Who is right here ? GPU-Z or AMD ?

[mooms]

I think AMD gives 4 times as much as GPU-Z because there are 4 LPDDR5 modules on this laptop, so in parallel it would effectively give 204 GB/s, but I'm not sure that games can really take advantage of this.

 

[Vya Domus]

I think neither are right, this is dual channel so it would work out to 2 X 51.2GB/s so about 100GB/s.

 

[damric]

lol weird

 

[W1zzard]

800 MHz x 4 (DDR5) x 128 bit / 8 bit per byte = 51.2 GB/s

Not sure if there's a mistake here somewhere, appreciate feedback

 

[AnotherReader]

800 MHz x 4 (DDR5) x 128 bit / 8 bit per byte = 51.2 GB/s

Not sure if there's a mistake here somewhere, appreciate feedback

Wouldn't 800 MHz for DDR5 correspond to DDR5 6400? In that case, the multiplier should be 8 rather than 4.

 

[mooms]

It's LPDDR5x 6400 so yes, it's x8.

GPU-Z displays DDR5 so maybe the issue is here.

ZenTimings is displaying as this, apparently its 4 single rank modules (LPDDR5X is soldered).

AIDA64 memory benchmark (scores are about the same with Hypervisor disabled).
I don't know why read speed is about half write speed.

 

[Vya Domus]

I don't know why read speed is about half write speed.

That's about right for dual channel DDR5 in aida64 memory benchmark.

 

[W1zzard]

So I looked into this. LPDDR5 (and LPDDR5X) clock the memory at one quarter the data clock, plus it's DDR, so x4 and x2 should be applied to be base clock of 800 MHz. (800 MHz x 4 x2 = DDR5-6400). Now multiply that by 128-bit bus width (6400 Mbps per pin x 128 pins = 819,200 Mbps) and divide by 8 to go from bits to byte (819,200 Mbps / 8 bit per byte = 102,400 MB/s)

Next release of GPU-Z will check whether the memory is DDR5 or LPDDR5 and apply the correct formula

 

[AMF]

It's LPDDR5x 6400 so yes, it's x8.

GPU-Z displays DDR5 so maybe the issue is here.

ZenTimings is displaying as this, apparently its 4 single rank modules (LPDDR5X is soldered).

AIDA64 memory benchmark (scores are about the same with Hypervisor disabled).
I don't know why read speed is about half write speed.

ouch very hi latency

 

[natr0n]

whatever number it is it wont get any faster.

 

[mooms]

ouch very hi latency

That's how LPDDR5 is.

So I looked into this. LPDDR5 (and LPDDR5X) clock the memory at one quarter the data clock, plus it's DDR, so x4 and x2 should be applied to be base clock of 800 MHz. (800 MHz x 4 x2 = DDR5-6400). Now multiply that by 128-bit bus width (6400 Mbps per pin x 128 pins = 819,200 Mbps) and divide by 8 to go from bits to byte (819,200 Mbps / 8 bit per byte = 102,400 MB/s)

Next release of GPU-Z will check whether the memory is DDR5 or LPDDR5 and apply the correct formula

Indeed:

 

[Wirko]

LPDDR5 (and LPDDR5X) clock the memory at one quarter the data clock

That's not always true, according to this document by Synopsys.

Thus, LPDDR5 adopts a new clocking scheme, where CK runs at one fourth the data-strobe frequency at speeds higher than 3200 Mbps, and at half the data-strobe frequency at speeds under 3200 Mbps. Hence, even at 6400 Mbps, this clocking scheme requires CK to operate only at 800 MHz.

Do you have a means to determine the true data clock frequency (WCK, RDQS) independently of CK?

 

[W1zzard]

That's not always true

Yup, but why would anyone put LPDDR5-3200 in a device that can run GPU-Z? If we see something like that I'll look into it

 

[mooms]

So, since we now know that's the correct bandwidth is 102 GB/s, do we have an idea why the AMD control panel is showing 204 GB/s ?

 

[Beginner Macro Device]

do we have an idea why the AMD control panel is showing 204 GB/s

Good ol' tradition. AMD software must contain at least one chunk of mocknuggetsery in order to be valid.

 

[mooms]

mocknuggetsery ? I'm not a native reader, can you reformulate ?

 

[Beginner Macro Device]

Tomfoolery, stupidity, BS, you name it. L'absurde absolu if we put it the French way.

//also not a native English speaker

 

[ThrashZone]

mocknuggetsery ? I'm not a native reader, can you reformulate ?

Hi,
Short version is contact amd and report bug hehe