cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

is2013
Journeyman III

Cache size in AMD Radeon HD7660D

Hi,

I am having a little confusion and couldn't find a clear cut solution. Is the cache (L1 and L2) size of AMD Radeon HD7660D in APU A10-5800K the same as that of CPU or is it different? The CPU has L1 = 64K and L2 = 4M. As far as i know, both CPU and GPU have their own caches even in case of APU. If they don't share the same caches, can anyone specify how to find these nos. for fused GPU?

0 Likes
1 Solution
ravkum
Staff

Hi,

CPU and GPU have their own caches, even on APU.

HD7660D is a northern island architecture device. For these devices, the L1 cache per compute unit is 8 kB and the L2 cache per memory channel is 64 kB.

For southern island devices, the L1 cache per compute unit is 16kB and L2 cache per memory channel is 64 kB.


Regards,

Ravi

View solution in original post

0 Likes
9 Replies
ravkum
Staff

Hi,

CPU and GPU have their own caches, even on APU.

HD7660D is a northern island architecture device. For these devices, the L1 cache per compute unit is 8 kB and the L2 cache per memory channel is 64 kB.

For southern island devices, the L1 cache per compute unit is 16kB and L2 cache per memory channel is 64 kB.


Regards,

Ravi

0 Likes

Hey thanks for the reply. Can you please share the link to the official documentation containing this information?

0 Likes

Hi,

This information is being updated in the AMD APP Programming guide. It will be available in the next release of the programming guide.

Regards,

0 Likes

Hi,

Can you please provide the private register size on these architectures?

0 Likes

Hi,

In Northern Island devices, VGPRs: 256 (128 - bit) SGPRs:    0, LDS Size: 32 kB

In Southern Island devices, VGPRs: 256 (32  - bit) SGPRs: 512, LDS Size: 32 kB

This information will also be available in the OpenCL Programming guide.

Regards,

So for the GPU in Trinity, the total VGPRs in bytes will be 256*128*16 (processing elements) * 6 CUs? and same goes for LDS size?

0 Likes

For Trinity

- size of vector register : 256KB per CU. So per GPU that would be 256KB * 6 CUs.

- LDS : 32KB per CU. So per GPU that would be 32KB * 6 CUs.

Regards

Pradeep

So i think the last question to sum this up would be whats the global memory in case of discrete GPUs and APUs? specifically in case of AMD Radeon HD7660D in APU A10-5800K and of HD 5870.

Is the global memory a small portion of the CPU memory??

0 Likes

Hi ifrah,

AMD APU has same physical memory which is partitioned based on device. You can use either CPU as a device for which memory allocation will be done on host side, or GPU as device for which memory allocation will be done on device side. In terms of OpenCL, memory is logically separate. Hence the global memory of GPU is a small portion of whole physical memory.

Size of global memory for discrete GPUs and APUs can be obtained using clinfo.

If you want to make GPU's global memory larger, you can try it with BIOS settings as explained here: Re: Is there are any way to make global memory of APU larger?

Thanks,

AMD_Support

0 Likes