Server Hardware on KnightLi Blog

Lenovo HR630x / HR650x Notes: LGA3647, 8259CL, Optane, and Common Pitfalls

Sat, 18 Apr 2026 23:08:00 +0800

Second-hand LGA3647 servers have become increasingly affordable, and Lenovo HR630x / HR650x machines retired from cloud deployments are now entering the homelab bargain-hunting scene. Their appeal is obvious: dual-socket Xeon Scalable, many memory slots, OCP networking, U.2 backplanes, IPMI management, plus the price advantage of some second-generation Xeon OEM CPUs and Optane PMem. It is easy to imagine a “thousand-yuan-class compute box” from these parts.

But these machines are not like upgrading a normal desktop PC. Before buying one, it is worth thinking through several pitfalls: motherboard versions, CPU generation, VRM power limits, memory compatibility, proprietary power supplies, fan noise, rare risers, hard-drive backplanes and trays, BMC passwords, and whether the BIOS is new enough.

This post reorganizes notes from two build reports. The goal is not to reproduce one exact build, but to make the trade-offs and traps around the HR630x / HR650x route clearer.

Platform Positioning

HR630x and HR650x are Lenovo hyperscale LGA3647 server platforms. In simple terms:

HR630x is a 1U machine, thinner and more constrained for expansion.
HR650x is a 2U machine, with more room for expansion, cooling, and installation.
The two platforms share many board-level references, so a lot of practical experience applies to both.
These machines are often cloud-retired units: cheap, but not always complete.

If you want a quiet, power-efficient server to sit next to your desk, these are not the best choice. If your goal is low-cost access to dual Xeons, more PCIe, many memory slots, and remote management, they are attractive.

Check Barebones Completeness First

When buying this kind of barebones system, do not look only at the base price. The real cost depends on what is missing.

Check these items carefully:

Whether both CPU heatsinks are included.
Whether all fans are present.
Whether the power supply count and wattage are enough.
Whether a U.2 / 2.5-inch drive backplane is included.
Whether the drive cables are included.
Whether drive trays are included.
Whether PCIe risers are included.
Whether the OCP NIC is included.
Whether the motherboard is the 24-DIMM or 16-DIMM version.

Some machines look cheap, but if they lack risers, trays, backplanes, or proprietary power supplies, later parts hunting can become more painful than the original purchase. This is especially true for HR650x risers, U.2 backplanes, and drive trays: they are not always easy to find on the used market, and the prices are not always friendly.

CPU: Why the Cheap 8259CL Needs Extra Work

One common value-oriented route is using a second-generation Xeon Scalable OEM CPU such as the Platinum 8259CL. It is cheap, has many cores and threads, and as a second-generation platform CPU, it can work with first-generation Optane persistent memory.

But cheap parts usually come with a reason. The 8259CL is an OEM model with a TDP of about 210W, slightly above the 205W limit many platforms expect by default. That difference looks small, but on some motherboards it can prevent the system from booting unless you modify the VRM controller’s current or power-related limit.

The common method is to use a USB-I2C tool such as MCP2221A, connect it to the motherboard’s VRM I2C interface, and write a new limit into VRM controllers such as PXE1610C. In the reference cases, the command form for HR630x / HR650x looks like this:

`1`	`ModTool.exe -PXE1610C 74 76`

The point is not to copy the command blindly. First confirm your motherboard’s VRM model, I2C header location, SCL, SDA, GND pin order, and addresses. Wiring it wrong or using the wrong platform command is riskier than the CPU itself.

Prepare a Known-Good Boot CPU

If the machine arrives with an old BIOS, or if the VRM modification has not been done yet, installing an 8259CL directly may produce no display at all. In that case, a cheap first-generation Xeon as a boot CPU can make troubleshooting much easier.

A boot CPU is useful for:

Entering BIOS and checking the version.
Updating BIOS and BMC.
Confirming that the motherboard, memory, power supply, and fans are healthy.
Excluding basic hardware faults before modifying the VRM.

If the seller has already updated the BIOS and the machine can boot directly, the boot CPU may not be needed. For beginners, though, it greatly reduces debugging difficulty.

Optane PMem Is the Highlight of This Platform

Second-generation Xeon Scalable supports first-generation Intel Optane DC Persistent Memory, also known as DCPMM / PMem. It installs into DIMM slots and can be configured in BIOS as memory mode or as persistent block storage.

This is one reason CPUs like the 8259CL are attractive: when large DDR4 RDIMM / LRDIMM modules become expensive, used Optane PMem can be a low-cost way to add capacity.

However, Optane is not a complete replacement for normal DRAM. Keep these points in mind:

It requires a second-generation Xeon with DCPMM support.
The BIOS must support and correctly identify Optane.
DRAM is usually still required as cache or as a companion memory tier.
Slot placement and channel pairing should follow the Lenovo manual.
Performance sits between DRAM and SSD, so do not expect normal DRAM behavior.
You can create namespaces and use them as block devices such as /dev/pmem0.

If the goal is low-cost high-capacity memory experimentation, Optane is interesting. If the goal is maximum memory bandwidth, a low-channel-count Optane setup may not be suitable.

Memory Slot Versions and Compatibility

HR630x / HR650x may come in 24-DIMM and 16-DIMM versions. Before ordering, ask the seller for clear motherboard photos instead of trusting the listing title.

For memory, it is safer to buy the same brand, frequency, capacity, and rank in one batch. The reference build notes mention unstable recognition when mixing modules, and in some cases even swapping CPU or memory positions affected which memory was detected.

Safer rules:

Follow the official manual’s slot population order.
Avoid mixing too many brands and specifications.
When unsure, boot with the minimum configuration first.
On dual-socket platforms, verify the memory channels attached to each CPU.
When using Optane, double-check the DRAM and PMem channel pairing.

Server memory is not “if it fits, it boots.” The larger and more mixed the capacity, the higher the debugging cost.

Also, memory cannot be inserted randomly. Lenovo’s official documentation defines the DIMM population order for independent mode. Before building, check the slots against the manual, then expand from the minimum bootable configuration. This matters even more with dual sockets, mixed capacity, mixed rank, or Optane PMem; an incorrect layout may cause no boot, missing memory, or only the channels under one CPU being detected.

Do Not Underestimate Fans and Noise

These machines were not designed for bedrooms or home offices. The 1U HR630x is especially obvious: fan speeds are high, the sound is sharp, and the default boot policy can be very conservative.

According to the build notes, the default fan speed is high, and IPMI / CLI control is needed to reduce noise. After tuning, idle noise can be much better, but under full load with two high-power CPUs, the system still needs enough airflow.

When tuning fans, monitor:

CPU temperature.
VRM temperature.
PCH temperature.
Memory temperature.
PSU temperature.
Inlet and outlet temperatures.

Do not only watch the CPU. Many chips on server boards depend on chassis airflow. If fan speed is lowered too much, the CPU may still look fine while the PCH, VRM, or NIC overheats.

Changing Fan Speed

The fans on HR650x / HR630x can be controlled through IPMI raw commands. The community script uses this command format:

`1`	`ipmitool -I lanplus -H <BMC_IP> -U <USER> -P '<PASSWORD>' raw 0x2e 0x30 00 00 <SPEED>`

<SPEED> can be treated as the target fan percentage. For example:

# Set to 10%
ipmitool -I lanplus -H 192.168.1.100 -U ADMIN -P 'password' raw 0x2e 0x30 00 00 10

# Set to 35%
ipmitool -I lanplus -H 192.168.1.100 -U ADMIN -P 'password' raw 0x2e 0x30 00 00 35

# Set to 100%, useful for full-speed testing or thermal fallback
ipmitool -I lanplus -H 192.168.1.100 -U ADMIN -P 'password' raw 0x2e 0x30 00 00 100

If you run the command from the server OS itself and the IPMI kernel modules are loaded, you can avoid the BMC network path:

`1`	`ipmitool raw 0x2e 0x30 00 00 20`

Before changing fan speed, confirm that ipmitool can read sensors:

1
2

ipmitool -I lanplus -H <BMC_IP> -U <USER> -P '<PASSWORD>' sensor
ipmitool -I lanplus -H <BMC_IP> -U <USER> -P '<PASSWORD>' sdr

If local ipmitool reports no usable interface, load these modules on Linux:

1
2
3

modprobe ipmi_devintf
modprobe ipmi_msghandler
modprobe ipmi_si

A safer approach is not to lock the fan at one very low speed, but to use temperature-based steps. For example:

CPU below 40°C: 10%
CPU 40°C to 45°C: 14%
CPU 45°C to 50°C: 20%
CPU 50°C to 60°C: 50%
CPU 60°C to 80°C: 80%
CPU above 80°C: 100%

This can be implemented with shell, Python, or a systemd timer: read CPU temperature every few seconds, then write the corresponding fan percentage. The community HR650X-IPMI-Auto-Fan script follows this idea.

For manual tuning, start conservatively. Try 20% at idle first, verify that CPU, PCH, VRM, memory, NIC, and PSU temperatures are stable, then gradually try 14% or 10%. For full-load testing, do not begin with low fan speeds; start at 50% or higher, confirm cooling headroom, then find a balance between noise and temperature.

IPMI raw commands are vendor OEM commands, and behavior may differ across BMC firmware versions. Before running them, make sure the machine can read sensors normally and keep a terminal ready to switch the fans back to high speed. If temperature readings are abnormal, sensors show na, or fan speed does not change as expected, do not continue lowering the speed.

Power Supplies, Risers, Backplanes, and Drive Trays

One big HR650x pitfall is that the power supply interface and many expansion parts are not standard PC parts. The PSU uses a Lenovo-specific form factor, so replacing or adding one can be expensive.

Check risers in advance as well. Different risers support different card layouts, such as full-height full-length, full-height half-length, and half-height half-length. If you plan to install GPUs, HBAs, 25G/40G NICs, or NVMe adapters later, confirm the riser type before buying the machine.

Drive backplanes also have multiple configurations. You may see 2U.2, 4U.2, 8U.2, or 2.5-inch bay backplanes. Backplanes, cables, trays, RAID cards, or HBAs can all become extra costs.

A practical suggestion: if you only want the machine to boot and run compute jobs, do not rush to complete every tray and backplane. If your goal is all-flash storage or high expansion, include those parts in the total budget from the beginning.

BMC, BIOS, and Management

Cloud-retired machines often come with unknown BMC passwords. If you can enter BIOS, you can usually create or reset a management user there. If you can boot into an OS, ipmitool can also manage BMC users.

It is usually best to update BIOS and BMC to a newer stable version, for three reasons:

Support for more second-generation Xeon models.
Better Optane PMem detection and management.
Fixes for BMC, fan policy, or hardware compatibility issues.

The references mention that HR630x / HR650x may require BIOS updates for 8259CL and Optane. Different machine batches vary: some sellers have already updated them, while others require manual work.

The BIOS and BMC downloads for HR650x can be found through Lenovo’s support page:

`1`	`https://datacentersupport.lenovo.com/cn/zc/products/servers/thinksystem-hyperscale/hr650x/7x57/7x57cto1ww/j300cvx2/downloads/driver-list/`

Also, HR650x supports Above 4G Decoding, but Resizable BAR support is not ideal. If you plan to install large-VRAM GPUs or use the machine for GPU compute, confirm BIOS options and power-cable plans first.

Who Should Try This

This kind of machine is better suited for people who:

Need many cheap x86 threads.
Can tolerate idle power draw and noise.
Have space for rackmount servers.
Are willing to read manuals, inspect board labels, and use a multimeter.
Can accept uncertainty in used-server parts.
Have patience for IPMI, BIOS, VRM, and DCPMM troubleshooting.

It is less suitable for people who:

Only want a quiet NAS.
Want a low-power 24/7 mini server.
Do not want to deal with BMC, fans, risers, backplanes, or proprietary PSUs.
Do not have spare CPUs, spare memory, or basic debugging tools.
Cannot accept BIOS updates, VRM modification, and fan tuning after purchase.

Summary

The main value of HR630x / HR650x is that they provide a low-cost LGA3647 dual-socket server platform. With cheap second-generation Xeons such as 8259CL and Optane PMem, they can become homelab compute nodes with impressive thread counts, memory capacity, and remote management.

The pitfalls are just as clear: high-power OEM CPUs may not be supported by default, so an MCP2221A VRM modification may be needed; memory slot versions and compatibility must be checked; fan noise and idle power cannot be treated like consumer hardware; risers, backplanes, drive trays, and power supplies can all add cost.

If the budget is tight and you enjoy tinkering, this is an interesting route. If you want something stable, quiet, and low-maintenance, calculate total power, noise, included parts, and future maintenance cost before jumping in.

References

lyc8503: AIO Ep19. Lenovo HR630x(HR650x) Server Build Log: https://blog.lyc8503.net/post/19-first-rack-server-hr630x/
一只白泽_沧渊: HR650x Troubleshooting Notes, source: bilibili: https://www.bilibili.com/read/cv36922131/?opus_fallback=1
Vision0220: HR650X-IPMI-Auto-Fan: https://github.com/Vision0220/HR650X-IPMI-Auto-Fan
Lenovo HR630x official support page: https://datacentersupport.lenovo.com/us/en/products/servers/thinksystem-hyperscale/hr630x
Lenovo HR650x official support page: https://datacentersupport.lenovo.com/cn/zc/products/servers/thinksystem-hyperscale/hr650x
Lenovo HR650x BIOS/BMC download page: https://datacentersupport.lenovo.com/cn/zc/products/servers/thinksystem-hyperscale/hr650x/7x57/7x57cto1ww/j300cvx2/downloads/driver-list/
Lenovo SR650 DIMM installation order for independent mode: https://pubs.lenovo.com/sr650/zh-CN/dimm_installation_dram_independent_mode_2

LGA3647 high TDC OEM CPU lighting idea: modify the ICC_MAX of VRM

Sat, 18 Apr 2026 22:32:00 +0800

On the LGA3647 platform, many OEM versions of Xeon Scalable processors are very affordable, but they may not light up when you get them on ordinary server motherboards or workstation motherboards. The typical phenomenon is power on, fans spinning, BMC or IPMI accessible, but the CPU initialization phase is stuck without even entering the real x86 execution process.

This kind of problem is not necessarily a bad CPU or just missing BIOS microcode. There is a long-term maintenance discussion in the ServeTheHome forum. The core idea is: the TDC requirements of some OEM CPUs are higher, and the ICC_MAX reported or restricted by the motherboard VRM by default does not meet the requirements, causing the platform to refuse to start in the early stages.

The solution is not to simply change the TDP, but to access the VRM controller through I2C/PMBus and change the VRM’s ICC_MAX reference value to 0xFF, which is 255A in common parlance.

This article will sort out the principles, processes, common motherboard wiring methods and command examples, but it is still not recommended to use it as a mindless copying tutorial. Different motherboards may have different VRM models, I2C pins, addresses, and BIOS restrictions. Be sure to return to the original post to check the latest information before taking action.

High TDC is not the same as high TDP

Many people will simply call this type of CPU “high-power CPU”, but a more accurate judgment point is TDC, which is the continuous current related limit, not just TDP.

For example some OEM models have a TDP that is only a few watts higher than a standard retail model, but the TDC may be set to 255A. If an ordinary motherboard is only prepared according to the current range of the standard SKU, it may think that the VRM capabilities do not match during power-on initialization and thus not continue to start.

This is why some OEM CPUs around 210W, which appear to be only a little more powerful than the 205W standard models, may still fail to light up on the default motherboard.

What does the ICC_MAX modification solve?

ICC_MAX can be understood as the current capability reference value declared by VRM for the platform. The method discussed in the forum is to use a USB-I2C tool to connect to the motherboard VRM controller, and use the tool to write ICC_MAX of the relevant controller as FF.

This modification mainly solves the following problems:

The motherboard does not accept high TDC OEM CPUs by default.
After the CPU is installed, the motherboard powers up but does not execute x86 code.
The BIOS itself already supports the corresponding generation of CPUs, but the VRM current capability statement stuck on booting.

It should be noted that according to the explanation of the author of the original post, this operation does not turn off all protection mechanisms. Hardware protections such as single-phase OCP and overheating protection will not disappear automatically just by changing ICC_MAX. But that doesn’t mean there’s no risk, because you’re still modifying the parameters of the motherboard’s power controller.

Commonly involved VRM controllers

There are several main types of common VRM controllers on the LGA3647/C620 platform:

PXE1610C
TPS53679
TPS53678
MP2955A

Different motherboards use different controllers and have different tool parameters. For example, the command form that appears many times in the original post is similar:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 50 52`

Here PXE1610C is the VRM controller type, 50, 52 is the I2C address. Dual-socket motherboards usually have two addresses, corresponding to different CPU power supply areas.

If the address is incorrect, the tool will usually return that the device cannot be found. Later versions of the tool also provide a scanning function that can be used to find possible VRM addresses:

`1`	`MCP2221a_iccmax_FF.exe -scan start end`

Don’t make random guesses about the specific address. It’s best to confirm it based on the original post, motherboard silk screen, VRM chip model and existing successful cases.

What to prepare

The most basic tools usually include:

MCP2221A USB-I2C adapter.
Dupont wire or fine wire.
Multimeter to confirm continuity between GND and pins.
Prepare a soldering iron, flux, and magnifying glass if necessary.
The corresponding MCP2221a_iccmax_FF tool.
A Windows computer used to execute the flash command.

If BIOS modification is also involved, you may need to:

External BIOS programmer.
A tool that can stably read and write 32MB SPI Flash.
Hot air station or suitable welding equipment.
Hex editor, or AMI BIOS related tools.

Only modifying VRM parameters and removing the BIOS chip to modify the BIOS are two completely different risk levels. The former is usually connected to I2C, while the latter involves desoldering and flashing SPI Flash, which has higher rollover costs.

Basic process

The overall idea can be broken down into 6 steps.

1. Confirm that the CPU is a high TDC OEM model

First make sure that the CPU is not a common incompatibility problem. Need to find out:

CPU specific model.
Which generation does Xeon Scalable belong to.
Does the BIOS support this generation of CPUs.
Is the microcode corresponding to stepping missing?
CPU TDP and TDC conditions.

If the BIOS doesn’t even support 2nd generation Xeon Scalable, there’s usually no point in just changing the VRM. For example, some users mentioned that CPUs such as 8259CL require at least the motherboard BIOS to support the second generation, and the old BIOS may not continue to initialize at all.

2. Confirm whether the motherboard has any successful cases

The original post has long compiled a batch of motherboards or barebones that have been tried by others. Common ranges include:

Supermicro X11SPA / X11SPW / X11SPM / X11SPi / X11SPL
Supermicro X11DPi / X11DPH / X11DAi / X11DPG / X11DDW
Intel S2600BP / S2600WF / S2600ST
Dell Precision T7820 / T7920 / R7920
Selected Dell PowerEdge, HPE ProLiant Gen10, Lenovo, Cisco, Inspur platforms

This list will be updated with forum discussions. Before purchasing a motherboard or CPU, it is best to check if there are any successful records of the same model, same PCB version, and same BIOS version.

3. Find the VRM model and I2C pins

This step is the most error-prone. You need to confirm:

VRM controller model number on the motherboard.
Where do SCL, SDA, and GND come from.
Do you need to unplug certain JVRM jumpers.
Which I2C addresses correspond to single or dual channels respectively.

The pins on different boards are completely different. For example, some Supermicro boards in the forum can connect SCL/SDA through the JVRM jumper cap position, some HPE models use specific connector positions, and some Dell servers need to be flashed in a specific standby state.

Don’t just look at “same brand” and then apply wiring. Even if they are all X11 series, the VRM models and wiring methods may be different.

4. Connect the MCP2221A

There are only three common connections:

SCL vs. SCL
SDA vs. SDA
GND vs. GND

Don’t connect the wrong power pin. In many scenarios, only I2C signals and ground wires are needed, and there is no need to supply power to the motherboard from MCP2221A.

Before connecting, it is recommended to cut off the power first, confirm GND with a multimeter, and then check the motherboard manual, forum pictures or silk screen. If you connect the wrong wires, the tool may not be able to find the device, or the VRM, BMC, or motherboard may be damaged.

5. Let the motherboard enter a state where VRM can be accessed

The status is different on different platforms. Some require installing the CPU and memory and booting into the BIOS, while others require just plugging in power and putting the VRM into standby or standby mode.

For example, in the case of HPE DL380/DL360 Gen10 in the forum, it is mentioned that the addresses of PXE1610C are 62 and 64, which can be connected to MCP2221A through the corresponding connectors and then written; Dell R640/R740/T640 platforms have special requirements such as plug-in only and no CPU installed.

This cannot be unified into one rule. Cases must be checked according to the motherboard model.

6. Execute the write command and verify

After confirming the controller type and address, execute a command similar to the following:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 50 52`

or:

`1`	`MCP2221a_iccmax_FF.exe -TPS53679 58 60`

After writing, it is recommended to read or repeat the execution to confirm that the parameters are indeed saved, then cut off the power, replace the CPU and memory, and test whether POST can be performed normally.

Common motherboard and platform modification methods

Common wiring and commands are organized below by platform. The pin here refers to the debugging interface, jumper or connector position in the original post or the compiled document, not the CPU Socket pin. Be sure to confirm with the motherboard silk screen, photos and multimeter before actual operation.

Supermicro X11DPi-N / X11DPi-NT rev.2.x

There are many cases of using MP2955A on this type of board. A common approach is to find the JVRM1, JVRM2 jumpers, remove the jumper caps and connect to the MCP2221A.

wiring:

1
2
3

JVRM1 pin 2 -> SCL
JVRM2 pin 2 -> SDA
USB2/USB3 pin 7/8 -> GND

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -MP2955A 20 21`

If it is a dual-circuit board, the two addresses usually correspond to the two CPU power supply controllers. After flashing, it is recommended to re-run the command or check the output to confirm that the parameters have been written.

Supermicro X11SPL-F / X11SPi-TF

A common controller for this type of board is TPS53679. The wiring usually goes through the JVR or JVRM interface.

wiring:

1
2
3

JVR(M)1 pin 1 -> SCL
JVR(M)1 pin 2 -> SDA
JVR(M)1 pin 3 -> GND

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -TPS53679 58`

Special note: There may also be a BIOS internal 165W limit on X11SPL. In other words, after the VRM is written, if the BIOS still limits power consumption or lacks microcode, the CPU may still not start normally.

Supermicro X11DPL-i / X11DPH-i

The common connection methods of this type of dual-circuit board also revolve around JVRM or related debugging interfaces.

wiring:

1
2
3

JVR(M)1 pin 1 -> SCL
JVR(M)1 pin 2 -> SDA
JVR(M)1 pin 3 -> GND

Common flash commands:

`1`	`MCP2221a_iccmax_FF.exe -TPS53679 58 60`

If you are using the version that introduces SCL/SDA separately through JVRM1/2, there is also a way to organize and write it as follows:

1
2
3

JVRM1 pin 2 -> SCL
JVRM2 pin 2 -> SDA
USB2 pin 7/8 -> GND

Commands are still executed per controller and address:

`1`	`MCP2221a_iccmax_FF.exe -TPS53679 58 60`

Supermicro X11DPH / X11DPG

The common connection method for this type of board is to remove the JVRM1/2 jumper cap and connect I2C.

wiring:

1
2
3

JVRM1 pin 2 -> SCL
JVRM2 pin 2 -> SDA
T-SGPIO1 pin 3/6 -> GND

The X11DPG related case in the original post mentioned that the PXE1610C addresses are available 50 and 52:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 50 52`

If the tool returns that the device cannot be found, do not continue to write, but first confirm the controller model and address. Some boards have different VRM controllers.

Supermicro X11DPU-G6

Common wiring:

1
2
3

JVRM1 pin 2 -> SCL
JVRM2 pin 2 -> SDA
T-SGPIO1 pin 3/6 -> GND

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 50 52`

Supermicro X11SPA-F / X11SPA-TF

The common controller for this type of single-channel workstation board is PXE1610C, and usually only one address needs to be written.

wiring:

1
2
3

JVR1 pin 1 -> SCL
JVR1 pin 2 -> SDA
JVR1 pin 3 -> GND

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 50`

Dell Precision T7920

A common case for Dell Precision T7920 is to connect the MCP2221A, boot the workstation into the BIOS, and then write two VRM addresses.

Flash status:

安装 CPU 和内存
开机进入 BIOS
保持机器运行
执行刷写命令

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 60 62`

Dell PowerEdge T640 / R640 / R740

These PowerEdge platforms are not exactly the same as Precision workstations. The compiled information emphasizes: Do not install the CPU when flashing, only plug in the power supply, and let the machine enter the VRM waiting state before writing.

Common wiring:

1
2
3

Pin 1 -> SCL
Pin 2 -> GND
Pin 3 -> SDA

Flash status:

不要安装 CPU
只接入电源
等待进入 VRM 可访问状态
执行刷写命令

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 60 62`

Dell server BIOS usually has a signature and verification mechanism, and it is not recommended to modify the BIOS at will. For this type of platform, it is preferable to only modify the VRM ICC_MAX and select a CPU that is already supported by the BIOS.

Lenovo SR650 / HR650 / SR630 / HR630

Platforms such as Lenovo need to confirm the specific model, VRM model and I2C pins first. Taking the finishing case of HR650X as an example, the VRM uses PXE1610C, and the I2C pins are located at the debugging position near the CPU1 slot. The sequence can be confirmed according to the data.

HR650X example wiring:

1
2
3

SCL
SDA
GND

Example flash status:

1
2
3

接好 MCP2221A
服务器进入 BIOS
执行刷写命令

Example command:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 74 76`

Lenovo platforms often also involve BIOS microcode issues. In the compiled information, SR630, HR630, SR650, HR650, P720, and P920 are classified as those that may require BIOS modification. Special attention should be paid especially when using early stepping CPUs such as P8124 and P8136.

Lenovo ThinkStation P920

The P920 is wired similarly to some Dell workstations.

wiring:

1
2
3

Pin 1 -> SCL
Pin 2 -> SDA
Pin 3 -> GND

Flash command:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 60 62`

HPE DL380 Gen10 / DL360 Gen10

In the forum case of HPE DL380/DL360 Gen10, the addresses of PXE1610C are 62 and 64, which are connected through the J226 type connector on the motherboard.

Flash status:

安装 CPU1 + CPU2
安装必要内存
开机进入 BIOS
连接 MCP2221A
执行刷写命令

Write respectively:

1
2

MCP2221a_iccmax_FF.exe -PXE1610C 62
MCP2221a_iccmax_FF.exe -PXE1610C 64

You can also write two addresses at once:

`1`	`MCP2221a_iccmax_FF.exe -PXE1610C 62 64`

It is also not recommended to change the BIOS of HPE ProLiant. Quad-socket machines such as DL560/DL580 Gen10 may also have CPU whitelist issues. Before selecting a CPU, check the support list of the corresponding model.

Motherboard not listed

Do not apply commands directly to motherboards that are not listed. The correct process is:

Check the VRM controller model number.
Find the I2C debug interface or JVRM jumper location.
Confirm SCL, SDA, GND.
Use the scan command to check the address.
Then press the controller model to write.

Scan command example:

`1`	`MCP2221a_iccmax_FF.exe -scan 20 7F`

If there is no result from the scan, first check whether SCL and SDA are connected reversely, whether GND is correct, and whether the motherboard is in VRM accessible state. Do not attempt to write without verifying the controller.

BIOS may also be the second threshold

The ICC_MAX modification only solves the VRM current declaration problem, which does not mean that all CPUs can be started directly.

Also need to pay attention to the BIOS:

Whether 1st or 2nd generation Xeon Scalable is supported.
Whether to include microcode corresponding to CPU stepping.
Is there a TDP/TDC whitelist or power consumption cap.
Whether there is manufacturer signature verification to prevent the modified BIOS from booting.

It was mentioned in the forum that some Supermicro X11SPL /

If it is an earlier stepping or special OEM model such as P8124 and P8136, simply changing the VRM may not be enough. You may also need to add microcode or bypass manufacturer restrictions.

BIOS 165W limit modification for X11SPL/X11SPM

There is an internal 165W limit in some Supermicro X11SPL and X11SPM BIOS. The way to organize the information is to use a hexadecimal editor such as HxD to modify the BIOS file, change the relevant location from A5 to FF, and at the same time correct the ProjectPeiDriver.ffs check related bytes.

It is not recommended for people without BIOS recovery capabilities to operate directly here. At least prepare:

Original BIOS backup.
IPMI BIOS recovery solution available, or external programmer.
Can confirm how the modified BIOS can be flashed.
The ability to desolder or offline flash the SPI Flash after a problem occurs.

There are two groups of general power consumption value replacements:

1
2

查找：6C 68 A5 00 00 00 68
替换：6C 68 FF 00 00 00 68

1
2

查找：FB B9 A5 00 00 00 5E
替换：FB B9 FF 00 00 00 5E

Different BIOS versions also need to correct the check byte accordingly. The common ones are summarized as follows.

X11SPL

X11SPL BIOS 3.6：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF C3 23
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF C3 6F

X11SPL BIOS 3.9：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 62 6C
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 62 B8

X11SPL BIOS 4.0：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 81 7A
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 81 C6

X11SPM

X11SPM BIOS 3.4：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 82 9D
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 82 E9

X11SPM BIOS 3.5：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 82 65
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 82 B1

X11SPM BIOS 3.8a：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF C3 EE
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF C3 3A

X11SPM BIOS 3.9：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 62 4A
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 62 96

X11SPM BIOS 4.0：

1
2

查找：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 81 7E
替换：26 22 9C 73 64 32 54 44 99 1C 8D C4 4A 73 D6 AF 81 CA

After modification, save it as a new BIOS file and do not overwrite the original file. It is best to compare the file size, checksum and modification location again before flashing.

Microcode completion process of early stepping CPU

If you use older stepping OEM CPUs such as P8124 and P8136, some motherboards may not have corresponding microcode in the BIOS. The general process in organizing data is:

Remove the BIOS SPI Flash and read the original BIOS.
Keep at least two original backups.
Open the BIOS with MMTool or a similar AMI BIOS tool.
Enter the CPU Patch area and check the existing microcode.
Xeon Scalable microcode for inserting missing stepping.
Save new BIOS.
Use a programmer to write back to SPI Flash.
Solder back the BIOS chip and test boot.

Special attention is needed: many LGA3647/C620 platform BIOS are 32MB, and cheap CH341A and ordinary burning folders are not necessarily reliable. The compilation of information also emphasized that it is not recommended to directly use the clip to read and write online, because the BMC or PCH may occupy the bus after the server is powered on, and the reading and writing results are unstable. A more stable way is to disassemble the chip to read and write offline, but this also means higher soldering risks.

Risk point

This modification seems to be just a few lines and a command, but the risk is not low.

The most common pitfalls are:

Wrong connection of SCL/SDA/GND.
Wrong VRM controller address found.
The motherboard version is different and someone else’s wiring is used.
The BIOS does not support the CPU and mistakenly believes that the VRM has not been changed successfully.
The VRM has insufficient heat dissipation and is unstable under full load for a long time.
The SPI Flash is damaged when modifying the BIOS.
The server manufacturer’s whitelist or signature mechanism causes the server to still not start after modification.

In addition, high TDC CPUs are not necessarily more cost-effective. Once the second-hand market gets hyped up, and with the cost of tools, welding, time and turnaround, it may be better to just buy an officially supported CPU or motherboard.

Suitable for who to try

More suitable for people who want to try:

There are already LGA3647 platforms and high TDC OEM CPUs.
Can understand the motherboard silk screen, chip model and forum wiring diagram.
Have basic welding and multimeter experience.
Can accept the cost of motherboard or CPU rollover.
I would like to check the latest feedback on the same model motherboard as in the original post.

Not recommended for those who try:

I just want to save money to install a stable workstation.
No experience in soldering or hardware troubleshooting.
I only have one motherboard on hand, and if it breaks there is no replacement.
It’s not clear the relationship between CPU stepping, BIOS microcode, and VRM models.

summary

LGA3647 high TDC OEM CPU cannot light up. In many cases, it is not simply that the TDP is too high, but that a higher TDC/current requirement is detected during early initialization of the platform. The approach in the ServeTheHome forum is to access the VRM controller through MCP2221A and adjust ICC_MAX to FF/255A to allow the motherboard to accept this type of OEM CPU.

The whole process can be understood as:

Confirmed CPU and BIOS generation support.
Confirm motherboard and VRM controller models.
Find the SCL, SDA, GND and I2C addresses.
Write ICC_MAX = FF with MCP2221A.
Address BIOS microcode or power consumption limitations when necessary.
Finally, the focus is on verifying the VRM temperature, overall machine stability and long-term load performance.

This is less of a regular upgrade tutorial and more of a detour for hardware gamers to the limitations of OEM platforms. The more detailed the information is, the slower the process is, and the higher the success rate.

Reference link

ServeTheHome forum original post: https://forums.servethehome.com/index.php?threads/vrm-modify-icc_max-to-run-high-tdc-oem-cpu.38686/
ServeTheHome Forum page 8 Summary: https://forums.servethehome.com/index.php?threads/vrm-modify-icc_max-to-run-high-tdc-oem-cpu.38686/page-8
JDDKCN/KCNVrmModTool：https://github.com/JDDKCN/KCNVrmModTool
Related Chinese compilation: https://aigcdaily.cn/news/b24egiog9ukwhyr/