0
$\begingroup$

The current enterprise server rack has, in most cases, more than 1 CPU socket per board and a bunch of graphics card together with a data processing unit on a single board in a node. Several such nodes in a rack are closely stitched together for data pathways and tightly integrated into the cluster for a tailored application as deemed fit by the system architect. My question is Are we still in the Von-Neumann architecture era? The reason I raise this question is 2 fold

  1. We no longer have a centralized processing unit, for processing a myriad of data, but rather separate discrete processing units other than CPU for handling specialized function including GPU, DPU and in some cases dedicated accelerators. Architecturally also, in some cases, GPU can have separate access to DRAM and SSD. So more than one processing unit (assuming entire socket as a processor which is not the case though). How come this classifies under Von-Neumann architecture?
  2. If the so called neuromorphic computing ever comes into dominance would the current heterogenous compute elements would still be there or We will be completely shifting towards the neural engine cores? In other words would we still be using the traditional multiply and accumulate based registers or we will start moving to the compute in memory thing?

Are we really facing a architectural change to incorporate NVME drives, most importantly intel pmem, into traditional architecture? Clearly pmem shows architectural choke points for utilizing its full potential

regards, eigenvalue

$\endgroup$
7
  • 2
    $\begingroup$ Why does it matter to you if it is or is not called a "von Neumann" architecture or not? What does this classification gain or lose you? $\endgroup$
    – Bill Barth
    Jun 25 at 15:52
  • $\begingroup$ This matters most in today's era as we have non volatile memory media closer to speeds of DRAM. With such a memory media close to the processor, many things can be accomplished provided we tweak the architecture of access. But again this brings the question I raised Are we still using the Von Neumann architecture or We want to build a new granular cores staggered within a dense matrix of non volatile memory? $\endgroup$
    – eigenvalue
    Jun 26 at 15:10
  • $\begingroup$ Maybe I’m stubborn, but unless you’re in the semiconductor design business, I don’t see how it matters whether we’re in the von Neumann era still or not. If you are, then you get a hand in choosing, and if not then you have to deal with what’s available. I think we’re likely to see some vN and some non-vN architectures going forward. $\endgroup$
    – Bill Barth
    Jun 26 at 15:42
  • $\begingroup$ Hardware architecture is the base of the pyramid of any computing system. The more the software stack takes advantage of the available architecture the more compute power the system has. This could be clearly seen in the case of apples m1 processor. The traditional gap between the hardware architects and the programmers not taking the fullest advantage of the architecture would lead to an under utilized opportunity. This seem to be the case with optane pmem, atleast for me, with the a new component in the memory hierarchy which was not there a decade earlier. $\endgroup$
    – eigenvalue
    Jun 26 at 16:38
  • $\begingroup$ Yes. Most computer parts go under used by most codes. Most coders do not have the resources to take the deep dive into their codes to try to achieve a few percent improvement. Most systems with these fantastic capabilities do not use them at all well. I think this is independent of whether they are vN or not. There are likely to be several more kinds of architectures in our times that require our current algos to be rewritten to suit them. It would be best if the algos matched the language matched the architecture well so that each part was always well used, but we don't live in that world. $\endgroup$
    – Bill Barth
    Jun 29 at 14:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.