QSFP compatibility of the Bluefield-2 DPU

Hi all,

I finally managed to order my Bluefield-2 CECOT DPUs but the provider could not offer cabling for purchase.

According to this NVidia post, QSFP ports are backward compatible; in fact, from QSFP-DD (400GB) all the way to QSFP+ (40G) are all backward compatible. However, the modulation used might be different. So, I am afraid now of purchasing a QSFP28 cable (using NRZ modulation) for the Bluefield-2 DPU having QSFP56 (PAM4 modulation).
Will it be compatible?

The second question is whether I should get a certain brand, or all these cables (should) strictly follow the standards and they don’t do some nasty tricks to do vendor-lockin? Like back in the days when certain mobile phones even though were using standard micro USB port, the pins were purposely swapped to disable charging it with other than the brand’s own charger. I know that this question is probably not a 100% sure Yes/No, as noone knows what random vendor might use; but any suggestions? For instance, can I buy QSFP cables from Mikrotik (originally sold for Mikrotik routers)?



You should use one of the supported cables/modules listed in the following page:

Best Regards,

Thanks, but all of these are NVIDIA cables only. I could not order them. In fact, even the Bluefield 2 DPU could not be ordered from nvidia store. I asked several times for a quotation, but no response. Luckily, someone was selling them on eBay.
I will be testing soon a cable from Mikrotik and share my experiences.

As I have no other option in my region, I have purchased a MikroTik XQ+DA0003 QSFP+ 100G 3M cable. Although, this specific MikroTik cable is not on the list of the compatible products by Nvidia, I can confirm it is working and the link is detected by the Bluefield-2 DPU.

One side note is that, even though I found other copper cables behaving similarly, once this DA cable is attached, I see dmesg reports as having insufficient power for the Bluefields. It means, because of the DA cables (which are active cables), the Bluefields need to draw more power/current. Using any random passive optical cables found in the lab, do not have this power-related issue and work as expected. For now, I am using those instead of the DA copper cable.