Jump to content
Search In
  • More options...
Find results that contain...
Find results in...

Hi,

 

I just watched the new Video about the new AMD GPUs: https://www.youtube.com/watch?v=0xUioPsLRKA

Because I am currently developing an own bare metal driver for AMD Rx570 GPUs, I find minutes 7:18 to 8:00 of the video very interesting.

 

I also encountered the fact that the GPU has only a 256MB BAR of accessible VRAM at boot time. However, most GPUs have a PCI Express register for resizing the BAR to it's full size. The CPU can access all of the VRAM after resizing. This is done in Linux Kernel already since version 4 for every supported AMD GPU and not just the latest Radeon 6000 cards and I assume the same is done in the windows driver too!

You can find the function for resizing the framebuffer BAR in the Linux Kernel here: https://elixir.bootlin.com/linux/v4.16/source/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c#L610

The driver will read the amount of VRAM from a GPU register called "mmCONFIG_MEMSIZE" first and then call the BAR resize function: https://elixir.bootlin.com/linux/v4.16/source/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c#L324

The mechanism of BAR resizing is explained in detail in the PCI Express Specification (https://cupdf.com/document/pci-express-base-specification-revision-30.html) in chapter 7.22 Resizable Bar Capability on page 757. Because the VRAM BAR is already a 64-bit BAR (for the Rx570 at least) it can already operate with 4GB or more.

There is no need to access VRAM in 256MB slices afterwards!

 

Did I misunderstood the video?

There are some more source which states the same like the video (https://www.wepc.com/news/what-is-amd-smart-access-memory/), but I could not find any sources where they have their information from.

If I am right the Smart Access Memory is nothing new and the concept is already used since years with older GPUs and even Intel CPUs.

Link to post
Share on other sites
12 hours ago, mluetkedreim said:

If I am right the Smart Access Memory is nothing new and the concept is already used since years with older GPUs and even Intel CPUs.

Then how do you explain their reporting of up to double-digit percentages of performance increase by enabling it?

CPURyzen 5 3600 Cooler: Arctic Liquid Freezer II 120mm AIO RAM: G.Skill Ripjaws V 4x8GB 3600 16-16-16-36 MoboASRock X570M Pro4

Graphics Card: ASRock Reference RX 5700 XT Case: Antec P5 PSU: Rosewill Capstone 750M Monitor: MSI Optix MAG241C Case Fans: 2x Arctic P12 PWM

Storage: HP EX950 1TB NVMe, HP EX900 1TB NVMe, dual Constellation ES 2GB in RAID1

https://hwbot.org/submission/4497882_btgbullseye_gpupi_v3.3___32b_radeon_rx_5700_xt_13min_37sec_848ms

Link to post
Share on other sites
38 minutes ago, BTGbullseye said:

Then how do you explain their reporting of up to double-digit percentages of performance increase by enabling it?

I think I expressed myself a bit misleading. I think there is something new called Smart Access Memory and it provides a performance increase, but it works different and not like the video explains it.

Link to post
Share on other sites
11 hours ago, mluetkedreim said:

I think I expressed myself a bit misleading. I think there is something new called Smart Access Memory and it provides a performance increase, but it works different and not like the video explains it.

That I will agree to.

CPURyzen 5 3600 Cooler: Arctic Liquid Freezer II 120mm AIO RAM: G.Skill Ripjaws V 4x8GB 3600 16-16-16-36 MoboASRock X570M Pro4

Graphics Card: ASRock Reference RX 5700 XT Case: Antec P5 PSU: Rosewill Capstone 750M Monitor: MSI Optix MAG241C Case Fans: 2x Arctic P12 PWM

Storage: HP EX950 1TB NVMe, HP EX900 1TB NVMe, dual Constellation ES 2GB in RAID1

https://hwbot.org/submission/4497882_btgbullseye_gpupi_v3.3___32b_radeon_rx_5700_xt_13min_37sec_848ms

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

×