Failed DRAM training/initialization on several DDR5 RDIMMs

[pjattke]

Hello, we encountered issues during the DRAM training on some of our DDR5 RDIMMs with the ddr5-tester platform.

The bitstreams were built with the following command:

make -j"$(nproc)" build TARGET_ARGS="--l2-size 256 \
   --build \
   --iodelay-clk-freq 400e6 \
   --bios-lto \
   --rw-bios \
   --from-spd $(realpath "$spd_file") \
   --no-sdram-hw-test"

The bitstream is based on commit 998d4afeb of the rowhammer-tester repository.

Testing Procedure

For each of the DIMMs listed in "Tested Devices", we flash the bitstream, then we run the DRAM training and memtest. We test each DIMM for 10 times but stop early, after five successful runs. We reboot and reflash the FPGA between test iterations.

We used the DDR5 tester board with SN18 and board revision 1.0.1 for all tests.

Tested Devices

✔ – devices that passed the memory training and/or memtest.
✘ – devices that failed the memory training and/or memtest.

Module Manuf.	Chip Manuf.	Size	Module Model No.	Int. ID	Result
Micron	Micron	16 GB	MTC10F1084S1RC48BA1	300	✘ (0/10)
Kingston	SK Hynix	16 GB	KSM48R40BS8KMM-16HMR	304	✔
Micron	Micron	32 GB	MTC20F1045S1RC48BA2	301	✔
Micron	Micron	32 GB	MTC18F1045S1PC48BA2	302	✘ (0/10)
Kingston	SK Hynix	32 GB	KSM48R40BS4TMM-32HMR	303	✘ (0/10)
Samsung	Samsung	32 GB	M321R4GA0BB0-CQKMS	308	✘ (0/10)
SK Hynix	SK Hynix	32 GB	HMCG88AGBRA188N	306	✘ (0/10)
Samsung	Samsung	32 GB	M321R4GA3BB6-CQKET	307	✔
Kingston	SK Hynix	32 GB	KSM48R40BD8KMM-32HMR	310	✔
Micron	Micron	32 GB	MTC20F2085S1RC48BA1	311	✘ (1/10)
Kingston	SK Hynix	64 GB	KTL-TS548D4-64G	313	✘ (0/10)
Kingston	SK Hynix	64 GB	KSM48R40BD4TMM-64HMR	312	✘ (0/10)

Interestingly, some of these devices (e.g., DIMM 300) have passed the memory training before (with a much older bitstream) but don't seem to work with the latest bitstream anymore. Has anything changed in the training procedure?

Logfiles

We attach the logfiles of the devices that failed for further analysis: logfiles.tar.gz.