Dgx sparks are not slow, that is fake news

utxo the webmaster 🧑💻
npub1utx00neqgqln72j22kej3ux7803c2k986henvvha4thuwfkper4s7r50e8
hex
00000459b0acbb70bc4fb761df2c76f992ce43fe8b55f21480674927d669961dnevent
nevent1qqsqqqqytxc2ewmsh38mwcwl93m0nykwg0lgk40jzjqxwjf86e5ev8gprpmhxue69uhhyetvv9ujuem4d36kwatvw5hx6mm9qgsw9n8heusyq0el9f99tveg7r0rhcu9tznatuekxt764m78ymqu36cgtvdzzKind-1 (TextNote)
Dgx sparks are not slow, that is fake news
Especially when you factor in prefill, which is half the speed in big context
New software optimizations drop almost every week that push the speeds up, little things like math kernels etc
Currently getting single stream 60 Tok/sec decode and 11,000 Tok/sec pp running deepseek on two sparks
People just looking at memory bandwidth and drawing conclusions don't know what they're talking about
Raw JSON
{
"kind": 1,
"id": "00000459b0acbb70bc4fb761df2c76f992ce43fe8b55f21480674927d669961d",
"pubkey": "e2ccf7cf20403f3f2a4a55b328f0de3be38558a7d5f33632fdaaefc726c1c8eb",
"created_at": 1787413605,
"tags": [
[
"client",
"Wisp"
],
[
"nonce",
"88502",
"16"
]
],
"content": "Dgx sparks are not slow, that is fake news\n\nEspecially when you factor in prefill, which is half the speed in big context \n\nNew software optimizations drop almost every week that push the speeds up, little things like math kernels etc\n\nCurrently getting single stream 60 Tok/sec decode and 11,000 Tok/sec pp running deepseek on two sparks \n\nPeople just looking at memory bandwidth and drawing conclusions don't know what they're talking about",
"sig": "6a4ec4bf11d79b09f37855c564aedcba13c913346933b107625dad7f980518ef3bf35e159252d93fd64953278e29a1c033524c055fde64cfe8648aebc013be44"
}