SysadminNews

Samsung LPDDR5X-PIM: internal bandwidth is not host I/O

On this page
  1. Follow where the arithmetic happens
  2. The presentation reports several different metrics
  3. Software and workload shape remain part of the result

Samsung’s Hot Chips presentation describes LPDDR5X-PIM, which places arithmetic beside DRAM banks. The advertised 614 GB/s concerns the internal PIM path. It should not be read as an ordinary 614 GB/s transfer link from the memory package to a processor.

Original matrix-vector example: stored rows [1, 2] and [3, 4] multiplied by input [5, 6] give 17 and 39. Bringing input to arithmetic near stored rows can reduce weight movement. Not Samsung’s exact layout or instruction sequence.
Original matrix-vector example: stored rows [1, 2] and [3, 4] multiplied by input [5, 6] give 17 and 39. Bringing input to arithmetic near stored rows can reduce weight movement. Not Samsung’s exact layout or instruction sequence. Chart : PeopleAreGeek. Data source.
View full-size image

Follow where the arithmetic happens

Samsung's technical paper explains a PIM block associated with each bank and a software layer that places data and controls execution. Moving the arithmetic near the stored weights can avoid repeatedly carrying those weights over the external memory interface.

In a small original matrix-vector example, two stored rows, [1, 2] and [3, 4], multiply the input [5, 6]. The outputs are 17 and 39. A PIM-style organization can bring the input to arithmetic near the rows and return the smaller result, rather than moving the entire matrix to a remote arithmetic unit for that operation.

This example explains data placement, not Samsung's exact instruction sequence or supported precision. Inputs, outputs, setup and synchronization still need to move or execute; processing in memory does not eliminate all communication.

The presentation reports several different metrics

The Samsung slides reproduced by ServeTheHome distinguish 76.8 GB/s conventional DRAM bandwidth from roughly 614 GB/s PIM bandwidth at the stated configuration. Their silicon demonstration uses Llama 3.1 8B with integer formats and a short context, not an arbitrary model and workload.

The reported output rate rises from 27.0 to 81.3 tokens/s, about 3.01 times. The separately reported runtime changes from 12.3 to 5.4 seconds, about 2.28 times. Those ratios are not interchangeable, and neither establishes an eight-times application speedup from the bandwidth ratio.

Software and workload shape remain part of the result

The earlier paper is a simulation study. Its matrix-operation results must not be relabeled as the later silicon demonstration. It shows why mapping and synchronization matter: tiling can leave compute blocks idle, and ordering overhead can consume part of the gain.

For a useful application comparison, keep the same task, numerical requirements and output-length policy, then measure the whole execution. Also identify the work left on the host. Compatibility with a memory controller does not mean an unchanged application automatically emits PIM operations.

PeopleAreGeek has not tested this memory. The explanatory diagram is original; the performance figures remain Samsung's reported demonstration, with the public slide source supplied so their conditions can be inspected.

Separate internal PIM bandwidth, model output rate and full runtime; attribute silicon slides and distinguish earlier simulator paper.