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) | ❌ | ❌ | ❌ |