By: Groo (charlie.delete@this.semiaccurate.com),
Room: Moderated Discussions
Ireland (boh.delete@this.outlook.ie) on March 13, 2017 9:12 am wrote:
> Thanks, I read the piece. But it left me wondering about it all, on another level. There might
> be a bit of flesh that ought to be added to the bones of that interpretation. I.e. The one that
> says, the existing fabric was there since Dec Alpha, and that's why AMD needed a new fabric.
>
> On the other hand, what's really going on, on top of server motherboards at AMD?
>
> Really, what is the big play here, in terms of the data center, cluster, HPC or whatever?
>
> Well, they're attempting to do something again, before Intel does. Intel is great at
> offering a multi-socket solution, with socket inter-connection between two or more
> x86 multi-cores. On the other hand, AMD is giving you the option to change that whole
> piece of hardware into a multi-socket ARM server in the future, if needed be.
>
> I mean, they've got a 'blank' sheet of paper, as far as taking the design from ARM goes, and baking
> that into some new kind of hardware interpretation. That 'infinity fabric' has to work inside
> this ambidextrous thing - where if the workload demanded it - I presume that those ARM chips,
> that fit into the same socket - could enable a lot more compute cores, to fit on the same socket.
> And you double or quadruple that by communicated between sockets on the same board.
>
> I.e. The kind of traffic, the kind of workload that's going to be done on the same board, using
> the same sockets, is nothing at all - like the kind of workload, that one would attempt to
> do, using that hardware and x86 cores. Unless I am wrong. I mean, an operator that wants to
> build a cluster of ARM-based AMD systems, is looking to optimize different things, from the
> one who wants to use the same cluster, to carry x86-based computational workload, no?
Two things stand out. First is the old Hypertranport was both hardware and protocol. THe new IF is physical layer agnostic which should provide a lot more flexibility. It works on-die, between MCM chips, and between sockets. I don't know if it also can work for inter-system or inter-rack comms, I will ask if I get in front of the right people.
The other thing is that IF is far more granular than HT ever was in that it isn't a chips to chip protocol or even a core to core protocol. From what I gathered from my chats with AMD personnel, there are multiple IF endpoints on every die, with multiple being a large number, not single digits. The idea is to both transport data between blocks and to have a separate control fabric as well. How this is exactly laid out and controlled, much less exact capabilities, hasn't been revealed yet.
It was strongly hinted at that a block can target another block directly for a transfer. My educated guess is for HSA type workloads and pointer passing, a CPU core can pass data directly to a shader on an APU that needs it for the nest instruction. I may be very wrong on this, but I suspect this is the long term goal of the system.
I am trying to find out more but getting anything more than bullet points is tough at the moment.
-Charlie
> Thanks, I read the piece. But it left me wondering about it all, on another level. There might
> be a bit of flesh that ought to be added to the bones of that interpretation. I.e. The one that
> says, the existing fabric was there since Dec Alpha, and that's why AMD needed a new fabric.
>
> On the other hand, what's really going on, on top of server motherboards at AMD?
>
> Really, what is the big play here, in terms of the data center, cluster, HPC or whatever?
>
> Well, they're attempting to do something again, before Intel does. Intel is great at
> offering a multi-socket solution, with socket inter-connection between two or more
> x86 multi-cores. On the other hand, AMD is giving you the option to change that whole
> piece of hardware into a multi-socket ARM server in the future, if needed be.
>
> I mean, they've got a 'blank' sheet of paper, as far as taking the design from ARM goes, and baking
> that into some new kind of hardware interpretation. That 'infinity fabric' has to work inside
> this ambidextrous thing - where if the workload demanded it - I presume that those ARM chips,
> that fit into the same socket - could enable a lot more compute cores, to fit on the same socket.
> And you double or quadruple that by communicated between sockets on the same board.
>
> I.e. The kind of traffic, the kind of workload that's going to be done on the same board, using
> the same sockets, is nothing at all - like the kind of workload, that one would attempt to
> do, using that hardware and x86 cores. Unless I am wrong. I mean, an operator that wants to
> build a cluster of ARM-based AMD systems, is looking to optimize different things, from the
> one who wants to use the same cluster, to carry x86-based computational workload, no?
Two things stand out. First is the old Hypertranport was both hardware and protocol. THe new IF is physical layer agnostic which should provide a lot more flexibility. It works on-die, between MCM chips, and between sockets. I don't know if it also can work for inter-system or inter-rack comms, I will ask if I get in front of the right people.
The other thing is that IF is far more granular than HT ever was in that it isn't a chips to chip protocol or even a core to core protocol. From what I gathered from my chats with AMD personnel, there are multiple IF endpoints on every die, with multiple being a large number, not single digits. The idea is to both transport data between blocks and to have a separate control fabric as well. How this is exactly laid out and controlled, much less exact capabilities, hasn't been revealed yet.
It was strongly hinted at that a block can target another block directly for a transfer. My educated guess is for HSA type workloads and pointer passing, a CPU core can pass data directly to a shader on an APU that needs it for the nest instruction. I may be very wrong on this, but I suspect this is the long term goal of the system.
I am trying to find out more but getting anything more than bullet points is tough at the moment.
-Charlie
