Variant comparison
From CMP Wiki
More actions
Reported 8 GB and 10 GB
The 8 GB card and the 10 GB card side by side. Prices are reported figures from the Discord on the dates given, not quotes. How to tell which card you have: Which card do I have.
| Property | 8 GB card | 10 GB card | Source |
|---|---|---|---|
| PCI ID | 10de:20c2 (subsystem 1585) | 10de:2082 (subsystem 1557) | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 |
| HBM | SK hynix HBM2e, 16 of 24 FBPA, 4096-bit | Samsung HBM2e, 20 of 24 FBPA, 5120-bit | Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20 Source: @darkmaniac7, 2026-10-30 (site owner, from his own cards); earlier posts that conflict on HBM2 versus HBM2e: Which card do I have#Memory vendor |
| Memory, stock | 8 GB | 10 GB | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 |
| Memory, unlocked | 64 GB (65536 MiB) Proven | 40 GB (40960 MiB) Proven; 48 GB Reported if the sixth stack is healthy, about 1 card in 5 (6 of 30 cards seen) | Source: @zorg33, 2026-10-07 Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 Source: @newman_08462, 2026-10-09 Source: @djpandaman, 2026-10-17 Source: @programmerboi, 2026-10-25 |
| Memory beyond that | 80 or 96 GB rumoured, nothing shown | 80 or 96 GB not achieved Contested: the evidence points to aliasing, not real memory | Source: @djpandaman, 2026-10-12 Source: @colo2442, 2026-10-14 |
| HBM clock, stock | 1,458 MHz (NDIV 54, VBIOS 92.00.67.00.01) or 1,728 MHz (NDIV 64, VBIOS 92.00.6D.00.0A) | 1,215 MHz | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| Bandwidth, theoretical | 1.49 TB/s at 1,458 MHz; 1.77 TB/s at 1,728 MHz; about 1.94 TB/s at NDIV 70 | 1.56 TB/s; about 1.87 TB/s with all six stacks | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18; higher values computed as bus width x 2 x clock |
| Bandwidth, reported | about 1,493 GB/s on an unlocked card (tool readout) | 1,741 GB/s on an unlocked card at stock clocks | Source: @cuddylac997, 2026-10-07 Source: @kio063218, 2026-10-16 |
| HBM overclock | NDIV 68-70 typical, 74 fails | not reported | Source: @alanoo, 2026-10-14 |
| SMs | 70 stock, 74 unlocked | 70 stock, 74 public; 112-128 after the RIR burn | Source: @lia_oc, 2026-10-22 Source: @wakeupmbrc1, 2026-10-27 |
| RIR fuse records | 15 of 16 used, last one partly burned: no burn path | 16 of 16 free | Source: @wakeupmbrc1, 2026-10-13 Source: @djpandaman, 2026-10-17 |
| PCIe link, stock | Gen1 x4 | Gen1 x4 | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 |
| PCIe link, unlocked | Gen2; Gen3 blocked by fuse | Gen2; Gen3 after the RIR burn; Gen4 reported, not public | Source: @djpandaman, 2026-10-09 Source: @wakeupmbrc1, 2026-10-25 |
| PCIe width | x16 needs the PCIe capacitor mod | same | Source: @lxtend, 2026-10-17 |
| NVLink | no path | not public; needs a board-level mod | Source: @lia_oc, 2026-10-22 |
| ECC | yes (cmpunlocker v0.5) | yes | Source: @reala7096, 2026-10-22 |
| Windows | no Windows unlock reported; cmpunlocker is Linux only (an A800 device-ID change is an untested idea) | yes, with the A100 40G / A800 40G driver | Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 Source: @wakeupmbrc1, 2026-10-11 |
| Power limit (VBIOS) | 250 W (92.00.67.00.01) or 300 W (92.00.6D.00.0A) | 250 W; 300 W reported with a no-flash VBIOS loader | Source: @zagitta, 2026-10-12 Source: @lxtend, 2026-10-20 |
| Power, in use | most owners cap at 140-200 W | power-bound above 80 SMs | Source: @djpandaman, 2026-10-13 Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| Power connector | one EPS12V 8-pin | same | Source: @civiloid, 2026-10-15 |
| Cooling | passive heatsink with IHS; forced airflow required | same; memory runs a few degrees warmer (41 C vs 37 C at about 33 W idle) | Source: @ghostinthedatacenter_87281, 2026-10-13 Source: @spectakkular, 2026-10-27 |
| Price, reported (USD per card unless noted) | 100-300 before the unlock; 1,900 (eBay, 2026-10-01); 3,000-3,500 (2026-10-13); 3,300-3,600 delivered from China (2026-10-25) | about 200 before the unlock news; 20,000 RMB in China, about the same as an 8 GB card (2026-10-13); cheaper than the 8 GB card on eBay (2026-10-13); 1,200 for a "for parts" card that worked (2026-10-25) | Source: @cuddylac997, 2026-10-07 Source: @djpandaman, 2026-10-12 Source: @lenxtroz, 2026-10-13 Source: @wakeupmbrc1, 2026-10-13 Source: @froggao, 2026-10-13 Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 Source: @kellertreppe123, 2026-10-25 Source: @programmerboi, 2026-10-25 |
| Good for | LLM inference: memory matters most | training, image and video generation, diffusion; compute-bound work | Source: @wakeupmbrc1, 2026-10-13 |
Which one to choose
Reported
- If you run LLMs, memory per card decides what fits. The 8 GB card gives 64 GB; the 10 GB card gives 40 GB (48 GB if lucky) plus 128 SMs and Gen3. Source: @68x0, 2026-10-26 PCIe only matters sometimes; VRAM matters basically all of the time. Source: @sojun80, 2026-10-20 In the same thread @shudza answered that "128 SM matters too", and @programmerboi that VRAM "only matters sometimes": his pretraining work is compute bound. Source: @shudza, 2026-10-20
Source: @programmerboi, 2026-10-20
Source: @programmerboi, 2026-10-20 - If you do not need tensor parallel or training, the 8 GB card serves better; think in memory per watt. Source: @djpandaman, 2026-10-07 The 10 GB card gets more compute and interconnect bandwidth today, the 8 GB card more total memory. Source: @_arcanum_, 2026-10-07
- If you already own 8 GB cards, trading them for 10 GB cards is not worth it for most users: "realistically for most of us, the bottleneck is still VRAM" (@lia_oc); @reala7096: "I wouldn't trade my 8GB cards for 10GB ones". Source: @lia_oc, 2026-10-22
Source: @lia_oc, 2026-10-22
Source: @reala7096, 2026-10-22 - Mixing both in one rig works with a per-card unlock profile, but a multi-card setup runs at the pace of the slowest card and link. Source: @djpandaman, 2026-10-07