View table: Confirmation

Four cards of an 8-card box (Gen2 x16, P2P verified on all pairs), served beside another 4-card instance; 1,024-token answers, his own bench. Smoke test passed (arithmetic, tool call, image). Speed matches the published release closely. Found the long-context copy defect listed on the page (85% of named paths wrong on 75k to 92k-token turns). ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @darkmaniac7, 2026-10-28 (lab notes)

Four cards on PCIe Gen2 x4 links: PP4 did not work well, TP4 with P2P worked and never felt slow. Quality a bit above GLM 5.3 Flash but less consistent; later he preferred the Morrowmake build for prefill and quality. ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @mountains1605, 2026-10-22
?'"`UNIQ--templatestyles-00000009-QINU`"'?Source: @mountains1605, 2026-10-23
?'"`UNIQ--templatestyles-0000000A-QINU`"'?Source: @mountains1605, 2026-10-27

Ran it for two days of agentic work on four 8 GB cards at PCIe Gen2 x4: better quality than the earlier GNmac recipe (which struggled with prefill and quality at times), handles complex work better, rated above GLM 5.3 Flash; staying on it. ?'"`UNIQ--templatestyles-00000006-QINU`"'?Source: @mountains1605, 2026-10-26
?'"`UNIQ--templatestyles-00000008-QINU`"'?Source: @mountains1605, 2026-10-27

Run on the same board as his own recipe: dual-socket Broadwell, 140 W per card, peer access off, so not the published 180 W / P2P conditions. Its prefill beat his own stack at every length. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: repo github.com/CurtisAccelerate/DeepSeekv41-Flash-1017-EXL3-CMP170HX by CurtisAccelerate on GitHub, posted by @anon33anon, 2026-10-20

Recipe 1.0.0 on four cards of an 8-card box with pqhaz's heretic patch written into the official shards, a miner co-resident (--gpu-memory-utilization 0.92), Engram on SSD, his own bench script. Tool calls work. Published speeds not reached. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @darkmaniac7, 2026-10-29 (lab notes)

Eight cards at Gen2 x16 with P2P, Engram from disk (125 GB host RAM at the time), 1,024-token answers. Prefill, aggregate throughput, 1M context and image input match the author; single-stream decode is about half the author's 117 tok/s. The first run ended with the NVMe dropping off (see the page's warning). ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @darkmaniac7, 2026-10-27 (lab notes)

Confirms one fan per pair of cards (first reported by @cibernox ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @cibernox, 2026-10-07): two Supermicro FAN-0074L4 fans, one per pair of cards, taped with VHB straight onto the card ends with no duct, running inference with 1M-token contexts. ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @c133, 2026-10-09

Same fans as the push-pull report by @wakeupmbrc1 ?'"`UNIQ--templatestyles-00000006-QINU`"'?Source: @wakeupmbrc1, 2026-10-16, but push only; @wakeupmbrc1 replied that adding pull fans lets it run quieter and cooler. ?'"`UNIQ--templatestyles-00000008-QINU`"'?Source: @alanoo, 2026-10-26
?'"`UNIQ--templatestyles-00000009-QINU`"'?Source: @wakeupmbrc1, 2026-10-26

Confirms the plenum and that S12038-4K fans are enough on it for inference (first reported by @reala7096 ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @reala7096, 2026-10-08). He uses his own horizontal remix of the plenum. ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @mountains1605, 2026-10-08

Confirms @port00 (RazK)'s integrated fan housing (Thingiverse 7409424 by PluraliT, updated on Cults3D), first reported in use by @68x0 at 60 C and 200 W ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @68x0, 2026-10-15. @moliminous used the Cults3D version unmodified on a 5-GPU PCIe switch board. ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @moliminous, 2026-10-10

mrpandagod and lxwx established the 10 GB SM unlock; GhostInTheDatacenter reproduced it ("Finally!!!" with a screenshot on 2026-10-19, a second screenshot on 2026-10-26), as did orrorin and Waypoint. The route through the RIR burn is 68x0's line ("You can unlock more SMs with the RIR burn"), which GhostInTheDatacenter confirmed ("Exactly"). ?'"`UNIQ--templatestyles-00000008-QINU`"'?Source: @ghostinthedatacenter_87281, 2026-10-19
?'"`UNIQ--templatestyles-0000000A-QINU`"'?Source: @ghostinthedatacenter_87281, 2026-10-26
?'"`UNIQ--templatestyles-0000000B-QINU`"'?Source: @68x0, 2026-10-26
?'"`UNIQ--templatestyles-0000000C-QINU`"'?Source: @ghostinthedatacenter_87281, 2026-10-26

Runs his cards in the same server; good BIOS and out-of-band management, cards stay cool with an aggressive fan tune; each GPU pair on its own switch. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @cmdrwaka324, 2026-10-26
?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @cmdrwaka324, 2026-10-26

Run on two x16-modded 8 GB cards (Threadripper 3970X host) as the baseline for his tensor-parallel extension of this repo, which was 28% faster. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: repo github.com/alexvnesta/glm53-flash-cmp170hx-tp by alexvnesta on GitHub, posted by @nestaa51, 2026-10-12

Four 8 GB cards on PCIe Gen2 x4 links (no x16 mod), PP4 layout. These are the field numbers for PP4 on x4 links, which the recipe itself has not measured. ?'"`UNIQ--templatestyles-00000006-QINU`"'?Source: @mountains1605, 2026-10-04
?'"`UNIQ--templatestyles-00000008-QINU`"'?Source: @mountains1605, 2026-10-08

PP4 reproduces the published numbers closely; TP4 lands at about 0.75x without x16 links and P2P, because the all-reduce runs through host memory. Draft acceptance matched the published regime. Vision and video worked. ?'"`UNIQ--templatestyles-00000006-QINU`"'?Source: @darkmaniac7, 2026-10-19 (lab notes)
?'"`UNIQ--templatestyles-00000008-QINU`"'?Source: @darkmaniac7, 2026-10-28 (lab notes)

@globalapathy used @matstatman's probe tools the day ?'"`UNIQ--nowiki-00000006-QINU`"'? was released and said they saved him from buying bad cards; only one card of the batch was good. He did not say which variant or which readout showed the problem. ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @globalapathy, 2026-10-20

lxwx first reported Gen3 working on a 10 GB card (2026-10-12); mrpandagod reports independently doing both 3.0 and 4.0 on 10 GB cards through the RIR route, and evanrodgers validated Gen3 x4 at boot on two cards with a clean 30-minute stress run. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @djpandaman, 2026-10-11

Changes on top of the repo: prefill batch 2,048 to 4,096, MTP draft tokens 1 to 2, PLE row cache 512 to 8,192 MiB, PLE AIO queue depth 256 to 512; NDIV 67, +250 MHz core offset, 200 W. He warns measurement differences may explain part of the gap. Used for one to two weeks of reverse-engineering work without issues. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @dottoreargus, 2026-10-07

Release 0.2.0 on Gen2 x16 cards with P2P, 1,024-token answers at temperature 0. Vision works. Four instances share one 48 GiB table through PLE_SHARED_DIR on tmpfs (77 GiB host RAM for all four); each keeps 90 to 100% of its lone speed. Custom all-reduce off cost 17 to 25% single-stream, so it stays on. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @darkmaniac7, 2026-10-28 (lab notes)

Three unlocked 8 GB cards, pulled the author's branch. Later runs used an uncensored model variant to get MTP working (MTP 1 token) and moved to 2 cards with the persistent NVMe KV cache, freeing the third card; prefill only slightly lower on 2 cards. ?'"`UNIQ--templatestyles-00000008-QINU`"'?Source: @68x0, 2026-10-08
?'"`UNIQ--templatestyles-0000000A-QINU`"'?Source: @68x0, 2026-10-08
?'"`UNIQ--templatestyles-0000000B-QINU`"'?Source: @68x0, 2026-10-09
?'"`UNIQ--templatestyles-0000000C-QINU`"'?Source: @68x0, 2026-10-11

orrorin proposed the record pair and Petri Krohn completed the plan; lxwx then burned 0x00150095 on his own 10 GB card and confirmed it clears the gate ("0x00150095 is right ... it does unlock the disable_sw_override fuse"). Reproduced again later by GhostInTheDatacenter, orrorin and Lycorascence with 170hx-rir-tool. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @lxtend, 2026-10-11

Has the same server (4028GR-TRT with the X10DRG-O-PCIE board) and confirms two PLX chips with five slots each; advises keeping cards behind the same switch in consecutive slots. Confirms the topology, not a performance result. ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @nathang6866, 2026-10-19
?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @nathang6866, 2026-10-20

Confirms that a card without forced airflow overheats and drops out even at idle (first reported by @analyticeth, off the bus at 95 C ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @analyticeth, 2026-10-07). ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @programmerboi, 2026-10-16

Confirms that a too-deep undervolt causes errors that a smaller offset removes (first reported by @ichundes, Xid 48/172/13 gone after lowering the offset by 50 ?'"`UNIQ--templatestyles-00000005-QINU`"'?Source: @ichundes, 2026-10-07). Both had passed 170tune's stability tests. ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @cibernox, 2026-10-07

Confirms that the Bykski N-TESLA-A100-40G-SR block fits the CMP 170HX (first reported installed by @lia_oc ?'"`UNIQ--templatestyles-00000007-QINU`"'?Source: @lia_oc, 2026-10-10): "i have 8 of the new bykski 40G waterblocks" ?'"`UNIQ--templatestyles-00000009-QINU`"'?Source: @djpandaman, 2026-10-08, "yes. they're the ideal block" ?'"`UNIQ--templatestyles-0000000A-QINU`"'?Source: @djpandaman, 2026-10-09, "my first one has 8 170hx cards in water blocks" ?'"`UNIQ--templatestyles-0000000B-QINU`"'?Source: @djpandaman, 2026-10-09.