5 ms·

By moving closer you don't gain bandwidth, you reduce latency.

• dom0 9y ago
  ... and not by that much ... ... which is why projects targeting this are actually about scheduling tasks across nodes, not moving processing actually closer to the memory in the "let's put the CPU in the memory" sense.
  • microcolonel 9y ago
    Well, for many types of compute, transaction latency is everything. Every time you cut the difference in half, you cut the latency in half and get twice as many dependent responses. Shrinking the process produces huge improvements. HBM exists for a good reason. You wouldn't stack dies (and hence reduce available thermal dissipation) if it didn't make a big difference.
    • dom0 9y ago
      It's in the name - HBM increases bandwidth. It does reduce latency a bit (since the electrical path is shorter), but the bigger and more fundamental latencies (row and columns) have not changed in many many years and are unlikely to. (It'd be possible to make them lower by proportionally higher power use and lower density, a trade-off which no one seems to be willing to make).
• adwn 9y ago
  No, you do actually gain bandwidth. By moving the memory into the same package as the processor, you can add more data lines: adding 1000 more pins to a CPU is practically infeasible, while 1000 more signals between dies in the same package is relatively simple with the right technology. In addition, due to the shorter distance and resulting benefits to signal integrity and power consumption, internal data lines can be clocked faster than those between chips.
• ptman 9y ago
  Typically, improving latency also improves bandwidth, whereas typically, improving bandwidth hurts latency.