GTX580 and How to Config in GPGPU-Sim

Introducing GTX580!

This is the fully unlocked version of GTX480 that have been provided by GPGPU-Sim. The only real difference is that we are going from 15 SM cores to 16 SM cores. This is done by setting gpgpu_n_clusters parameter in the config file, and everything else remains the same.

======

How to Configure GPU Cores to share L1I cache

There are two files that need to be changed:

• gpgpusim.config
  
• config_fermi_islip.icnt
  

Considering how default GTX580 has total of 16SM cores, let NSCALE be the number of SM cores we want to group together. (ex. if NSCALE = 2, we are having every two cores share a signle L1I, resulting in 8 total L1I caches)

Remember that we are having the cores share common caches by merging cores together and scaling a single core to match the pipeline of several cores

In gpgpusim.config

Here's the list of parameters that need to be changed:

• gpgpu_n_clusters 16/NSCALE - Number of SIMT clusters
  
• gpgpu_shader_registers 32768*NSCALE - Number of Shader Core registers
  
• gpgpu_shader_core_pipeline 1536*NSCALE:32 - Number of threads in the pipeline
  
• gpgpu_shader_cta 8*NSCALE - Number of Cooperative Thread Arrays per shader core
  
• gpgpu_num_sched_per_core 2*NSCALE - Number of schedulers per core
  
• gpgpu_cache:il1 - Instruction cache architecture
  

gpgpu_n_clusters - For cores to share instruction cache, we need them to be in the same cluster, so we reduce the cluster count by how many caches there are.

gpgpu_shader_registers - Each core contains 32768 registers, so scale it by number of cores we want to represent.

gpgpu_shader_core-pipeline - Each Fermi core contains 1536 threads, so scale that by NSCALE. 32 is the number of threads per warp. (resulting in 48 warps per core)

gpgpu_shader_cta - Each Fermi core contains 8 CTA, so scale it by NSCALE.

gpgpu_num_sched_per_core - Each Fermi core has two schedulers, so we scale that by NSCALE.

Finally, we configure gpgpu_cache:il1 based on the cache architecture we want to simulate. We have discussed this on a previous blog post.

In config_fermi_islip.icnt

There's only one thing that needs to be changed in this file.

• k = 28 - This is the topology of the simulated GPU.
  

This variable defines the Topology configuration. k should equal gpgpu_n_clusters + gpgpu_n_mem*gpgpu_n_sub_partition_per_mchannel from gpgpu_config. For GTX580 configuration, since n_mem = 6 and sub_partition = 2, we need k = n_clusters + 12.

======

Current status of testing:

These are the cache parameters that have been simulated for each core configuration.
The L1I configurations are specified as: SetNum_BlockSize_Associativity

• 01SM
  • 1_128_24 , 1_128_32
  • 2_128_24 , 2_128_4 , 2_128_16 2_128_32
  • 4_128_2 , 4_128_4 , 4_128_6 , 4_128_8 , 4_128_10 , 4_128_12 , 4_128_14 , 4_128_16 , 4_128_20 , 4_128_24 , 4_128_32
  • 6_128_4
  • 8_128_4
• 02SM
  • 1_128_32
  • 2_128_4 , 2_128_16
  • 4_128_2 , 4_128_4 , 4_128_6 , 4_128_8 , 4_128_10 , 4_128_16
  • 6_128_4
  • 8_128_4
  • 12_128_4
  • 16_128_4
• 04SM
  • 1_128_4
  • 2_128_4 , 2_128_16 , 2_128_32
  • 4_128_1 , 4_128_2 , 4_128_4 , 4_128_6 , 4_128_8 , 4_128_12 , 4_128_10 , 4_128_16 , 4_128_24
    
  • 6_128_4
    
  • 8_128_4
  • 16_128_4
• 08SM
  • 2_128_4
  • 4_128_2 , 4_128_4 , 4_128_8
  • 8_128_4
• 16SM (default)
  • 2_128_4
  • 4_128_4

Fabhia's currently processing all the data we got from all the runs.