PCIe Gen4 on the 10 GB card
More actions
| Reported 10 GB | |
| Driver | nvidia-open 615.71.09 (cmpunlocker, 10gb profile) |
|---|---|
| Unlock build | 170hx_tips Gen3 route (Variant B) + lab Gen4 patch + runtime PHY init |
| Steps | PCIe Gen4 on the 10 GB card |
| Source | lab notes |
| Confirmations | None yet. |
Warning 10 GB card only, after the RIR fuse burn and from a working Gen3 link. One card at a time, one retrain attempt per boot. A failed 16 GT/s transition can drop the card off the bus, and recovery is a secondary bus reset, a reboot or a power cycle (#Recovery). Nothing on this page is burned: a power cycle returns the card to its boot link. Do not burn a Gen3 or Gen4 FF record (#Risks).
On 2026-11-17 the lab trained PCIe Gen4 on both of its gate-burned 10 GB cards (10de:2082, 122 and 108 SMs) on bare metal: 16 GT/s x4 at the card and at the root port, equalization complete, 0 AER and 0 Xid, and pinned copies at 6.37 GB/s host to device and 6.76 GB/s device to host, 2.0x the Gen3 x4 figures. It is volatile and needs no fuse write beyond the RIR gate. The route is the Gen3 driver build plus an additive Gen4 patch that clears bit 25 of PCIE_MAGIC_BITS_D at probe, then a runtime PHY init and one config-space retrain. The other Gen4 bits of that register stayed set: A100 values were not needed. Width stays x4 without the PCIe x16 capacitor mod. Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Results
| Card A (122 SM) | Card B (108 SM) | |
|---|---|---|
| Upstream port | CPU root port, up to 32 GT/s x8, no switch | its own CPU root port, same type |
| Link before (Gen3 route) | 8 GT/s x4 at both ends | 8 GT/s x4 at both ends |
LnkCap2 after cmp_fuse_gen4=0 |
2.5-16 GT/s (was 2.5-8 GT/s) | 2.5-16 GT/s (was 2.5-8 GT/s) |
LnkSta after the PHY init and one retrain |
16 GT/s x4 at both ends, equalization complete | 16 GT/s x4 at both ends, equalization complete |
| AER and Xid after the retrain | 0 (correctable 0), 0 | 0 (correctable 0), 0 |
| Pinned 256 MiB copies at Gen4 x4, host to device / device to host | 6.37 / 6.76 GB/s | 6.37 / 6.76 GB/s |
| Same copies at Gen3 x4 (two-card soak average) | 3.24 / 3.38 GB/s | 3.24 / 3.38 GB/s |
| 30-minute Gen4 soak, card alone | PASS: 120 of 120 samples at 16 GT/s x4 at both ends, 0 AER (correctable 0), 0 Xid, gpu-burn 0 errors, 6.37 / 6.76 GB/s average under concurrent compute (minimum 6.27), still 16 GT/s at the end, at most 56 C | not run alone |
| 30-minute Gen4 soak, both cards at once | PASS: 120 of 120 samples at Gen4 x4 on each card, 0 AER (correctable 0), 0 Xid, gpu-burn PASS, 6.31 / 6.76 GB/s per card under load (minimum 6.30), still 16 GT/s at the end, 54 to 56 C | |
| SMs and memory with Gen4 | 122 SM, 40 GiB | 108 SM, 40 GiB |
Host: a desktop board, UEFI, Ubuntu 24.04, kernel 6.8.0-149, Secure Boot off, cmpunlocker on nvidia-open 615.71.09 with the 10gb profile. Card A went first; card B trained while card A stayed at 16 GT/s, so both ran Gen4 x4 at the same time. Neither card needed a second attempt. Gen4 x4 raw is 7.88 GB/s per direction, so the copies run at about 81 to 86 % of raw, as Gen3 did on the same cards. Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Code
The lab's four driver patches (the Gen3 stack and the Gen4 patch, described below) and its scripts (set_pcie_speed.c, cmp-gen3.sh, cmp_gen4_e1.py, cmp-gen4.sh) are on GitHub: darkmaniac7/cmp170hx-10gb-pcie-gen4. Its README gives the exact base (cmpunlocker commit 2c7e2ec on nvidia-open 615.71.09), the patch order, the build and install commands, the module options, the steps for every boot and the pass gates. The four patches apply to a fresh 615.71.09 tree with no fuzz and give the same source tree the tested modules were built from. 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: repo github.com/darkmaniac7/cmp170hx-10gb-pcie-gen4 by darkmaniac7 on GitHub
Requirements
- A 10 GB card (
10de:2082) with the RIR fuse burn done and active:DISABLE_SW_OVERRIDE_STATUS(0x820084) reads 0. - Gen3 working first on the same build: the Variant B route on nvidia-open 615.71.09 with cmpunlocker's 10gb profile. Gen4 starts from a trained 8 GT/s link.
- A Gen4-capable port at the other end. The lab's cards sit directly on CPU root ports. Behind a PCIe switch, both sides of the switch must do 16 GT/s.
- Bare metal. The retrain writes the root port's config space, which a passthrough guest cannot do (the same limit as Gen3).
- A pinned-memory copy test for host to device and device to host.
LnkStaalone is not the pass signal. - Console or power access: a failed attempt can need a power cycle.
The driver build
The lab built one module set for kernel 6.8.0-149 from these parts, in this order:
- cmpunlocker's core patches for 615.71.09 (10gb profile).
- The Gen3 route as "Variant B": cmpunlocker's Gen2 patch kept for its booter code in
kernel_gsp.c, its Gen2 block inkernel_gsp_tu102.creplaced by the 170hx_tips post-GSP Gen3 block, plus the 170hx_tips probe-time fuse override (cmp_fuse_gen3). Details: PCIe Gen3 on the 10 GB card. - The in-memory VBIOS override patch from the vix loader, used for the V/F offset (Undervolting and power limits). Gen4 does not use it and the two patches touch different files, but it was in the tested build; Gen4 without it was not tested.
- The Gen4 patch. It is additive (it removes no line of the stack above) and touches two files:
nv-pci.c: two module parameters.cmp_fuse_gen4=0clears bit 25 (GEN4_SPEED_DISABLED) ofPCIE_MAGIC_BITS_D(0x820520) at probe, by read-modify-write, under the same software override that clears the Gen3 fuses.cmp_fuse_magicd_clrcan also clear bits 19, 22, 26 and 28 (mask0x14480000; together with bit 25 that gives the A100 value0x00200000). The lab did not need it.kernel_gsp_tu102.c: an optional boot-time PHY init after the Gen3 block, behind the registry keyCmpGen4=1, which runs only on a card whose bit 25 is already clear. The lab did not use it; the PHY init ran at runtime instead.- Left out on purpose: the
0x881C0(SYSTEM_MAX_SPEED) Booter write from CMPGlitch's 8 GB Gen4 driver patch, which wedged GSP boot in that author's tests (FECS PRI timeout, RmInitAdapter failure).
With cmp_fuse_gen4, cmp_fuse_magicd_clr and CmpGen4 unset, the build behaves exactly as the Gen3 build. The patches and the scripts are posted on GitHub; see #Code. Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: repo github.com/lixiangwuxian/170hx_tips by lixiangwuxian on GitHub, posted by @lxtend, 2026-10-19
Source: @._.patrik, 2026-11-16
Source: repo github.com/Panchovix/nvidia-linux-vbios-vix-loader by Panchovix on GitHub
Module options
options nvidia NVreg_RegistryDwords="RmForceEnableGen2=1;RMPcieLinkSpeed=0x5;CmpGen3=1" cmp_fuse_gen3=0 cmp_fuse_gen4=0
- No
cmp_fuse_bdf: the fuse overrides then apply to every 2082 card in the host. Addcmp_fuse_bdf=with one card's bus address to limit allcmp_fuse_*overrides, Gen4 included, to that card. - No
CmpGen4key: the PHY init runs at runtime from a trained Gen3 link, the timing CMPGlitch validated on the 8 GB card. RMPcieLinkSpeedstays0x5, the Gen3 value. The Gen4 retrain goes through config space and does not read it.- Keep every registry key in one
NVreg_RegistryDwordsstring. A secondoptions nvidia NVreg_RegistryDwords=line replaces the first and silently drops the Gen3 keys. The per-device VBIOS override parameters are separate parameters; see Kernel parameters and module options.
Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Steps
One card at a time, one attempt per boot. Read LnkSta, never LnkCap, as the result.
- Back up the installed modules and the modprobe file, install the build and the options above, run
depmod -aand update the initramfs. Cold boot, or reload the driver with no GPU clients running. - Check the probe line on every card:
dmesg | grep "CMP fuse: magic_d"showsbefore=0x16680000 clear=0x02000000andafter=0x14680000 (rmw), with no "DID NOT STICK". - Check the go signal:
lspci -s <bdf> -vv | grep LnkCap2lists 2.5-16 GT/s on the card, and the root port lists 16 GT/s too. If the card still stops at 8 GT/s, Gen4 is a no-go this boot; Gen3 still works. - Train Gen3 first with the Gen3 script (PCIe Gen3 on the 10 GB card: RM speed 1, then 2, target link speed 3 on both ends, then RM speed 3). Go on only with 8 GT/s at both ends. After a driver reload the lab's links were still at Gen3 and this step was already done.
- Run the PHY init (below) with the driver bound and the link at Gen3. It does not retrain: the link stays at 8 GT/s.
- Retrain once through config space. Set target link speed 4 on the root port and on the card, then set Retrain Link on the root port, and wait up to about 3 s for Link Training to clear:
setpci -s <root port> CAP_EXP+30.w=0004:000f setpci -s <card> CAP_EXP+30.w=0004:000f setpci -s <root port> CAP_EXP+10.w=0020:0020
- Check the four gates (below). If one fails, set the target link speed back to 3 on both ends (
CAP_EXP+30.w=0003:000f) without retraining, and do not try again this boot. - Do the next card. A card already at 16 GT/s stays there while the next one trains.
Do not use RM's speed control for step 6. On this driver NV2080_CTRL_CMD_BUS_SET_PCIE_SPEED is passed to the closed GSP-RM, which works from a link capability cached at boot; on CMPGlitch's 8 GB card it returned 0x56 (not supported) for Gen4. The lab did not try it on the 10 GB card. Training at 16 GT/s through the LTSSM, set up this way, is the method CMPGlitch posted for the 8 GB card. The whole route is volatile: run steps 4 to 7 after every boot. Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: @._.patrik, 2026-11-16
The runtime PHY init (E1)
The CMP VBIOS leaves out the Gen4 part of the A100 DevInit PHY setup. CMPGlitch worked out the sequence that replays it on an 8 GB card (x16, capacitor mod); the lab ran the same writes on the 10 GB card, through BAR0, from a trained Gen3 link with the driver bound. The register changes were identical on both cards:
| Register (BAR0) | Write | Before | After |
|---|---|---|---|
0x88CE4 |
0x14 |
0x3f |
0x14
|
0x88CE0 |
bits 5:0 = 6 | bits 5:0 = 2 | bits 5:0 = 6 |
CYA_0 (0x8C2C0) |
clear bits 0, 2, 3, 13 and 23 (the A100 value) | 0x060731b2 |
0x060711b2
|
PL_LINK_RATE (0x8C1C0) |
bits 19:16 = 4 | 0x00130036 |
0x00140036
|
LINK_CONFIG_0 (0x8C040) |
bits 19:18 = 0 (reversed encoding: 0 = Gen4, 1 = Gen3, 2 = Gen2, 3 = Gen1) | 0x80544c00 |
0x80504c00
|
PRIV_MISC_1 (0x8841C) |
set bits 11 and 13, clear bits 12, 14, 30 and 31 (as the Gen3 block) | 0x20b40d00 |
0x20b42d00
|
VSEC hierarchy (0x88610) |
set bits 0 and 12 (0x1001) |
no change on these cards | |
After the writes the card's internal link capability (0x88084) reported 16 GT/s as its maximum (speed field 3 to 4). The lab's script read every register back, refused to run unless the device was a 2082 with the driver bound, bit 25 of PCIE_MAGIC_BITS_D clear and the link at Gen3, and stopped on any 0xBADFxxxx or all-ones read. It does not touch 0x8B8C0 or 0x8872C: they stay as the Gen3 block leaves them, and no Gen4 values were needed for them. Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: @._.patrik, 2026-11-16
Pass gates
All four, on each card:
LnkSta16 GT/s at both ends, width kept.lspci -vvon the card and on its root port both readSpeed 16GT/swith the width from before the retrain (x4 here), and sysfscurrent_link_speedreads16.0 GT/s PCIe.- At least 5 GB/s in both directions in pinned 256 MiB host-to-device and device-to-host copies. Gen3 x4 runs at about 3.3 GB/s on these cards, and Gen4 x4 is expected at about 6.4 to 6.8 GB/s; the lab measured 6.37 and 6.76. A link that reads 16 GT/s but copies at Gen3 speed, or anywhere below 5 GB/s, is Gen4 in name only and fails. The 5 GB/s line is the lab's threshold (1.5x Gen3), not a community number.
- No new Xid and no new uncorrectable AER at either end, and the correctable AER count (
aer_dev_correctable) does not climb under load. A climbing correctable count means marginal 16 GT/s equalization, and the replays eat the bandwidth. - A 30-minute soak with gpu-burn and continuous pinned copies at the same time, reading
LnkStaat both ends every 15 s (120 samples), and a check after an idle period. RM still believes the link is Gen3 (#What nvidia-smi shows), so watch for a silent fall back to 8 GT/s.
Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Why the bandwidth gate: Gen4 in name only
On 2026-11-16 several members reported links labelled Gen4 that moved data at Gen3 speed. kingsizexxxl said of his cap-modded card that it was "allegedly gen4x16, but it clearly isnt"; his OpenCL-Benchmark run showed 11.00 GB/s send and 10.91 GB/s receive and labelled the link Gen3 x16. deejay9983: "24gb round trip is gen 3 speeds". patrik read the numbers in that exchange as Gen3 or Gen4 x8. On his RTX 5090 at Gen4, OpenCL-Benchmark also showed low bandwidth, while a Python copy test showed about twice what he had measured on his CMP at "Gen4 x16". Use a pinned copy test you trust, and compare it with Gen3 on the same card. Reported
Source: @kingsizexxxl, 2026-11-16
Source: @kingsizexxxl, 2026-11-16
Source: @deejay9983, 2026-11-16
Source: @._.patrik, 2026-11-16
Source: @._.patrik, 2026-11-16
What nvidia-smi shows
nvidia-smi (pcie.link.gen.current) keeps reporting Gen3 on a card at 16 GT/s. RM caches the link capability at boot, and the Gen4 retrain goes around RM. That is expected and not a failure: judge the link by LnkSta at both ends, sysfs current_link_speed and the copy bandwidth. On CMPGlitch's 8 GB Gen4 card the Gen3 display was likewise described as a display error. Proven 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Source: @wakeupmbrc1, 2026-11-16
Recovery
Neither lab card needed it. These are the planned steps, not yet exercised on a Gen4 failure:
- Gen4 trained but failed a gate: set the target link speed to 3 on the root port and on the card (
CAP_EXP+30.w=0003:000f), set Retrain Link once on the root port, and confirm 8 GT/s inLnkSta. RM's speed control (cmp-pcie-speed 3) is the second choice; it has not been tried from a Gen4 link. - The card fell off the bus (
LnkStareadsffff, Xid 79, or config space unreadable): stop all writes and retries. Do a secondary bus reset on the root port (set bit 6 of its Bridge Control, then clear it), then reboot; if the card is still missing, power cycle. - Back to plain Gen3: remove
cmp_fuse_gen4(andcmp_fuse_magicd_clrorCmpGen4if you added them), update the initramfs and reboot. The build then behaves as the Gen3 build. A power cycle clears every volatile change regardless.
Theory 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes)
Risks
- Warning RM does not know about the Gen4 link. If RM runs its own speed change (a performance-state change or an error recovery), it may retrain to the Gen3 it has cached and drop the link to 8 GT/s without a log line. It did not happen in the lab's 30-minute soak; the idle re-check in the pass gates is there to catch it. 10 GB
Source: @darkmaniac7, 2026-11-17 (lab notes) - Warning One attempt per boot, never a loop. A failed 16 GT/s transition is more likely than a failed Gen3 one to take the card off the bus. A card that dropped off the bus in a Gen3 transition on another machine was later judged to have a damaged PCIe controller. 10 GB
Source: @lxtend, 2026-10-19
Source: @darkmaniac7, 2026-11-17 (lab notes) - Warning Do not burn a Gen3 or Gen4 FF record. Which bit of chain 2 block 39 disables Gen3 is unresolved, the meaning of
PCIE_MAGIC_BITS_Dbits 19, 22, 26 and 28 is not decoded, and the route above needs none of it to be permanent. 10 GB
Source: repo github.com/sayhellotojungle/cmp170hx-ffcalc by sayhellotojungle on GitHub
Source: @darkmaniac7, 2026-11-17 (lab notes) - 16 GT/s equalization held on the lab's two unmodified x4 links to CPU root ports. Risers, slots behind chipsets and switches can be worse: watch the correctable AER count. Source: @darkmaniac7, 2026-11-17 (lab notes)
- The boot-time PHY init (
CmpGen4=1) has not been tested on its own on any card; CMPGlitch's boot-time builds always carried the0x881C0write that wedged GSP boot. The runtime route is the tested one. Source: @._.patrik, 2026-11-16
Source: @darkmaniac7, 2026-11-17 (lab notes) - Gen4 inside a passthrough VM is open, for the same reason as Gen3: the guest cannot write the host root port. Source: @darkmaniac7, 2026-11-17 (lab notes)
- x16 needs the PCIe x16 capacitor mod. Without the caps, mrpandagod's Gen4 attempts fell back to 8 GT/s; Gen4 at x16 on a cap-modded 10 GB card has not been measured here. Source: @djpandaman, 2026-10-08
Other reports
- Reported bruno063185 runs his gate-burned 10 GB card at PCIe Gen4 x4 at boot on 615.71.09, with 80 SMs and 40 GiB, all volatile and no burns beyond the gate. He offered to post the recipe; it had not been posted by 2026-11-16, and no transfer rate was given. It was the first report of Gen4 x4 on this card with no fuse write beyond the gate. 10 GB
Source: @bruno063185, 2026-11-16 - Reported mrpandagod reported Gen3 and Gen4 on 10 GB cards through the RIR route in October 2026, with Gen4 validated behind a PLX switch (his motherboard is Gen3), and x16 needing the capacitor mod. No log or method was posted. 10 GB
Source: @djpandaman, 2026-10-08
Source: @djpandaman, 2026-10-09
Source: @djpandaman, 2026-10-11 - lxwx, who published the Gen3 route, said in October that Gen4 needs "some tricky way to retrain" and that he had no Gen4 slot to test it. matstatman told a new 10 GB owner on 2026-11-10 that his cards "should be able to unlock gen4 x4", x16 with the hardware mod. The lab's result matches both. Source: @lxtend, 2026-10-15
Source: @matstatman, 2026-11-10 - On the 8 GB card, CMPGlitch reports Gen4 x16 on one glitched, cap-modded card; see CMPGlitch fuse research archive and Unlock internals: status and what is next#PCIe Gen3 and Gen4. Source: @._.patrik, 2026-11-16
Credits
- The Gen3 route this builds on: lxwx (@lxtend), 170hx_tips, and evanrodgers' wrapper with its 30-minute two-card acceptance run, posted by @hezz1891. Source: repo github.com/lixiangwuxian/170hx_tips by lixiangwuxian on GitHub, posted by @lxtend, 2026-10-19
Source: repo github.com/evanrodgers/cmp-170hx-rir-burn-util by evanrodgers on GitHub, posted by @hezz1891, 2026-10-26 - The Gen4 PHY init sequence, the LTSSM retrain and the warning about the
0x881C0write: CMPGlitch by @._.patrik, worked out on an 8 GB card. Source: @._.patrik, 2026-11-16 - The first report of volatile Gen4 x4 on a gate-burned 10 GB card: @bruno063185. Source: @bruno063185, 2026-11-16
- The "Gen4 in name only" reports behind the bandwidth gate: @kingsizexxxl, @deejay9983 and @._.patrik. Source: @kingsizexxxl, 2026-11-16
Source: @deejay9983, 2026-11-16 - The in-memory VBIOS override patch in the tested build: Panchovix, vix loader. Source: repo github.com/Panchovix/nvidia-linux-vbios-vix-loader by Panchovix on GitHub
- The port to the 10 GB card, the finding that bit 25 alone is enough, the runtime route, the bandwidth gate and the soaks: the lab. Source: @darkmaniac7, 2026-11-17 (lab notes)
Common questions
Does PCIe Gen4 work on the 10 GB CMP 170HX?
Why does nvidia-smi still say Gen3 after the Gen4 retrain?
Is a link that reads 16 GT/s real Gen4?
Do I need to burn a fuse for Gen4?
Related
PCIe Gen3 on the 10 GB card, RIR fuse burn, PCIe x16 capacitor mod, Gen2 link retrain, Kernel parameters and module options, Verifying an unlock, Unlock without bricking, SW-IFF unlock on the 10 GB card, 10 GB card, Unlock internals: status and what is next, CMPGlitch fuse research archive, Undervolting and power limits, Rumours and claims to verify#PCIe Gen4 on the 10 GB card.