Current status for one test (eth-l2xcbase 1c 64b Intel-xxv710 drv_vfio_pci MRR), interval of values in Mpps bidirectional (in brackets value seen before latest firmware upgrade):
TB212: 30.24 - 30.30 (30.6)
TB213: 29.76 - 29.84 (30.0)
TB214: 30.83 - 30.90 (31.1)
TB215: 30.82 - 30.89 (31.1)

TBs 214 and 215 can be considered good and identical. TB212 has different PCI address of port2 on DUT (0000:17:00.1 instead of 0000:19:00.0 on other testbeds), but I am not sure that can explain the performance difference. I do not see any clear signal from telemetry that might explain why TB213 is so slow.
NDR and PDR values are always quite close to MRR, soak values are too noisy.

The interval values come from runs with 1000x1s trials. The values in trending [0] are higher, basically the same as before firmware upgrade. As the difference between good and bad testbeds is around 3%, this creates false progressions and regressions in trending.

[0] https://csit.fd.io/trending/#eNp1kM8OgjAMxp9mXkwTNsV58SDyHmZsVRb5M7ZpwKd3GJPCwUub5td-X9sQe4_XgM2J5QWTBRPSmhTY7rxN6eUciA6sHoFn2R2F43jkmR7AOPOARsBhXwHXgLFO1agrFRBa72cJcZklzDOu9Ii4eiLy34UGlEdFE40gEjEsXJbLUM_NqxaDfSM1puWJ6_QJQlyv1ePkFvR3oSxZXm663rff9-XlB9n2WlM