diff options
Diffstat (limited to 'src/mem/DRAMCtrl.py')
-rw-r--r-- | src/mem/DRAMCtrl.py | 38 |
1 files changed, 27 insertions, 11 deletions
diff --git a/src/mem/DRAMCtrl.py b/src/mem/DRAMCtrl.py index 171d99686..4eb8ac18a 100644 --- a/src/mem/DRAMCtrl.py +++ b/src/mem/DRAMCtrl.py @@ -12,6 +12,7 @@ # # Copyright (c) 2013 Amin Farmahini-Farahani # Copyright (c) 2015 University of Kaiserslautern +# Copyright (c) 2015 The University of Bologna # All rights reserved. # # Redistribution and use in source and binary forms, with or without @@ -41,6 +42,7 @@ # Ani Udipi # Omar Naji # Matthias Jung +# Erfan Azarkhish from m5.params import * from AbstractMemory import * @@ -382,8 +384,8 @@ class DDR3_1600_x64(DRAMCtrl): # developed at the University of Kaiserslautern. This high level tool # uses RC (resistance-capacitance) and CV (capacitance-voltage) models to # estimate the DRAM bank latency and power numbers. -# [2] A Logic-base Interconnect for Supporting Near Memory Computation in the -# Hybrid Memory Cube (E. Azarkhish et. al) +# [2] High performance AXI-4.0 based interconnect for extensible smart memory +# cubes (E. Azarkhish et. al) # Assumed for the HMC model is a 30 nm technology node. # The modelled HMC consists of 4 Gbit layers which sum up to 2GB of memory (4 # layers). @@ -397,8 +399,8 @@ class DDR3_1600_x64(DRAMCtrl): # devices per rank (DDR) => devices per layer ( 1 for HMC). # The parameters for which no input is available are inherited from the DDR3 # configuration. -# This configuration includes the latencies from the DRAM to the logic layer of -# the HMC +# This configuration includes the latencies from the DRAM to the logic layer +# of the HMC class HMC_2500_x32(DDR3_1600_x64): # size of device # two banks per device with each bank 4MB [2] @@ -441,7 +443,8 @@ class HMC_2500_x32(DDR3_1600_x64): # cycles (Assumption) tRRD = '3.2ns' - # activation limit is set to 0 since there are only 2 banks per vault layer. + # activation limit is set to 0 since there are only 2 banks per vault + # layer. activation_limit = 0 # Values using DRAMSpec HMC model [1] @@ -449,20 +452,33 @@ class HMC_2500_x32(DDR3_1600_x64): tWR = '8ns' tRTP = '4.9ns' - # Default different rank bus delay assumed to 1 CK for TSVs, @1250 MHz = 0.8 - # ns (Assumption) + # Default different rank bus delay assumed to 1 CK for TSVs, @1250 MHz = + # 0.8 ns (Assumption) tCS = '0.8ns' # Value using DRAMSpec HMC model [1] tREFI = '3.9us' - # Set default controller parameters - page_policy = 'close' - write_buffer_size = 8 - read_buffer_size = 8 + # The default page policy in the vault controllers is simple closed page + # [2] nevertheless 'close' policy opens and closes the row multiple times + # for bursts largers than 32Bytes. For this reason we use 'close_adaptive' + page_policy = 'close_adaptive' + + # RoCoRaBaCh resembles the default address mapping in HMC addr_mapping = 'RoCoRaBaCh' min_writes_per_switch = 8 + # These parameters do not directly correlate with buffer_size in real + # hardware. Nevertheless, their value has been tuned to achieve a + # bandwidth similar to the cycle-accurate model in [2] + write_buffer_size = 32 + read_buffer_size = 32 + + # The static latency of the vault controllers is estimated to be smaller + # than a full DRAM channel controller + static_backend_latency='4ns' + static_frontend_latency='4ns' + # A single DDR3-2133 x64 channel refining a selected subset of the # options for the DDR-1600 configuration, based on the same DDR3-1600 # 4 Gbit datasheet (Micron MT41J512M8). Most parameters are kept |