Introduction

This datasheet provides the performance numbers of various device drivers in MCU PLUS SDK for AM243X

Generic Setup details

SOC Details	Values
Core	R5F
Core Operating Speed	800 MHz
Cache Status	Enabled

Optimization Details	Values
Build Profile	Release
R5F Compiler flags	-mcpu=cortex-r5 -mfloat-abi=hard -mfpu=vfpv3-d16 -Wall -Werror -g -mthumb -Wno-gnu-variable-sized-type-not-at-end -Wno-unused-function
R5F Linker flags	-Wl,–diag_suppress=10063 -Wl,–ram_model -Wl,–reread_libs
Code Placement	MSRAM
Data Placement	MSRAM

Performance Numbers

AM243X-EVM

Software/Application used : sbl_ospi and ipc_notify_echo
Size of sbl_ospi appimage : 321 KB
Size of ipc_notify_echo appimage : 81 KB
Cores present in the appimage : m4f0-0, r5f0-0, r5f0-1, r5f1-0, r5f1-1
Boot Media Clock : 166.667 MHz (Octal DDR mode)

Boot time breakdown	Time (ms)
ROM : init + SBL load from flash	12.00
SBL : SYSFW Load	10.823
SBL : System Init	14.251
SBL : Drivers_open	1.664
SBL : Board_driversOpen	28.916
SBL : Sciclient Get Version	9.937
SBL : CPU Load	19.159
SBL : Total time taken	84.755

The time taken for Board_driversOpen (around 20 ms) is mostly for the PHY tuning of OSPI. If this needs to be further reduced, one can pre-train the PHY, note down delay values and set it directly instead of the tuning procedure.
Most of the time taken for the CPU load is in loading the M4F. Other CPUs take ~1.5ms each. This needs more investigation. Possible reason is the M4F IRAM and DRAM initialization time.
Here the CPU load or section copy takes place from the OSPI memory to the respective memory directly, this would be slower that mem to mem copy.
The time taken for Sciclient Get Version can be avoided if the version check is disabled

SBL SD performance

Software/Application used : sbl_sd and adc_singleshot
Size of sbl_sd appimage : 339 KB
Size of adc_singleshot appimage : 54 KB
Cores present in the appimage : r5f0-0
SD card read speed : 12.5 MBps mode

Boot time breakdown	Time (ms)
ROM : init + SBL load from flash	12.00
SBL : SYSFW Load	10.846
SBL : System Init	13.899
SBL : Drivers_open	106.915
SBL : Board_driversOpen	0.00
SBL : Sciclient Get Version	9.937
SBL : File read from SD Card	6.760
SBL : CPU Load	13.801
-------------------------------------—	-----------—
SBL : Total time taken	162.161

The MMCSD driver and SD card initialization is done as part of Drivers_open, so Board_driversOpen happens instantaneously.
Most of the time taken for the CPU load is in loading the M4F. Other CPUs take ~1.5ms each. This needs more investigation. Possible reason is the M4F IRAM and DRAM initialization time.
The CPU load is faster in this case because SD bootloader is at its core a memory bootloader. Appimage file from SD card is read into the memory (this takes 15 ms in this case). The CPU load now is mostly memcpy in the OCRAM. This would be quite fast. The only caveat here is that the buffer size to receive appimage is allocated in OCRAM, so it's limited. This is not the case in OSPI.
The time taken for Sciclient Get Version can be avoided if the version check is disabled

XIP Benchmark

Software/Application used : xip_benchmark
FIR operation

Caching status	Cycles taken
Code/Data fully cached	23961
Code/Data not cached	48850
Code/Data not cached 1 of 10 iterations	26492

MEMCPY operation

Caching status	Cycles taken
Code/Data fully cached	1565
Code/Data not cached	7612
Code/Data not cached 1 of 10 iterations	2466

MCAN performance

Internal loopback mode of operation
Memory Mode : TX Buffer Mode, RX FIFO Mode
MCAN CLK : 80MHz
CPU : R5F

Frame Type	Arbitration BitRate(Mbps)	Data BitRate(Mbps)	Arb Phase bits	Data Phase bits	Throughput (Msg/Sec)
CAN FD STANDARD FORMAT	1	5	27	538	6666
CAN FD EXTENDED FORMAT	1	5	46	538	5780

MCSPI performance

MCSPI driver performance in different modes

measurement is provided for Master mode in Loopback configuration with TX and RX.
Transfer time captured with driver API in different operation mode.
Throughput comparison is added for transfer of 400 Bytes with different word width and corresponding number of words.
Time
CPU : R5F
Transfer Length : 400 Bytes
SPI Clock speed : 50MHz

Number of Words	Word Width (Bits)	Polled mode Throughput / Transfer time	Interrupt mode (Mbps) Throughput / Transfer time	Dma mode (Mbps) Throughput / Transfer time
400	08	10.55 Mbps / 303.29 us	23.51 Mbps / 136.09 us	0.91 Mbps / 3502.98 us
200	16	21.13 Mbps / 151.43 us	31.06 Mbps / 103.02 us	0.95 Mbps / 3359.05 us
100	32	38.22 Mbps / 83.73 us	36.95 Mbps / 86.60 us	0.97 Mbps / 3287.17 us

Theoretically for 400 Bytes at 50MHz time required for clocks is 64us.
Additionally hardware adds 160ns + 1bit time delay between each word transfer as measured in the scope

CPSW Performance

For CPSW performance refer Ethernet Performance

ICSSG Performance

TCP Test

Software/Application used : enet_lwip_icssg
iperf test type : TCP iperf
iperf command used : iperf -c 192.168.0.158 -i 10 -t 100 -r
Packet allocation : 32 buffers per each Rx flow, 16 buffers for Tx channel

TCP Performance(am64x/243x emvs):

TCP direction	B/W (Mb/s)	CPU load(%)
TCP RX	93.2	52.1
TCP TX	92.9	87.57

TCP Performance(am243x-lp):

TCP direction	B/W (Mb/s)	CPU load(%)
TCP RX	93.20	36.82
TCP TX	93.30	52.54

UDP Test

Software/Application used : enet_lwip_icssg
iperf test type : UDP iperf
iperf command used : iperf -c 192.168.0.158 -u -b 60M -l 256 -i 10 -t 100 -r
Packet allocation : 32 buffers per each Rx flow, 16 buffers for Tx channel

UDP TX Performance(am64x/243x emvs):

For Packet Size: 1500 B

Tx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	100	0	-
50 Mbps	100	0.03	-
100 Mbps	100	0.07	-
Max(220 Mbps)	100	0.07	-

UDP TX Performance(am243x-lp):

For Packet Size: 1500 B

Tx Side BW/Packet Size	CPU load(%)	Latency
25 Mbps	100	-
50 Mbps	100	-
100 Mbps	100	-

UDP RX Performance(am64x/243x emvs):

For Packet Size: 256 B

Rx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	44.35	0.11	-
50 Mbps	60.34	41	-
100 Mbps	-	-	-

For Packet Size: 512 B

Rx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	21.83	0	-
50 Mbps	56.74	0.09	-
100 Mbps	65.61	41	-

For Packet Size: 1500 B

Rx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	16.84	0	-
50 Mbps	28.93	0	-
90 Mbps	39	0.7	-

UDP RX Performance(am243x-lp):

For Packet Size: 256 B

Rx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	37.35	0.12	-
50 Mbps	48	36	-
100 Mbps	-	-	-

For Packet Size: 512 B

Rx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	21.83	0	-
50 Mbps	40	0.12	-
100 Mbps	50	36	-

For Packet Size: 1500 B

Rx Side BW/Packet Size	CPU load(%)	Packet loss(%)	Latency
25 Mbps	15	0	-
50 Mbps	26.5	0.08	-
89 Mbps	33	0.1	-

IPC performance

IPC NOTIFY

10000 messages are sent and average one way message latency is measured

Local Core	Remote Core	Average Message Latency (us)
r5f0-0	m4f0-0	1.52
r5f0-0	r5f0-1	0.81
r5f0-0	r5f1-0	0.88
r5f0-0	r5f1-1	0.93
r5f0-0	a530-0	0.00
r5f0-0	a530-1	0.00

IPC RPMSG

10000 messages are sent and average one way message latency is measured

Local Core	Remote Core	Message Size	Average Message Latency (us)
r5f0-0	r5f0-1	4	6.345
r5f0-0	r5f1-0	4	6.347
r5f0-0	r5f1-1	4	6.343
r5f0-0	m4f0-0	4	11.751
r5f0-0	r5f0-1	32	8.546
r5f0-0	r5f0-1	64	11.094
r5f0-0	r5f0-1	112	14.888
r5f0-0	m4f0-0	32	17.012
r5f0-0	m4f0-0	64	22.996
r5f0-0	m4f0-0	112	32.060

MATHLIB

MATHLIB BENCHMARK

Calculated for the 500 samples taken between 0 and 2 * Pi
Trignometric function timings compared between the optimized Mathlib mcusdk implementation and the compiler mathlib version

The max error for each operation between the optimized Mathlib mcusdk functions and the compiler mathlib version is printed

Function	Err	Max Cycles Mathlib (mcusdk)	avg cycles Mathlib (mcusdk)	max cycles mathlib (clang)	avg cycles mathlib (clang)
sin	0.0000007150	52	52.139999	757	278.600006
cos	0.0000002870	65	65.220001	951	280.640015
sincos sin	0.0000001790	79	78.895996	477	275.376007
sincos cos	0.0000001900
asin	0.0000003430	74	74.071999	1175	431.036011
acos	0.0000004770	76	76.000000	1098	385.519989
atan	0.0000005360	80	80.082001	922	373.182007
atan2	0.0000007150	117	104.704002	681	480.558014

SA2UL

PKA RSA SIGN VERIFY

Software/Application used : test_sa2ul_pka
Supported keys : 4K and 2K
CPU with operating speed : R5F with 800MHZ
OS used : nortos

ECDSA Sign/sec Verify/sec Sign and verify/sec

2048 631 8771 589

4096 97 2770 94

SHA

Software/Application used : test_sa2ul_sha
Code Placement : OCMC
Data Placement : OCMC
Input Data sizes : 512B, 1KB, 2KB, 4KB, 8KB, 16KB and 32KB

CPU with operating speed : R5F with 800MHZ

SHA	Size	Performance (Mbps)
512	32.00 KB	2037.890154
512	16.00 KB	1811.293638
512	8.00 KB	1417.569285
512	4.00 KB	980.344054
512	2.00 KB	602.076252
512	1024.00 B	341.511204
512	512.00 B	180.182558
256	32.00 KB	1588.546930
256	16.00 KB	1411.482184
256	8.00 KB	1156.424113
256	4.00 KB	837.386999
256	2.00 KB	549.683372
256	1024.00 B	325.111618
256	512.00 B	175.652640

PKA ECDSA

Software/Application used : test_sa2ul_pka
Supported Curves : p-256 and p-384
CPU with operating speed : R5F with 800MHZ
OS used : nortos

ECDSA Sign/sec Verify/sec Sign and verify/sec

256 1135 693 430

384 585 349 218

AES

Software/Application used : test_sa2ul_aes
Code Placement : OCMC
Data Placement : OCMC
Input Data sizes : 512B, 1KB, 2KB, 4KB, 8KB, 16KB and 32KB