Skip to content

GPU Support Matrix (3.4.x)

Ken Raffenetti edited this page Nov 10, 2020 · 6 revisions

This page describes which types of MPI communication are supported with device memory buffers. Note that --enable-ch4-am-only forces all MPI semantics to be implemented by the ch4 device (software emulation), and will work with any ch4 device configuration. The support level key is as follows:

  • ✅ well supported
  • *️⃣ supported, may incur extra copies
  • ❌ not supported, may crash
NVIDIA --with-device=ch4:ofi --with-device=ch4:ucx --enable-ch4-am-only
pt2pt *️⃣ *️⃣
collectives (movement) *️⃣ *️⃣
collectives (ops) *️⃣ *️⃣ *️⃣
one-sided (movement) *️⃣ *️⃣ *️⃣
one-sided (ops) *️⃣ *️⃣ *️⃣
Intel --with-device=ch4:ofi --with-device=ch4:ucx --enable-ch4-am-only
pt2pt *️⃣ *️⃣
collectives (movement) *️⃣ *️⃣
collectives (ops) *️⃣ *️⃣ *️⃣
one-sided (movement) *️⃣ *️⃣ *️⃣
one-sided (ops) *️⃣ *️⃣ *️⃣
AMD --with-device=ch4:ofi --with-device=ch4:ucx --enable-ch4-am-only
pt2pt
collectives (movement)
collectives (ops)
one-sided (movement)
one-sided (ops)