

Just look at the task manager while you play a game. You always use a portion of RAM for the GPU (and it's not only for textures). Thing is, if VRAM is not filled, you never see the performance drop.
Now, if you have a low amount of VRAM, it reaches 100%, you PCIe bus is slower due to "cost reducing electrical limitations", and to make it worse, you use DDR4 2400MHz RAM, or worse, DDR3... The low end users just got one hell of a punch to the guts.

 

[Vya Domus]

That's my 270x, you can have access to some "fast driver level 256MB", the rest of the 2GB, and some "additional" 8GB of RAM. All that "VRAM" is available to developers.
Check other cards, you will see similar stuff. The RAM heap will vary in size according to what the PC has.

How does any of this prove the GPU can fetch and execute stuff directly from RAM without having to make copies first to their local memory ?

You do realize these are still associated with local and host memory right ? That's why they're distinct regions in the first place, they are not all VRAM.

See that DEVICE_LOCAL_BIT on the 2 GB region that doesn't exist for the 8 GB one ? Local means GPU visible only memory, that's why the heap associated with the system memory doesn't have that flag

... because it isn't VRAM ...

View attachment 140402
Just look at the task manager while you play a game. You always use a portion of RAM for the GPU (and it's not only for textures). Thing is, if VRAM is not filled, you never see the performance drop.
Now, if you have a low amount of VRAM, it reaches 100%, you PCIe bus is slower due to "cost reducing electrical limitations", and to make it worse, you use DDR4 2400MHz RAM, or worse, DDR3... The low end users just got one hell of a punch to the guts.

You can show me this all day, unless you can prove that shared memory is accessed directly be the GPU, which it can't do, you're wasting your time. At one point you'll need to realize that not all memory that gets managed by the API actually ends up on the GPU.

 

[GoldenX]

It's managed by the driver (or in Vulkan's case, the developer), and it doesn't change the fact that PCIe and RAM performance is needed when you fill your VRAM.

 

[Vya Domus]

It's managed by the driver

Exactly but those copies still need to be made, they're just not exposed to the programmer.

 

[Flaky]

This was true with up to PCIe 3.0, but what about with PCIe 4.0 slots?

Nothing has changed in regard to how lane count and pcie generation is negotiated, and affects resulting link and it's bandwidth.

Does a x4 card (electrically and or / for whatever other reason functioning @ x4) work @ x4 or @ x8 in a PCIe 4.0 slot?

This whole issue stems from the fact that x8 is working as x16 in a PCIe 4.0 slot: does the same apply to x4?

This is confusing and nonsense at the same time. Lane count is the number of active physical, electrical transmit+receive pairs between the controller and the device.
If some diagnostic program reports wider link than a certain port or device supports, it simply got it wrong.

The whole "as if"/"is working as" simplification causes more confusion than it is worth. This stuff is really simple.
PCIe gen + PCIe lane count => resulting bandwidth
Not the other way around.

 

[HTC]

Nothing has changed in regard to how lane count and pcie generation is negotiated, and affects resulting link and it's bandwidth.


This is confusing and nonsense at the same time. Lane count is the number of active physical, electrical transmit+receive pairs between the controller and the device.
If some diagnostic program reports wider link than a certain port or device supports, it simply got it wrong.

The whole "as if"/"is working as" simplification causes more confusion than it is worth. This stuff is really simple.
PCIe gen + PCIe lane count => resulting bandwidth
Not the other way around.

There's obviously some confusion here and i don't know if it's from you, me or both.

Take a look @ these two pics (from the article in the OP):

These pics show the EXACT SAME CARD: once in a PCIe 3.0 board and the another in a PCIe 4.0 board. As can be seen, the one in a PCIe 4.0 board has nearly double the bandwidth in GPU read / write and this is supposed to be BECAUSE the x8 PCIe 3.0 card is actually working as x16 PCIe 3.0 in an X570 board, which is why the bandwidth nearly doubles, EVEN thought it only is x8 electrically.

My question was: does an x4 PCIe 3.0 card behave the same way?

If it does, is this by accident from the board dudes or something in the drivers that enables this? Perhaps a combination of both? Or is it supposed to be working as intended?

EDIT

Strange: it's supposed to be the same card but there are slight differences, such as 20MHz speed difference, among others.

 

[Flaky]

(...) BECAUSE the x8 PCIe 3.0 card is actually working as x16 PCIe 3.0 in an X570 board, which is why the bandwidth nearly doubles, EVEN thought it only is x8 electrically.

Just stop.
On the left you have PCIe 3.0 x8 connection.
On the right you have PCIe 4.0 x8 connection.
That's it, and that's all. There is no x16 anywhere here.

 

[HTC]

Just stop.
On the left you have PCIe 3.0 x8 connection.
On the right you have PCIe 4.0 x8 connection.
That's it, and that's all. There is no x16 anywhere here.

It seems i'm expressing myself wrong: they BOTH use x8 connection but one has x8 speed while the other has THE EQUIVALENT of x16 speed.

EDIT

It's as if the BOARD'S PCIe 4.0 slot is doubling the card's PCIe 3.0 connection speed. It's not supposed to do that, I THINK, but it appears that's what's happening with the 5500XT. Dunno how else to explain that read / write GPU memory speed doubling.

EDIT #2

Wait.

I think i got it backwards: it's not the X570 board that doubles the speed but ALL other NON-PCIe 4.0 capable boards that LIMIT the x8 PCIe 4.0 to PCIe 3.0, thus HALVING the speed. This actually makes a lot more sense.

 

[Waldorf]

well, because 3.0 is slower, not because its "limiting" it.
See it like sata. 3 is faster than 2 with the same amount of connections.

Everytime you go "up" one generation, it doubles the bandwidth each lane can carry, nothing more.

ppl might say "4.0 x8 has the same speed (as 3.0 x16)", but thats only true in regards to performance, not physical/electrical lanes.

 

[HTC]

well, because 3.0 is slower, not because its "limiting" it.
See it like sata. 3 is faster than 2 with the same amount of connections.

Everytime you go "up" one generation, it doubles the bandwidth each lane can carry, nothing more.

ppl might say "4.0 x8 has the same speed (as 3.0 x16)", but thats only true in regards to performance, not physical/electrical lanes.

That was my point exactly.

I think i figured it out though, like i said in the 2nd edit of my previous post.

It's not the X570 board that is doubling the speed but, instead, ALL the boards that are NOT capable of PCIe 4.0 are HALVING the bandwidth speed of the PCIe capable 4.0 cards.

With the 5500XT being a PCIe 4.0 capable card, it runs @ double the speed of PCIe 3.0 so long as the board it's placed in is ALSO PCIe 4.0 capable. If not, it will run @ PCIe 3.0 speed instead, effectively cutting it's bandwidth speed in half.

This being the case, my earlier question is already answered: a PCIe 4.0 card will run in an X570 board with twice the bandwidth VS the exact same card in a PCIe 3.0 board.

 

[hat]

Yeah, that's how PCI-E works. Every generation doubles the bandwidth, so a PCI-E 4.0 card in a 3.0 slot will only have half the bandwidth available to it. This usually isn't a big deal, but when you put a small amount of vram on a card that's wired for 4.0 x8 and stick it in a 3.0 slot (where it will run, of course, at 3.0 x8), you run into serious limitations when the vram runs out. Sending data back and forth between the card and system ram over PCI-E, even if it's 4.0 x16, is going to be a lot slower than just having everything on the card ready to go.

 

[GoldenX]

As simple as that.