Jump to content
Toggle menu
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

cmp170hx-ffcalc

From CMP Wiki

Reported 8 GB and 10 GB

cmp170hx-ffcalc reads a CMP 170HX's floorsweep fuses and predicts what Kate's GA100VFI hardware unlock could reach on that card: the framebuffer now and at most (64 or 80 GiB on the 8 GB card) and the SM count. The scan itself only reads. The repository also describes a plan for the permanent fuse writes; no member reported using it by 2026-11-15. Author: sayhellotojungle (GitHub). Repository: https://github.com/sayhellotojungle/cmp170hx-ffcalc. @truededinfo posted it in the Discord on 2026-11-14. Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub Source: @truededinfo, 2026-11-14

What it does

The README is in Chinese; this is a translation of the author's own claims, purpose and status only.

  • It is an offline helper for Kate's GA100VFI research (GA100VFI results): plain Python plus a read-only kernel module, independent of cmpunlocker.
  • The module reads 61 fuse registers on each CMP 170HX on a Linux host with the patched NVIDIA driver loaded, then unloads itself. The README says nothing is written to the card and a running model is not disturbed.
  • It prints one row per card: the framebuffer now and at most (the README's example reads 64 to 80 GiB) and the SM count.
  • It also generates a per-card plan for the GA100VFI fuse writes and a check of that plan against the card, but it performs no write. The README warns that the write is physically irreversible and that a wrong target value can permanently ruin a memory partition or the whole card. This wiki does not describe the write.

Reported 8 GB and 10 GB
Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub

Status (README)

  • Data from three real cards: two with 80 GiB potential, one capped at 64 GiB.
  • Whether the fuse controller recomputes the FF chain CRC on a write is unverified.
  • The MIG, ECC and NVDEC unlock is marked untested by the author.
  • The 10 GB card's capacity is an estimate at 4 GiB per FBP.
  • Tested on Debian 12 with nvidia-open 610.x (the cmpunlocker P2P patch version).
  • The register addresses and bit map come from Kate's community research, not from NVIDIA documentation. Licence: MIT.

Reported 8 GB and 10 GB
Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub

What owners report

  • It is the eligibility check people point to. Asked by @hamuki whether running it on his GPU server shows the potential for the unlock, @truededinfo answered yes. When @lenxtroz asked how to check whether his card can be unlocked further, @seriousrising linked it. Source: @hamuki, 2026-11-14
    Source: @truededinfo, 2026-11-14
    Source: @lenxtroz, 2026-11-14
    Source: @seriousrising, 2026-11-14
  • Scans posted on 2026-11-14 and 15. Most cards came out at an 80 GB potential and a few at 64 GB, with SM counts from 102 (one of @mountains1605's cards) to 128: @68x0 all 3 cards at 80 GB with 124, 126 and 128 SMs; @gnasherrr 7 at 80 GB and 1 at 64 GB; @deced_123's only card at 64 GB; one of @sojun80's cards at 64 GB. These are predictions from the fuse readout, not unlock results, and not every post names the tool. The full per-owner list is on Full SM unlock. Source: @deced_123, 2026-11-14
    Source: @mountains1605, 2026-11-14
    Source: @gnasherrr, 2026-11-14
    Source: @68x0, 2026-11-14
    Source: @sojun80, 2026-11-15
  • Before ffcalc: the SM spreads members quoted (2026-11-05 and 11-06). On the clean room server @_orrorin said none of his cards are golden samples, "only 80SM work"; @mbelli answered: "Through 15 the lowest I saw was 96 with most in the 120s", and he had heard 110 was the average. How these counts were read (a fuse read or a count after an unlock) and which variant is not stated. Source: @_orrorin, 2026-11-05 (CMP 170hx Clean Room Discussion server)
    Source: @mbelli, 2026-11-06 (CMP 170hx Clean Room Discussion server)
  • Nobody reported using the write plan. @.shakdiesel asked whether anyone had used the burn plan from ffcalc successfully, worried that something would go wrong because of rows already populated on the card; @kingsizexxxl: "Quite possible" and "Bleeding edge". Earlier that day, in a separate exchange, @djpandaman said the only case where a working card would have trouble is one whose blank FF rows were all rewritten already, and even then it would only stop the unlock from being made persistent. No write from the plan was reported by 2026-11-15. Source: @djpandaman, 2026-11-14
    Source: @.shakdiesel, 2026-11-14
    Source: @.shakdiesel, 2026-11-14
    Source: @kingsizexxxl, 2026-11-14
    Source: @kingsizexxxl, 2026-11-14
  • A second FF row dumper (2026-11-16). @killred found that ffcalc's capture "cannot read the ff_rows" and posted his own ffdump.py as a zip: run it as root with --enable-memory and the card's PCI address (--bdf). He had not written any fuse on his card at that point. The author's change of the same day makes ffcalc read all 512 rows. Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub Source: @killred, 2026-11-16
    Source: @killred, 2026-11-16
    Source: @killred, 2026-11-16
    Source: @killred, 2026-11-16
    Source: @killred, 2026-11-16

Reported 8 GB Is its burn plan safe? (15 to 17 November) In the order posted:

  • @nemz38_is_my_waifu_91412 (11-15): people seem to use only ffcalc to check unlock potential. @lia_oc: "that ffcalc repo still has issues". Source: @nemz38_is_my_waifu_91412, 2026-11-15
    Source: @lia_oc, 2026-11-15
  • @68x0 (11-16): "That repo is broken and doesn't dump FF rows"; "The burn plan it generates is incorrect, it doesn't account for existing data in the FF rows". "Use gpcprobe": it presents the GPCs and TPCs, and "I wouldn't trust that data it spits out". Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-16
    Source: @68x0, 2026-11-16
  • @nemz38_is_my_waifu_91412 (11-16): ffcalc at least seemed consistent: the card that gpcprobe shows with a defective GPC came out at 112 SMs, and the memory and SM figures looked right, "probably not the rest". Source: @nemz38_is_my_waifu_91412, 2026-11-16
    Source: @nemz38_is_my_waifu_91412, 2026-11-16
  • @fallengod9962 (11-16) asked whether to apply a plan his AI assistant had built from ffcalc's readout. @lia_oc: "please do not yolo ffcalc without doing your research first or waiting for other people to post more tools"; @samuel_hayden.: "If you yolo ffcalcs you will brick your card". Source: @fallengod9962, 2026-11-16
    Source: @lia_oc, 2026-11-16
    Source: @samuel_hayden., 2026-11-17

Contested 8 GB Issue #1 (planning over rows already burned) was fixed on 2026-11-16 Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub; the posts do not say whether @68x0's criticism came before or after that fix. More: Voltage-glitch unlock (GA100VFI)#Is cmp170hx-ffcalc safe to plan with?.

Caveats

  • The capacity and SM figures are what the fuses allow, not what a card does after GA100VFI. Whether a GPC the scan shows as cooked is really dead, and how members check it after the unlock, is on 8 GB compute and HBM.
  • @jonathanyjoerrington asked whether the check has to be run on a stock card without cmpunlocker; nobody answered in chat. The README asks for the patched NVIDIA driver and says the tool has no code dependency on cmpunlocker. Source: @jonathanyjoerrington, 2026-11-14
    Source: @jonathanyjoerrington, 2026-11-14
    Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub
  • Before any burn, check ffcalc's GPC and TPC reading against gpcprobe and plan only from a full capture of the card you will burn; members warn that applying an unchecked plan bricks the card (see #What owners report and Unlock without bricking). Source: @68x0, 2026-11-16
    Source: @samuel_hayden., 2026-11-17

On the 10 GB card: the lab's permanent SM burn (2026-11-17)

  • It reads the 10 GB card correctly. On two 10 GB cards (10de:2082) ffcalc decoded the floorsweep from the raw fuse rows and reproduced the floorsweep masks the lab had read directly from the hardware (GPC disable and the eight TPC masks on one card, GPC, TPC and PES on the other). Proven 10 GB
    Source: @darkmaniac7, 2026-11-17 (lab notes)
  • Issue #1 is fixed. The first plans did not account for FF rows already burned on the card. On 2026-11-16 the author made ffcalc read all 512 rows and refuse to plan from an incomplete capture. Plan only from a full capture of the card you will burn. Reported
    Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub
  • A burner was added on 2026-11-17. burntool programs blank main-array rows one at a time and verifies each; its procedure document reports 17 FF rows on one card that survived a cold boot. That example uses the glitch window of the 8 GB route. Reported
    Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub
  • What the lab fixed before using it on 2082 cards (commit 4e089b54fc3c): the capture accessor calls int(addr, 0) on integer offsets, so the first status call fails with a TypeError; the burner writes in plan order and sends a SENSE at the start of every run, so the GPC-record-last order has to be enforced by the plan or the caller; after a controller fault it still read other registers and restored FUSEADDR, where it should stop all access; and it does not tie the live device to the card the plan was made for (only the target rows are compared). Do not use --force. With those fixes the lab burned seven compute-only records per card (GPC, TPC and PES; no memory, PCIe or RECONFIG records) and both cards came up at 122 and 108 SMs from a cold boot. Drop the PCIe records the planner can add: which bit is the Gen3 disable is still unresolved (resolved on 2026-11-19: bit 9, below). Steps: SW-IFF unlock on the 10 GB card#Making it permanent: the FF burn. Proven 10 GB
    Source: @darkmaniac7, 2026-11-17 (lab notes)
    Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub

On the 8 GB card: the lab's burns after the glitch (2026-11-19)

  • The same patched burn tool on two glitched 8 GB cards. With the fixes above the lab burned 14 and 16 records (compute, memory and PCIe) on two 8 GB cards, one row at a time with every row read back. Both cards came up as planned after a cold boot, and a fresh plan from the new read was a NO-OP. The plans came from the lab's own planner and were simulated with ffcalc's simulate (0 errors) before each glitch. Proven 8 GB
    Source: @darkmaniac7, 2026-11-19 (lab notes)
  • Send CMD_SENSE first. In the glitch window EFP_STATUS (0x820078) often reads 0, and the burner wants EFP = 1 before its own sense. One CMD_SENSE sets EFP to 1 and leaves DSOV at 0; then the window check passes. Proven 8 GB
    Source: @darkmaniac7, 2026-11-19 (lab notes)
  • The Gen3 bit is bit 9. OPT_DISABLE_GEN3_SPEED is chain 2 block 39 bit 9; a record clearing bit 1 left it at 1 after a cold boot. A planner should clear bit 9 alone and keep bit 1. Proven 8 GB
    Source: @darkmaniac7, 2026-11-19 (lab notes)
  • The in-session read-back is not a verification. CMD_SENSE does not re-apply the FF records to the option registers; the GPC, FBP and PCIe registers changed only after a cold boot. Proven 8 GB
    Source: @darkmaniac7, 2026-11-19 (lab notes)
  • Setup, flow and the records: Voltage-glitch unlock: lab method on a Pico 2 W, Voltage-glitch unlock: lab results and fuse findings.

See also