mcpat_template.xml for CPU-GPU heterogeneous architectures

[ttungl]

Hi, It's likely an off-topic but I would like to know that can McPAT simulate the CPU-GPU heterogeneous architecture? I looked into the examples of McPAT's ProcessorDescriptionFiles, but there is no configuration for GPU cores session. Can I just modify a CPU core for GPU cores session or does it have any specific configuration for GPU cores in the xml file for McPAT? Thanks.

Update: First, sorry for a confusing question. I am trying to modify the input file (.xml) for McPAT to simulate a HSA. My example for the CPU-GPU heterogeneous architecture is 16 cores CPUs, 16 compute units GPUs, and 4 Memory controllers. I use two cores for one CPU, therefore the number of CPUs totally is 8. I use four(4) compute units for one GPU, thus the total number of GPUs is 4. We have 4 MCs, each one is linked to each switch of the NOC. This forms a 4-by-4 NOC, with 8 CPUs, 4GPUs, 4 MCs in the network.
After simulating in m2s, the pipeline.out is generated. I would update the parameters in this file to the .xml file format for McPAT's simulation. The xml file in McPAT consists of the cores's parameters in the system, however, should the number of cores be 16 (8 CPUs + 4GPUs+4MCs) or 36 (16 cores CPUs + 16 compute units GPUs + 4MCs) -- I am confused about this number due to how does McPAT define the core (a single core in CPU or a whole CPU with two cores; (or) a compute unit or a whole GPU with 4 cus)? Is there any examples for GPU's configuration in McPAT ? I found one paper that mentioned about this configuration. According to the pipeline.out, it shows 16 cores in the Config.general. This means, McPAT only considers the nodes/switches as the cores. Please correct me if this speculation is wrong. Any helps would be appreciated. Thanks a lot. Best, Tung

[marcusnchow]

Tung,

Are you still looking for an example mcpat file? gpgpu-sim is a gpu simulator that is integrated with mcpat. The config file they use is this. gpuwattch_gtx480.txt This file is model after the GTX 480, which has 15 SM. Which I believe is similar to AMDs compute units.

I am curious to know how you configured M2S to work with heterogeneous architecture. Since I have had trouble with that before.

[trinayan]

Hi,

Multi2sim does not have a power model for the GPU yet. But there are efforts underway to do that. However the power model may be designed only for the new Kepler simulator and not the Southern Islands one. It should not be that hard to port but validation would have to be done. We might integrate GPUWaatch or GpuPow by making necessary changes. But if you want one as of now you would have to use gpgpu sim or macsim that have gpu power models. So you can look at those if you want

As for a fused architecture like AMD APU the current release code will not suppport this configuration. A basic implementation is ready that simulates this configuration but there are some problems that have not been solved. So we cannot release that code now until those issues are resolved.

On Apr 27, 2017 6:13 PM, "mchow24" notifications@github.com wrote:

Tung,

Are you still looking for an example mcpat file? gpgpu-sim is a gpu simulator that is integrated with mcpat. The config file they use is this. gpuwattch_gtx480.txt https://github.com/Multi2Sim/multi2sim/files/963043/gpuwattch_gtx480.txt This file is model after the GTX 480, which has 15 SM. Which I believe is similar to AMDs compute units.

I am curious to know how you configured M2S to work with heterogeneous architecture. Since I have had trouble with that before.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Multi2Sim/multi2sim/issues/46#issuecomment-297853782, or mute the thread https://github.com/notifications/unsubscribe-auth/AIoqKmpRfDFatEGVL2PHneTShKyGQEv9ks5r0RLzgaJpZM4MeFnP .