Radeon 780M bandwidth readings

[UmpaCumpa]

I noticed that GPU-Z provides a bandwidth reading of 204.8 GB/s for Radeon 780M integrated GPU in Ryzen 8700G. It seems to me that it's too high since the theoretical bandwidth of dual-channel DDR5 6400 should be 102.4 GB/s. Is this a bug or am I doing any mistake here?

 

[Contra]

He thinks it's GDDR5.
Bw GDDR5=4b/c x 3200MHz x 128b = 204.8 GB/s

 

[DirtyDingusMcgee]

Has more to do with the way the software sees it. 4x32 instead of 2x64. Nothing wrong.

 

[UmpaCumpa]

Has more to do with the way the software sees it. 4x32 instead of 2x64. Nothing wrong.

Both ways still work out to the same bus width of 128 bits, so it shouldn't affect bandwidth calculation. And the bus width is indicated correctly.

 

[UmpaCumpa]

Does anyone else have an idea about this? I guess that the displayed value is incorrect, do you agree?

 

[FanlessOC]

I agree that the calculation is off by a factor two.

As shown my memory runs at 8.4 gigatransfers per second, which, on a 128 bit (16 byte) wide data bus, comes down to 134.4 gigabytes per second maximum theoretical bandwidth. GPU-Z displays exactly double that value.

In terms of memory read speed, the theoretical maximum is far higher than what is realistically achievable on AMD's memory controller though...

*) the read speed might actually be a lot lower, I don't have a recent version of aida

 

[mczkrt]

I can confirm that bandwidth reading at least for AMD GPUs using DDR5 system memory is off by a factor of 2.

(2400*2*128)/8=76,8GB

Intel with DDR4 is fine.

(1333*2*64)/8=21,3GB

 

[PowerBuilder]

102.4GB/s bandwidth is correct. (AMD R7 8700G main memory interface = 128bit, 64bit slot x 2 6400MHz DDR5 SDRAM module card, real electronic signal 32bit x quad-channel) 204.8GB/s is a doubled figure, so it is an error in the GPU-Z program notation. And the actual bandwidth is slightly lower than the theoretical bandwidth due to the I/O structure and loss. Instead, the effective bandwidth may be slightly larger due to the effect of the GPU's internal high-speed L1, L2 cache memory and color compression technology.

 

[FanlessOC]

102.4GB/s bandwidth is correct. (AMD R7 8700G main memory interface = 128bit, 64bit slot x 2 6400MHz DDR5 SDRAM module card, real electronic signal 32bit x quad-channel) 204.8GB/s is a doubled figure, so it is an error in the GPU-Z program notation. And the actual bandwidth is slightly lower than the theoretical bandwidth due to the I/O structure and loss. Instead, the effective bandwidth may be slightly larger due to the effect of the GPU's internal high-speed L1, L2 cache memory and color compression technology.

Sorry, but no to that last part. Caches don't increase memory bandwidth, neither does compression. Like, my internet bandwidth doesn't go up if I download a zip file. Sure, the download might be faster, but the bandwidth stays the same.

I noted another difference between GPU-Z and the Techpowerup GPU database for this GPU: GPU-Z reports a PCIe 4.0 x16 bus, where the database says it's PCIe 4.0 x8. Not sure what it is in reality though. (see https://www.techpowerup.com/gpu-specs/radeon-780m.c4020 )