GPU Support Matrix (3.4.x)

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

CUDA	--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)	❌	❌	❌