The AMD MI60 allegedly has an infinity fabric link. It is rumored to connect each of 4 MI60's in a hive to two others, in a ring topology. The MI60 is not compatible with the infinity fabric link used for the MI100. It might or might not be compatible with the one used for the MI50.
Anyone may submit answers and bet on them. For an answer to resolve "Yes", you must convince me that a part number is valid and compatible with the MI60, for example with a link or a screenshot from a reputable source. The part must be able to connect with 4 MI60's, both physically and electronically. There may be multiple valid part numbers, so there may be multiple correct answers.
1,000🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ9,720 | |
| 2 | Ṁ42 | |
| 3 | Ṁ2 |
First of all, apologies for not resolving on January 1 like I planned to. I had a lot on my mind this past week.
I am happy to say that I conclusively know the part number of the infinity fabric link for AMD's MI60. It is 199-999764, also known as the "4P Adjacent" variety. In October, someone sold 2 units of these, and I bought one. I tested it with my MI60's and rocm-bandwidth-test showed that it increased communication bandwidth. For me it was not quite as much bandwidth as I expected from the specifications, but who knows if that's something wrong with my setup or a misunderstanding of the specifications. But I don't regret the purchase. The part looks very cool and I may buy more if I see a good deal.
I don't know how many people will see this, but if anyone needs tips on getting MI60 or similar cards to work for pytorch or llama.cpp, let me know and I'll share more about how I got them to work.
If the correct part number turns out to be slightly different than a submitted answer, for example if it is a substring of the submitted number, or if it has symbols like hyphens, underscores, or spaces added or omitted, or if it has different capitalization, but it is otherwise correct and trustworthy, I will resolve that submitted answer as yes.