Suspiciously-discounted second 7900xt delivered, installed, underclocked, and undervolted.
16x -> 8x pcie switch for dual gpus is rough. As one would expect, it halves token throughput.
However I can now run qwen3.8-27b with headroom as opposed to having to close applications to conserve vram in order to run a heavily-quantized qwen3.6-35b-a3b.
I've read "people" say qwen3.8-27b was a big jump, so I guess I'll find out.
Matt Hamilton
npub1gx…74lmh
2026-10-28 01:25:24 UTC
Author Public Key
npub1gx2gqstph00y56zgeshhme8dyt4kk4tgt0dypqdk49qwravk37jse74lmhPublished at
2026-10-28 01:25:24 UTCEvent JSON
{
"id": "84030c28dde9ec0e138cfd85c09316fcfd5e347f6f55edb85225c099839b3d96",
"pubkey": "4194804161bbde4a6848cc2f7de4ed22eb6b55685bda4081b6a940e1f5968fa5",
"created_at": 1793150724,
"kind": 1,
"tags": [
[
"proxy",
"https://noauthority.social/@eriner/117515925905248562",
"web"
],
[
"proxy",
"https://noauthority.social/users/eriner/statuses/117515925905248562",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://noauthority.social/users/eriner/statuses/117515925905248562",
"pink.momostr"
],
[
"-"
]
],
"content": "Suspiciously-discounted second 7900xt delivered, installed, underclocked, and undervolted.\n\n16x -\u003e 8x pcie switch for dual gpus is rough. As one would expect, it halves token throughput.\n\nHowever I can now run qwen3.8-27b with headroom as opposed to having to close applications to conserve vram in order to run a heavily-quantized qwen3.6-35b-a3b.\n\nI've read \"people\" say qwen3.8-27b was a big jump, so I guess I'll find out.",
"sig": "925a473b23f00a5aa541ce1906dc2b3519350761c8d59889b80a38e48e878e97fd2fac481050177e10bdb2b3acc7c0d25f96414b820b5c61405ac73ac96e7a62"
}