1 # Copyright (c) 2012-2020 ARM Limited
4 # The license below extends only to copyright in the software and shall
5 # not be construed as granting a license to any other intellectual
6 # property including but not limited to intellectual property relating
7 # to a hardware implementation of the functionality of the software
8 # licensed hereunder. You may use the software subject to the license
9 # terms below provided that you ensure that this notice is replicated
10 # unmodified and in its entirety in all distributions of the software,
11 # modified or unmodified, in source code or in binary form.
13 # Copyright (c) 2013 Amin Farmahini-Farahani
14 # Copyright (c) 2015 University of Kaiserslautern
15 # Copyright (c) 2015 The University of Bologna
16 # All rights reserved.
18 # Redistribution and use in source and binary forms, with or without
19 # modification, are permitted provided that the following conditions are
20 # met: redistributions of source code must retain the above copyright
21 # notice, this list of conditions and the following disclaimer;
22 # redistributions in binary form must reproduce the above copyright
23 # notice, this list of conditions and the following disclaimer in the
24 # documentation and/or other materials provided with the distribution;
25 # neither the name of the copyright holders nor the names of its
26 # contributors may be used to endorse or promote products derived from
27 # this software without specific prior written permission.
29 # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 # "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 # LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 # A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 # OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 # SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 # LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 # DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 # THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41 from m5
.params
import *
42 from m5
.proxy
import *
43 from m5
.objects
.AbstractMemory
import *
44 from m5
.objects
.QoSMemCtrl
import *
46 # Enum for memory scheduling algorithms, currently First-Come
47 # First-Served and a First-Row Hit then First-Come First-Served
48 class MemSched(Enum
): vals
= ['fcfs', 'frfcfs']
50 # Enum for the address mapping. With Ch, Ra, Ba, Ro and Co denoting
51 # channel, rank, bank, row and column, respectively, and going from
52 # MSB to LSB. Available are RoRaBaChCo and RoRaBaCoCh, that are
53 # suitable for an open-page policy, optimising for sequential accesses
54 # hitting in the open row. For a closed-page policy, RoCoRaBaCh
55 # maximises parallelism.
56 class AddrMap(Enum
): vals
= ['RoRaBaChCo', 'RoRaBaCoCh', 'RoCoRaBaCh']
58 # Enum for the page policy, either open, open_adaptive, close, or
60 class PageManage(Enum
): vals
= ['open', 'open_adaptive', 'close',
63 # DRAMCtrl is a single-channel single-ported DRAM controller model
64 # that aims to model the most important system-level performance
65 # effects of a DRAM without getting into too much detail of the DRAM
67 class DRAMCtrl(QoSMemCtrl
):
69 cxx_header
= "mem/dram_ctrl.hh"
71 # single-ported on the system interface side, instantiate with a
72 # bus in front of the controller for multiple ports
73 port
= SlavePort("Slave port")
75 # the basic configuration of the controller architecture, note
76 # that each entry corresponds to a burst for the specific DRAM
77 # configuration (e.g. x32 with burst length 8 is 32 bytes) and not
78 # the cacheline size or request/packet size
79 write_buffer_size
= Param
.Unsigned(64, "Number of write queue entries")
80 read_buffer_size
= Param
.Unsigned(32, "Number of read queue entries")
82 # threshold in percent for when to forcefully trigger writes and
83 # start emptying the write buffer
84 write_high_thresh_perc
= Param
.Percent(85, "Threshold to force writes")
86 # threshold in percentage for when to start writes if the read
88 write_low_thresh_perc
= Param
.Percent(50, "Threshold to start writes")
90 # minimum write bursts to schedule before switching back to reads
91 min_writes_per_switch
= Param
.Unsigned(16, "Minimum write bursts before "
94 # scheduler, address map and page policy
95 mem_sched_policy
= Param
.MemSched('frfcfs', "Memory scheduling policy")
96 addr_mapping
= Param
.AddrMap('RoRaBaCoCh', "Address mapping policy")
97 page_policy
= Param
.PageManage('open_adaptive', "Page management policy")
99 # enforce a limit on the number of accesses per row
100 max_accesses_per_row
= Param
.Unsigned(16, "Max accesses per row before "
103 # size of DRAM Chip in Bytes
104 device_size
= Param
.MemorySize("Size of DRAM chip")
106 # pipeline latency of the controller and PHY, split into a
107 # frontend part and a backend part, with reads and writes serviced
108 # by the queues only seeing the frontend contribution, and reads
109 # serviced by the memory seeing the sum of the two
110 static_frontend_latency
= Param
.Latency("10ns", "Static frontend latency")
111 static_backend_latency
= Param
.Latency("10ns", "Static backend latency")
113 # the physical organisation of the DRAM
114 device_bus_width
= Param
.Unsigned("data bus width in bits for each DRAM "\
116 burst_length
= Param
.Unsigned("Burst lenght (BL) in beats")
117 device_rowbuffer_size
= Param
.MemorySize("Page (row buffer) size per "\
119 devices_per_rank
= Param
.Unsigned("Number of devices/chips per rank")
120 ranks_per_channel
= Param
.Unsigned("Number of ranks per channel")
122 # default to 0 bank groups per rank, indicating bank group architecture
124 # update per memory class when bank group architecture is supported
125 bank_groups_per_rank
= Param
.Unsigned(0, "Number of bank groups per rank")
126 banks_per_rank
= Param
.Unsigned("Number of banks per rank")
128 # Enable DRAM powerdown states if True. This is False by default due to
129 # performance being lower when enabled
130 enable_dram_powerdown
= Param
.Bool(False, "Enable powerdown states")
132 # For power modelling we need to know if the DRAM has a DLL or not
133 dll
= Param
.Bool(True, "DRAM has DLL or not")
135 # DRAMPower provides in addition to the core power, the possibility to
136 # include RD/WR termination and IO power. This calculation assumes some
137 # default values. The integration of DRAMPower with gem5 does not include
138 # IO and RD/WR termination power by default. This might be added as an
139 # additional feature in the future.
141 # timing behaviour and constraints - all in nanoseconds
143 # the base clock period of the DRAM
144 tCK
= Param
.Latency("Clock period")
146 # the amount of time in nanoseconds from issuing an activate command
147 # to the data being available in the row buffer for a read/write
148 tRCD
= Param
.Latency("RAS to CAS delay")
150 # the time from issuing a read/write command to seeing the actual data
151 tCL
= Param
.Latency("CAS latency")
153 # minimum time between a precharge and subsequent activate
154 tRP
= Param
.Latency("Row precharge time")
156 # minimum time between an activate and a precharge to the same row
157 tRAS
= Param
.Latency("ACT to PRE delay")
159 # minimum time between a write data transfer and a precharge
160 tWR
= Param
.Latency("Write recovery time")
162 # minimum time between a read and precharge command
163 tRTP
= Param
.Latency("Read to precharge")
165 # time to complete a burst transfer, typically the burst length
166 # divided by two due to the DDR bus, but by making it a parameter
167 # it is easier to also evaluate SDR memories like WideIO.
168 # This parameter has to account for burst length.
169 # Read/Write requests with data size larger than one full burst are broken
170 # down into multiple requests in the controller
171 # tBURST is equivalent to the CAS-to-CAS delay (tCCD)
172 # With bank group architectures, tBURST represents the CAS-to-CAS
173 # delay for bursts to different bank groups (tCCD_S)
174 tBURST
= Param
.Latency("Burst duration "
175 "(typically burst length / 2 cycles)")
177 # tBURST_MAX is the column array cycle delay required before next access,
178 # which could be greater than tBURST when the memory access time is greater
180 tBURST_MAX
= Param
.Latency(Self
.tBURST
, "Column access delay")
182 # tBURST_MIN is the minimum delay between bursts, which could be less than
183 # tBURST when interleaving is supported
184 tBURST_MIN
= Param
.Latency(Self
.tBURST
, "Minimim delay between bursts")
186 # CAS-to-CAS delay for bursts to the same bank group
187 # only utilized with bank group architectures; set to 0 for default case
188 # tBURST is equivalent to tCCD_S; no explicit parameter required
189 # for CAS-to-CAS delay for bursts to different bank groups
190 tCCD_L
= Param
.Latency("0ns", "Same bank group CAS to CAS delay")
192 # Write-to-Write delay for bursts to the same bank group
193 # only utilized with bank group architectures; set to 0 for default case
194 # This will be used to enable different same bank group delays
195 # for writes versus reads
196 tCCD_L_WR
= Param
.Latency(Self
.tCCD_L
,
197 "Same bank group Write to Write delay")
199 # time taken to complete one refresh cycle (N rows in all banks)
200 tRFC
= Param
.Latency("Refresh cycle time")
202 # refresh command interval, how often a "ref" command needs
203 # to be sent. It is 7.8 us for a 64ms refresh requirement
204 tREFI
= Param
.Latency("Refresh command interval")
206 # write-to-read, same rank turnaround penalty
207 tWTR
= Param
.Latency("Write to read, same rank switching time")
209 # write-to-read, same rank turnaround penalty for same bank group
210 tWTR_L
= Param
.Latency(Self
.tWTR
, "Write to read, same rank switching "
211 "time, same bank group")
213 # read-to-write, same rank turnaround penalty
214 tRTW
= Param
.Latency("Read to write, same rank switching time")
216 # rank-to-rank bus delay penalty
217 # this does not correlate to a memory timing parameter and encompasses:
218 # 1) RD-to-RD, 2) WR-to-WR, 3) RD-to-WR, and 4) WR-to-RD
219 # different rank bus delay
220 tCS
= Param
.Latency("Rank to rank switching time")
222 # minimum precharge to precharge delay time
223 tPPD
= Param
.Latency("0ns", "PRE to PRE delay")
225 # maximum delay between two-cycle ACT command phases
226 tAAD
= Param
.Latency(Self
.tCK
,
227 "Maximum delay between two-cycle ACT commands")
229 two_cycle_activate
= Param
.Bool(False,
230 "Two cycles required to send activate")
232 # minimum row activate to row activate delay time
233 tRRD
= Param
.Latency("ACT to ACT delay")
235 # only utilized with bank group architectures; set to 0 for default case
236 tRRD_L
= Param
.Latency("0ns", "Same bank group ACT to ACT delay")
238 # time window in which a maximum number of activates are allowed
239 # to take place, set to 0 to disable
240 tXAW
= Param
.Latency("X activation window")
241 activation_limit
= Param
.Unsigned("Max number of activates in window")
243 # time to exit power-down mode
244 # Exit power-down to next valid command delay
245 tXP
= Param
.Latency("0ns", "Power-up Delay")
247 # Exit Powerdown to commands requiring a locked DLL
248 tXPDLL
= Param
.Latency("0ns", "Power-up Delay with locked DLL")
250 # time to exit self-refresh mode
251 tXS
= Param
.Latency("0ns", "Self-refresh exit latency")
253 # time to exit self-refresh mode with locked DLL
254 tXSDLL
= Param
.Latency("0ns", "Self-refresh exit latency DLL")
256 # number of data beats per clock. with DDR, default is 2, one per edge
257 beats_per_clock
= Param
.Unsigned(2, "Data beats per clock")
259 data_clock_sync
= Param
.Bool(False, "Synchronization commands required")
261 # Currently rolled into other params
262 ######################################################################
264 # tRC - assumed to be tRAS + tRP
266 # Power Behaviour and Constraints
267 # DRAMs like LPDDR and WideIO have 2 external voltage domains. These are
268 # defined as VDD and VDD2. Each current is defined for each voltage domain
269 # separately. For example, current IDD0 is active-precharge current for
270 # voltage domain VDD and current IDD02 is active-precharge current for
271 # voltage domain VDD2.
272 # By default all currents are set to 0mA. Users who are only interested in
273 # the performance of DRAMs can leave them at 0.
275 # Operating 1 Bank Active-Precharge current
276 IDD0
= Param
.Current("0mA", "Active precharge current")
278 # Operating 1 Bank Active-Precharge current multiple voltage Range
279 IDD02
= Param
.Current("0mA", "Active precharge current VDD2")
281 # Precharge Power-down Current: Slow exit
282 IDD2P0
= Param
.Current("0mA", "Precharge Powerdown slow")
284 # Precharge Power-down Current: Slow exit multiple voltage Range
285 IDD2P02
= Param
.Current("0mA", "Precharge Powerdown slow VDD2")
287 # Precharge Power-down Current: Fast exit
288 IDD2P1
= Param
.Current("0mA", "Precharge Powerdown fast")
290 # Precharge Power-down Current: Fast exit multiple voltage Range
291 IDD2P12
= Param
.Current("0mA", "Precharge Powerdown fast VDD2")
293 # Precharge Standby current
294 IDD2N
= Param
.Current("0mA", "Precharge Standby current")
296 # Precharge Standby current multiple voltage range
297 IDD2N2
= Param
.Current("0mA", "Precharge Standby current VDD2")
299 # Active Power-down current: slow exit
300 IDD3P0
= Param
.Current("0mA", "Active Powerdown slow")
302 # Active Power-down current: slow exit multiple voltage range
303 IDD3P02
= Param
.Current("0mA", "Active Powerdown slow VDD2")
305 # Active Power-down current : fast exit
306 IDD3P1
= Param
.Current("0mA", "Active Powerdown fast")
308 # Active Power-down current : fast exit multiple voltage range
309 IDD3P12
= Param
.Current("0mA", "Active Powerdown fast VDD2")
311 # Active Standby current
312 IDD3N
= Param
.Current("0mA", "Active Standby current")
314 # Active Standby current multiple voltage range
315 IDD3N2
= Param
.Current("0mA", "Active Standby current VDD2")
317 # Burst Read Operating Current
318 IDD4R
= Param
.Current("0mA", "READ current")
320 # Burst Read Operating Current multiple voltage range
321 IDD4R2
= Param
.Current("0mA", "READ current VDD2")
323 # Burst Write Operating Current
324 IDD4W
= Param
.Current("0mA", "WRITE current")
326 # Burst Write Operating Current multiple voltage range
327 IDD4W2
= Param
.Current("0mA", "WRITE current VDD2")
330 IDD5
= Param
.Current("0mA", "Refresh current")
332 # Refresh Current multiple voltage range
333 IDD52
= Param
.Current("0mA", "Refresh current VDD2")
335 # Self-Refresh Current
336 IDD6
= Param
.Current("0mA", "Self-refresh Current")
338 # Self-Refresh Current multiple voltage range
339 IDD62
= Param
.Current("0mA", "Self-refresh Current VDD2")
341 # Main voltage range of the DRAM
342 VDD
= Param
.Voltage("0V", "Main Voltage Range")
344 # Second voltage range defined by some DRAMs
345 VDD2
= Param
.Voltage("0V", "2nd Voltage Range")
347 # A single DDR3-1600 x64 channel (one command and address bus), with
348 # timings based on a DDR3-1600 4 Gbit datasheet (Micron MT41J512M8) in
349 # an 8x8 configuration.
350 class DDR3_1600_8x8(DRAMCtrl
):
351 # size of device in bytes
352 device_size
= '512MB'
354 # 8x8 configuration, 8 devices each with an 8-bit interface
357 # DDR3 is a BL8 device
360 # Each device has a page (row buffer) size of 1 Kbyte (1K columns x8)
361 device_rowbuffer_size
= '1kB'
363 # 8x8 configuration, so 8 devices
367 ranks_per_channel
= 2
369 # DDR3 has 8 banks in all configurations
375 # 8 beats across an x64 interface translates to 4 clocks @ 800 MHz
390 # Greater of 4 CK or 7.5 ns
393 # Greater of 4 CK or 7.5 ns
396 # Default same rank rd-to-wr bus turnaround to 2 CK, @800 MHz = 2.5 ns
399 # Default different rank bus delay to 2 CK, @800 MHz = 2.5 ns
402 # <=85C, half for >85C
405 # active powerdown and precharge powerdown exit time
408 # self refresh exit time
411 # Current values from datasheet Die Rev E,J
423 # A single HMC-2500 x32 model based on:
424 # [1] DRAMSpec: a high-level DRAM bank modelling tool
425 # developed at the University of Kaiserslautern. This high level tool
426 # uses RC (resistance-capacitance) and CV (capacitance-voltage) models to
427 # estimate the DRAM bank latency and power numbers.
428 # [2] High performance AXI-4.0 based interconnect for extensible smart memory
429 # cubes (E. Azarkhish et. al)
430 # Assumed for the HMC model is a 30 nm technology node.
431 # The modelled HMC consists of 4 Gbit layers which sum up to 2GB of memory (4
433 # Each layer has 16 vaults and each vault consists of 2 banks per layer.
434 # In order to be able to use the same controller used for 2D DRAM generations
435 # for HMC, the following analogy is done:
436 # Channel (DDR) => Vault (HMC)
437 # device_size (DDR) => size of a single layer in a vault
438 # ranks per channel (DDR) => number of layers
439 # banks per rank (DDR) => banks per layer
440 # devices per rank (DDR) => devices per layer ( 1 for HMC).
441 # The parameters for which no input is available are inherited from the DDR3
443 # This configuration includes the latencies from the DRAM to the logic layer
445 class HMC_2500_1x32(DDR3_1600_8x8
):
447 # two banks per device with each bank 4MB [2]
450 # 1x32 configuration, 1 device with 32 TSVs [2]
451 device_bus_width
= 32
453 # HMC is a BL8 device [2]
456 # Each device has a page (row buffer) size of 256 bytes [2]
457 device_rowbuffer_size
= '256B'
459 # 1x32 configuration, so 1 device [2]
462 # 4 layers so 4 ranks [2]
463 ranks_per_channel
= 4
465 # HMC has 2 banks per layer [2]
466 # Each layer represents a rank. With 4 layers and 8 banks in total, each
467 # layer has 2 banks; thus 2 banks per rank.
473 # 8 beats across an x32 interface translates to 4 clocks @ 1250 MHz
476 # Values using DRAMSpec HMC model [1]
482 # tRRD depends on the power supply network for each vendor.
483 # We assume a tRRD of a double bank approach to be equal to 4 clock
484 # cycles (Assumption)
487 # activation limit is set to 0 since there are only 2 banks per vault
491 # Values using DRAMSpec HMC model [1]
496 # Default different rank bus delay assumed to 1 CK for TSVs, @1250 MHz =
497 # 0.8 ns (Assumption)
500 # Value using DRAMSpec HMC model [1]
503 # The default page policy in the vault controllers is simple closed page
504 # [2] nevertheless 'close' policy opens and closes the row multiple times
505 # for bursts largers than 32Bytes. For this reason we use 'close_adaptive'
506 page_policy
= 'close_adaptive'
508 # RoCoRaBaCh resembles the default address mapping in HMC
509 addr_mapping
= 'RoCoRaBaCh'
510 min_writes_per_switch
= 8
512 # These parameters do not directly correlate with buffer_size in real
513 # hardware. Nevertheless, their value has been tuned to achieve a
514 # bandwidth similar to the cycle-accurate model in [2]
515 write_buffer_size
= 32
516 read_buffer_size
= 32
518 # The static latency of the vault controllers is estimated to be smaller
519 # than a full DRAM channel controller
520 static_backend_latency
='4ns'
521 static_frontend_latency
='4ns'
523 # A single DDR3-2133 x64 channel refining a selected subset of the
524 # options for the DDR-1600 configuration, based on the same DDR3-1600
525 # 4 Gbit datasheet (Micron MT41J512M8). Most parameters are kept
526 # consistent across the two configurations.
527 class DDR3_2133_8x8(DDR3_1600_8x8
):
531 # 8 beats across an x64 interface translates to 4 clocks @ 1066 MHz
542 # Current values from datasheet
554 # A single DDR4-2400 x64 channel (one command and address bus), with
555 # timings based on a DDR4-2400 8 Gbit datasheet (Micron MT40A2G4)
556 # in an 16x4 configuration.
557 # Total channel capacity is 32GB
558 # 16 devices/rank * 2 ranks/channel * 1GB/device = 32GB/channel
559 class DDR4_2400_16x4(DRAMCtrl
):
563 # 16x4 configuration, 16 devices each with a 4-bit interface
566 # DDR4 is a BL8 device
569 # Each device has a page (row buffer) size of 512 byte (1K columns x4)
570 device_rowbuffer_size
= '512B'
572 # 16x4 configuration, so 16 devices
573 devices_per_rank
= 16
575 # Match our DDR3 configurations which is dual rank
576 ranks_per_channel
= 2
578 # DDR4 has 2 (x16) or 4 (x4 and x8) bank groups
579 # Set to 4 for x4 case
580 bank_groups_per_rank
= 4
582 # DDR4 has 16 banks(x4,x8) and 8 banks(x16) (4 bank groups in all
583 # configurations). Currently we do not capture the additional
584 # constraints incurred by the bank groups
587 # override the default buffer sizes and go for something larger to
588 # accommodate the larger bank count
589 write_buffer_size
= 128
590 read_buffer_size
= 64
595 # 8 beats across an x64 interface translates to 4 clocks @ 1200 MHz
596 # tBURST is equivalent to the CAS-to-CAS delay (tCCD)
597 # With bank group architectures, tBURST represents the CAS-to-CAS
598 # delay for bursts to different bank groups (tCCD_S)
601 # @2400 data rate, tCCD_L is 6 CK
602 # CAS-to-CAS delay for bursts to the same bank group
603 # tBURST is equivalent to tCCD_S; no explicit parameter required
604 # for CAS-to-CAS delay for bursts to different bank groups
613 # RRD_S (different bank group) for 512B page is MAX(4 CK, 3.3ns)
616 # RRD_L (same bank group) for 512B page is MAX(4 CK, 4.9ns)
619 # tFAW for 512B page is MAX(16 CK, 13ns)
627 # Here using the average of WTR_S and WTR_L
630 # Greater of 4 CK or 7.5 ns
633 # Default same rank rd-to-wr bus turnaround to 2 CK, @1200 MHz = 1.666 ns
636 # Default different rank bus delay to 2 CK, @1200 MHz = 1.666 ns
639 # <=85C, half for >85C
642 # active powerdown and precharge powerdown exit time
645 # self refresh exit time
646 # exit delay to ACT, PRE, PREALL, REF, SREF Enter, and PD Enter is:
647 # tRFC + 10ns = 340ns
650 # Current values from datasheet
665 # A single DDR4-2400 x64 channel (one command and address bus), with
666 # timings based on a DDR4-2400 8 Gbit datasheet (Micron MT40A1G8)
667 # in an 8x8 configuration.
668 # Total channel capacity is 16GB
669 # 8 devices/rank * 2 ranks/channel * 1GB/device = 16GB/channel
670 class DDR4_2400_8x8(DDR4_2400_16x4
):
671 # 8x8 configuration, 8 devices each with an 8-bit interface
674 # Each device has a page (row buffer) size of 1 Kbyte (1K columns x8)
675 device_rowbuffer_size
= '1kB'
677 # 8x8 configuration, so 8 devices
680 # RRD_L (same bank group) for 1K page is MAX(4 CK, 4.9ns)
685 # Current values from datasheet
692 # A single DDR4-2400 x64 channel (one command and address bus), with
693 # timings based on a DDR4-2400 8 Gbit datasheet (Micron MT40A512M16)
694 # in an 4x16 configuration.
695 # Total channel capacity is 4GB
696 # 4 devices/rank * 1 ranks/channel * 1GB/device = 4GB/channel
697 class DDR4_2400_4x16(DDR4_2400_16x4
):
698 # 4x16 configuration, 4 devices each with an 16-bit interface
699 device_bus_width
= 16
701 # Each device has a page (row buffer) size of 2 Kbyte (1K columns x16)
702 device_rowbuffer_size
= '2kB'
704 # 4x16 configuration, so 4 devices
707 # Single rank for x16
708 ranks_per_channel
= 1
710 # DDR4 has 2 (x16) or 4 (x4 and x8) bank groups
711 # Set to 2 for x16 case
712 bank_groups_per_rank
= 2
714 # DDR4 has 16 banks(x4,x8) and 8 banks(x16) (4 bank groups in all
715 # configurations). Currently we do not capture the additional
716 # constraints incurred by the bank groups
719 # RRD_S (different bank group) for 2K page is MAX(4 CK, 5.3ns)
722 # RRD_L (same bank group) for 2K page is MAX(4 CK, 6.4ns)
727 # Current values from datasheet
737 # A single LPDDR2-S4 x32 interface (one command/address bus), with
738 # default timings based on a LPDDR2-1066 4 Gbit part (Micron MT42L128M32D1)
739 # in a 1x32 configuration.
740 class LPDDR2_S4_1066_1x32(DRAMCtrl
):
745 device_size
= '512MB'
747 # 1x32 configuration, 1 device with a 32-bit interface
748 device_bus_width
= 32
750 # LPDDR2_S4 is a BL4 and BL8 device
753 # Each device has a page (row buffer) size of 1KB
754 # (this depends on the memory density)
755 device_rowbuffer_size
= '1kB'
757 # 1x32 configuration, so 1 device
761 ranks_per_channel
= 1
763 # LPDDR2-S4 has 8 banks in all configurations
772 # 8 CK read latency, 4 CK write latency @ 533 MHz, 1.876 ns cycle time
775 # Pre-charge one bank 15 ns (all banks 18 ns)
783 # 8 beats across an x32 DDR interface translates to 4 clocks @ 533 MHz.
784 # Note this is a BL8 DDR device.
785 # Requests larger than 32 bytes are broken down into multiple requests
793 # active powerdown and precharge powerdown exit time
796 # self refresh exit time
799 # Irrespective of speed grade, tWTR is 7.5 ns
802 # Default same rank rd-to-wr bus turnaround to 2 CK, @533 MHz = 3.75 ns
805 # Default different rank bus delay to 2 CK, @533 MHz = 3.75 ns
808 # Activate to activate irrespective of density and speed grade
811 # Irrespective of density, tFAW is 50 ns
815 # Current values from datasheet
837 # A single WideIO x128 interface (one command and address bus), with
838 # default timings based on an estimated WIO-200 8 Gbit part.
839 class WideIO_200_1x128(DRAMCtrl
):
844 device_size
= '1024MB'
846 # 1x128 configuration, 1 device with a 128-bit interface
847 device_bus_width
= 128
849 # This is a BL4 device
852 # Each device has a page (row buffer) size of 4KB
853 # (this depends on the memory density)
854 device_rowbuffer_size
= '4kB'
856 # 1x128 configuration, so 1 device
859 # Use one rank for a one-high die stack
860 ranks_per_channel
= 1
862 # WideIO has 4 banks in all configurations
874 # Read to precharge is same as the burst
877 # 4 beats across an x128 SDR interface translates to 4 clocks @ 200 MHz.
878 # Note this is a BL4 SDR device.
884 # WIO 8 Gb, <=85C, half for >85C
887 # Greater of 2 CK or 15 ns, 2 CK @ 200 MHz = 10 ns
890 # Default same rank rd-to-wr bus turnaround to 2 CK, @200 MHz = 10 ns
893 # Default different rank bus delay to 2 CK, @200 MHz = 10 ns
896 # Activate to activate irrespective of density and speed grade
899 # Two instead of four activation window
903 # The WideIO specification does not provide current information
905 # A single LPDDR3 x32 interface (one command/address bus), with
906 # default timings based on a LPDDR3-1600 4 Gbit part (Micron
907 # EDF8132A1MC) in a 1x32 configuration.
908 class LPDDR3_1600_1x32(DRAMCtrl
):
913 device_size
= '512MB'
915 # 1x32 configuration, 1 device with a 32-bit interface
916 device_bus_width
= 32
918 # LPDDR3 is a BL8 device
921 # Each device has a page (row buffer) size of 4KB
922 device_rowbuffer_size
= '4kB'
924 # 1x32 configuration, so 1 device
927 # Technically the datasheet is a dual-rank package, but for
928 # comparison with the LPDDR2 config we stick to a single rank
929 ranks_per_channel
= 1
931 # LPDDR3 has 8 banks in all configurations
939 # 12 CK read latency, 6 CK write latency @ 800 MHz, 1.25 ns cycle time
945 # Greater of 4 CK or 7.5 ns, 4 CK @ 800 MHz = 5 ns
948 # Pre-charge one bank 18 ns (all banks 21 ns)
951 # 8 beats across a x32 DDR interface translates to 4 clocks @ 800 MHz.
952 # Note this is a BL8 DDR device.
953 # Requests larger than 32 bytes are broken down into multiple requests
961 # active powerdown and precharge powerdown exit time
964 # self refresh exit time
967 # Irrespective of speed grade, tWTR is 7.5 ns
970 # Default same rank rd-to-wr bus turnaround to 2 CK, @800 MHz = 2.5 ns
973 # Default different rank bus delay to 2 CK, @800 MHz = 2.5 ns
976 # Activate to activate irrespective of density and speed grade
979 # Irrespective of size, tFAW is 50 ns
983 # Current values from datasheet
1005 # A single GDDR5 x64 interface, with
1006 # default timings based on a GDDR5-4000 1 Gbit part (SK Hynix
1007 # H5GQ1H24AFR) in a 2x32 configuration.
1008 class GDDR5_4000_2x32(DRAMCtrl
):
1010 device_size
= '128MB'
1012 # 2x32 configuration, 1 device with a 32-bit interface
1013 device_bus_width
= 32
1015 # GDDR5 is a BL8 device
1018 # Each device has a page (row buffer) size of 2Kbits (256Bytes)
1019 device_rowbuffer_size
= '256B'
1021 # 2x32 configuration, so 2 devices
1022 devices_per_rank
= 2
1024 # assume single rank
1025 ranks_per_channel
= 1
1027 # GDDR5 has 4 bank groups
1028 bank_groups_per_rank
= 4
1030 # GDDR5 has 16 banks with 4 bank groups
1036 # 8 beats across an x64 interface translates to 2 clocks @ 1000 MHz
1037 # Data bus runs @2000 Mhz => DDR ( data runs at 4000 MHz )
1038 # 8 beats at 4000 MHz = 2 beats at 1000 MHz
1039 # tBURST is equivalent to the CAS-to-CAS delay (tCCD)
1040 # With bank group architectures, tBURST represents the CAS-to-CAS
1041 # delay for bursts to different bank groups (tCCD_S)
1044 # @1000MHz data rate, tCCD_L is 3 CK
1045 # CAS-to-CAS delay for bursts to the same bank group
1046 # tBURST is equivalent to tCCD_S; no explicit parameter required
1047 # for CAS-to-CAS delay for bursts to different bank groups
1052 # tCL is not directly found in datasheet and assumed equal tRCD
1058 # RRD_S (different bank group)
1059 # RRD_S is 5.5 ns in datasheet.
1060 # rounded to the next multiple of tCK
1063 # RRD_L (same bank group)
1064 # RRD_L is 5.5 ns in datasheet.
1065 # rounded to the next multiple of tCK
1071 # Therefore, activation limit is set to 0
1072 activation_limit
= 0
1077 # Here using the average of WTR_S and WTR_L
1080 # Read-to-Precharge 2 CK
1086 # A single HBM x128 interface (one command and address bus), with
1087 # default timings based on data publically released
1088 # ("HBM: Memory Solution for High Performance Processors", MemCon, 2014),
1089 # IDD measurement values, and by extrapolating data from other classes.
1090 # Architecture values based on published HBM spec
1091 # A 4H stack is defined, 2Gb per die for a total of 1GB of memory.
1092 class HBM_1000_4H_1x128(DRAMCtrl
):
1093 # HBM gen1 supports up to 8 128-bit physical channels
1094 # Configuration defines a single channel, with the capacity
1095 # set to (full_ stack_capacity / 8) based on 2Gb dies
1096 # To use all 8 channels, set 'channels' parameter to 8 in
1097 # system configuration
1099 # 128-bit interface legacy mode
1100 device_bus_width
= 128
1102 # HBM supports BL4 and BL2 (legacy mode only)
1105 # size of channel in bytes, 4H stack of 2Gb dies is 1GB per stack;
1106 # with 8 channels, 128MB per channel
1107 device_size
= '128MB'
1109 device_rowbuffer_size
= '2kB'
1111 # 1x128 configuration
1112 devices_per_rank
= 1
1114 # HBM does not have a CS pin; set rank to 1
1115 ranks_per_channel
= 1
1117 # HBM has 8 or 16 banks depending on capacity
1118 # 2Gb dies have 8 banks
1121 # depending on frequency, bank groups may be required
1122 # will always have 4 bank groups when enabled
1123 # current specifications do not define the minimum frequency for
1124 # bank group architecture
1125 # setting bank_groups_per_rank to 0 to disable until range is defined
1126 bank_groups_per_rank
= 0
1128 # 500 MHz for 1Gbps DDR data rate
1131 # use values from IDD measurement in JEDEC spec
1132 # use tRP value for tRCD and tCL similar to other classes
1138 # BL2 and BL4 supported, default to BL4
1139 # DDR @ 500 MHz means 4 * 2ns / 2 = 4ns
1142 # value for 2Gb device from JEDEC spec
1145 # value for 2Gb device from JEDEC spec
1148 # extrapolate the following from LPDDR configs, using ns values
1149 # to minimize burst length, prefetch differences
1154 # start with 2 cycles turnaround, similar to other memory classes
1155 # could be more with variations across the stack
1158 # single rank device, set to 0
1161 # from MemCon example, tRRD is 4ns with 2ns tCK
1164 # from MemCon example, tFAW is 30ns with 2ns tCK
1166 activation_limit
= 4
1171 # start with tRFC + tXP -> 160ns + 8ns = 168ns
1174 # A single HBM x64 interface (one command and address bus), with
1175 # default timings based on HBM gen1 and data publically released
1176 # A 4H stack is defined, 8Gb per die for a total of 4GB of memory.
1177 # Note: This defines a pseudo-channel with a unique controller
1178 # instantiated per pseudo-channel
1179 # Stay at same IO rate (1Gbps) to maintain timing relationship with
1180 # HBM gen1 class (HBM_1000_4H_x128) where possible
1181 class HBM_1000_4H_1x64(HBM_1000_4H_1x128
):
1182 # For HBM gen2 with pseudo-channel mode, configure 2X channels.
1183 # Configuration defines a single pseudo channel, with the capacity
1184 # set to (full_ stack_capacity / 16) based on 8Gb dies
1185 # To use all 16 pseudo channels, set 'channels' parameter to 16 in
1186 # system configuration
1188 # 64-bit pseudo-channle interface
1189 device_bus_width
= 64
1191 # HBM pseudo-channel only supports BL4
1194 # size of channel in bytes, 4H stack of 8Gb dies is 4GB per stack;
1195 # with 16 channels, 256MB per channel
1196 device_size
= '256MB'
1198 # page size is halved with pseudo-channel; maintaining the same same number
1199 # of rows per pseudo-channel with 2X banks across 2 channels
1200 device_rowbuffer_size
= '1kB'
1202 # HBM has 8 or 16 banks depending on capacity
1203 # Starting with 4Gb dies, 16 banks are defined
1206 # reset tRFC for larger, 8Gb device
1207 # use HBM1 4Gb value as a starting point
1210 # start with tRFC + tXP -> 160ns + 8ns = 168ns
1212 # Default different rank bus delay to 2 CK, @1000 MHz = 2 ns
1216 # active powerdown and precharge powerdown exit time
1219 # self refresh exit time
1222 # A single LPDDR5 x16 interface (one command/address bus)
1223 # for a single x16 channel with default timings based on
1224 # initial JEDEC specification
1225 # Starting with 5.5Gbps data rates and 8Gbit die
1226 # Configuring for 16-bank mode with bank-group architecture
1227 # burst of 32, which means bursts can be interleaved
1228 class LPDDR5_5500_1x16_BG_BL32(DRAMCtrl
):
1230 # Increase buffer size to account for more bank resources
1231 read_buffer_size
= 64
1233 # Set page policy to better suit DMC Huxley
1234 page_policy
= 'close_adaptive'
1236 # 16-bit channel interface
1237 device_bus_width
= 16
1239 # LPDDR5 is a BL16 or BL32 device
1240 # With BG mode, BL16 and BL32 are supported
1241 # Use BL32 for higher command bandwidth
1244 # size of device in bytes
1247 # 2kB page with BG mode
1248 device_rowbuffer_size
= '2kB'
1250 # Use a 1x16 configuration
1251 devices_per_rank
= 1
1254 ranks_per_channel
= 1
1256 # LPDDR5 supports configurable bank options
1257 # 8B : BL32, all frequencies
1258 # 16B : BL32 or BL16, <=3.2Gbps
1259 # 16B with Bank Group Arch (4B/BG): BL32 or BL16, >3.2Gbps
1260 # Initial configuration will have 16 banks with Bank Group Arch
1261 # to maximim resources and enable higher data rates
1263 bank_groups_per_rank
= 4
1265 # 5.5Gb/s DDR with 4:1 WCK:CK ratio for 687.5 MHz CK
1268 # Greater of 2 CK or 18ns
1271 # Base RL is 16 CK @ 687.5 MHz = 23.28ns
1274 # Greater of 2 CK or 18ns
1277 # Greater of 3 CK or 42ns
1280 # Greater of 3 CK or 34ns
1283 # active powerdown and precharge powerdown exit time
1284 # Greater of 3 CK or 7ns
1287 # self refresh exit time (tRFCab + 7.5ns)
1290 # Greater of 2 CK or 7.5 ns minus 2 CK
1293 # With BG architecture, burst of 32 transferred in two 16-beat
1294 # sub-bursts, with a 16-beat gap in between.
1295 # Each 16-beat sub-burst is 8 WCK @2.75 GHz or 2 CK @ 687.5 MHz
1296 # tBURST is the delay to transfer the Bstof32 = 6 CK @ 687.5 MHz
1298 # can interleave a Bstof32 from another bank group at tBURST_MIN
1299 # 16-beats is 8 WCK @2.75 GHz or 2 CK @ 687.5 MHz
1300 tBURST_MIN
= '2.91ns'
1301 # tBURST_MAX is the maximum burst delay for same bank group timing
1302 # this is 8 CK @ 687.5 MHz
1303 tBURST_MAX
= '11.64ns'
1308 # LPDDR5, 8 Gbit/channel for 280ns tRFCab
1312 # Greater of 4 CK or 6.25 ns
1314 # Greater of 4 CK or 12 ns
1317 # Required RD-to-WR timing is RL+ BL/n + tWCKDQ0/tCK - WL
1318 # tWCKDQ0/tCK will be 1 CK for most cases
1319 # For gem5 RL = WL and BL/n is already accounted for with tBURST
1320 # Result is and additional 1 CK is required
1323 # Default different rank bus delay to 2 CK, @687.5 MHz = 2.91 ns
1329 # Greater of 2 CK or 5 ns
1333 # With Bank Group Arch mode tFAW is 20 ns
1335 activation_limit
= 4
1337 # at 5Gbps, 4:1 WCK to CK ratio required
1338 # 2 data beats per WCK (DDR) -> 8 per CK
1341 # 2 cycles required to send activate command
1342 # 2 command phases can be sent back-to-back or
1343 # with a gap up to tAAD = 8 CK
1344 two_cycle_activate
= True
1347 data_clock_sync
= True
1349 # A single LPDDR5 x16 interface (one command/address bus)
1350 # for a single x16 channel with default timings based on
1351 # initial JEDEC specification
1352 # Starting with 5.5Gbps data rates and 8Gbit die
1353 # Configuring for 16-bank mode with bank-group architecture, burst of 16
1354 class LPDDR5_5500_1x16_BG_BL16(LPDDR5_5500_1x16_BG_BL32
):
1356 # LPDDR5 is a BL16 or BL32 device
1357 # With BG mode, BL16 and BL32 are supported
1358 # Use BL16 for smaller access granularity
1361 # For Bstof16 with BG arch, 2 CK @ 687.5 MHz with 4:1 clock ratio
1363 tBURST_MIN
= '2.91ns'
1364 # For Bstof16 with BG arch, 4 CK @ 687.5 MHz with 4:1 clock ratio
1365 tBURST_MAX
= '5.82ns'
1371 # A single LPDDR5 x16 interface (one command/address bus)
1372 # for a single x16 channel with default timings based on
1373 # initial JEDEC specification
1374 # Starting with 5.5Gbps data rates and 8Gbit die
1375 # Configuring for 8-bank mode, burst of 32
1376 class LPDDR5_5500_1x16_8B_BL32(LPDDR5_5500_1x16_BG_BL32
):
1378 # 4kB page with 8B mode
1379 device_rowbuffer_size
= '4kB'
1381 # LPDDR5 supports configurable bank options
1382 # 8B : BL32, all frequencies
1383 # 16B : BL32 or BL16, <=3.2Gbps
1384 # 16B with Bank Group Arch (4B/BG): BL32 or BL16, >3.2Gbps
1387 bank_groups_per_rank
= 0
1389 # For Bstof32 with 8B mode, 4 CK @ 687.5 MHz with 4:1 clock ratio
1391 tBURST_MIN
= '5.82ns'
1392 tBURST_MAX
= '5.82ns'
1394 # Greater of 4 CK or 12 ns
1397 # Greater of 2 CK or 10 ns
1400 # With 8B mode tFAW is 40 ns
1402 activation_limit
= 4
1404 # Reset BG arch timing for 8B mode
1409 # A single LPDDR5 x16 interface (one command/address bus)
1410 # for a single x16 channel with default timings based on
1411 # initial JEDEC specification
1412 # 6.4Gbps data rates and 8Gbit die
1413 # Configuring for 16-bank mode with bank-group architecture
1414 # burst of 32, which means bursts can be interleaved
1415 class LPDDR5_6400_1x16_BG_BL32(LPDDR5_5500_1x16_BG_BL32
):
1417 # 5.5Gb/s DDR with 4:1 WCK:CK ratio for 687.5 MHz CK
1420 # Base RL is 17 CK @ 800 MHz = 21.25ns
1423 # With BG architecture, burst of 32 transferred in two 16-beat
1424 # sub-bursts, with a 16-beat gap in between.
1425 # Each 16-beat sub-burst is 8 WCK @3.2 GHz or 2 CK @ 800 MHz
1426 # tBURST is the delay to transfer the Bstof32 = 6 CK @ 800 MHz
1428 # can interleave a Bstof32 from another bank group at tBURST_MIN
1429 # 16-beats is 8 WCK @2.3 GHz or 2 CK @ 800 MHz
1430 tBURST_MIN
= '2.5ns'
1431 # tBURST_MAX is the maximum burst delay for same bank group timing
1432 # this is 8 CK @ 800 MHz
1438 # Required RD-to-WR timing is RL+ BL/n + tWCKDQ0/tCK - WL
1439 # tWCKDQ0/tCK will be 1 CK for most cases
1440 # For gem5 RL = WL and BL/n is already accounted for with tBURST
1441 # Result is and additional 1 CK is required
1444 # Default different rank bus delay to 2 CK, @687.5 MHz = 2.5 ns
1450 # 2 command phases can be sent back-to-back or
1451 # with a gap up to tAAD = 8 CK
1454 # A single LPDDR5 x16 interface (one command/address bus)
1455 # for a single x16 channel with default timings based on initial
1456 # JEDEC specifcation
1457 # 6.4Gbps data rates and 8Gbit die
1458 # Configuring for 16-bank mode with bank-group architecture, burst of 16
1459 class LPDDR5_6400_1x16_BG_BL16(LPDDR5_6400_1x16_BG_BL32
):
1461 # LPDDR5 is a BL16 or BL32 device
1462 # With BG mode, BL16 and BL32 are supported
1463 # Use BL16 for smaller access granularity
1466 # For Bstof16 with BG arch, 2 CK @ 800 MHz with 4:1 clock ratio
1468 tBURST_MIN
= '2.5ns'
1469 # For Bstof16 with BG arch, 4 CK @ 800 MHz with 4:1 clock ratio
1476 # A single LPDDR5 x16 interface (one command/address bus)
1477 # for a single x16 channel with default timings based on
1478 # initial JEDEC specification
1479 # 6.4Gbps data rates and 8Gbit die
1480 # Configuring for 8-bank mode, burst of 32
1481 class LPDDR5_6400_1x16_8B_BL32(LPDDR5_6400_1x16_BG_BL32
):
1483 # 4kB page with 8B mode
1484 device_rowbuffer_size
= '4kB'
1486 # LPDDR5 supports configurable bank options
1487 # 8B : BL32, all frequencies
1488 # 16B : BL32 or BL16, <=3.2Gbps
1489 # 16B with Bank Group Arch (4B/BG): BL32 or BL16, >3.2Gbps
1492 bank_groups_per_rank
= 0
1494 # For Bstof32 with 8B mode, 4 CK @ 800 MHz with 4:1 clock ratio
1499 # Greater of 4 CK or 12 ns
1502 # Greater of 2 CK or 10 ns
1505 # With 8B mode tFAW is 40 ns
1507 activation_limit
= 4
1509 # Reset BG arch timing for 8B mode