Skip to content

Documenting the Catapult v3 SmartNIC FPGA boards (Dragontails Peak & Longs Peak)

Notifications You must be signed in to change notification settings

tow3rs/catapult-v3-smartnic-re

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 

Repository files navigation

Project overview

This project aims to document the Catapult v3 SmartNIC cards that can be found occasionally on eBay.
These cards seem to be used in several Microsoft projects and Azure servers.
There are at least two versions of these cards:

  • PCI Express "standard" card format, codename 'Longs Peak'
  • Open Compute Project (OCP) Mezzanine card format, codename 'Dragontails Peak', for this form factor there are at least three variants
    • Board with 10AXF40GAA
    • Board with 10AX090N3F40E2SG
    • Board with 10AS066N3F40I2SGES

Each of these SmartNIC features a large Arria 10 FPGA, 4 GB of DDR4-SDRAM, 128MB of flash memory for FPGA configuration, one Mellanox NIC ASIC, one QSFP+ port for 50Gb Ethernet connectivity, three 8x PCIe Gen3 interfaces, an onboard USB programmer and several LEDs.
As usual in these strange pieces of hardware, it's hard to find any kind of documentation.

The main goal is to recycle these boards and use them as a cheap platform for FPGA learning and development from a hobbyist point of view.

Hardware overview

HW diagram

Both variants share the same main hardware components with minor differences:

  • FPGA: Altera Arria 10 with non-standard P/N: 10AXF40GAA.
  • 1 Gb (128 MB) QSPI Flash memory for FPGA configuration: Micron N25Q00AA.
  • Mellanox NIC ASIC.
    • ConnectX-3 Pro for the OCP version.
    • ConnectX-4 Lx for the PCIe version.
  • 5 GB (4.5 GB usable) DDR4 SDRAM, organized as two independent 2.25 GB 72bit interfaces (64 data + 8 ECC): SK hynix H5AN4G6NAFR-UHC.
  • Three I2C Busses.
    • Bus for U17 retimer and QSFP interface management.
    • Bus to manage the following parts: several power distribution components, I2C EEPROM, PCIe clock generation chip, I2C to GPIO chip, temperature sensor and Mellanox NIC ASIC.
    • Bus for Mellanox NIC ASIC management.
  • Two independent 8x PCIe Gen3 interfaces for FPGA.
    • First interface is routed to the first 8x interface of PCIe bifurcation in SAMTEC connector (OCP board) and to 1-8 lanes of the PCIe edge (PCIe board).
    • Second interface is routed to an external connector J8 (OCP board) and to 9-16 lanes of the PCIe edge (PCIe board).
  • One PCIe 8x Gen3 interface for Mellanox NIC ASIC, routed to the second 8x PCIe bifurcation in the SAMTEC connector (OCP board) and to the OCuLink connector (PCIe board).
  • One QSFP+ port connected to the FPGA through the U17 retimer.
  • Onboard USB JTAG programmer based on FTDI FT232H.
  • Several oscillators.
    • Five oscillators for FPGA.
    • One oscillator for Mellanox NIC ASIC.
    • One oscillator for U17 retimer DS250DF810.
    • One oscillator for FT232H.
    • One PCIe clock distribution chip.
  • Several LEDs (9 of them accessible from the FPGA).
  • Several external headers.
    • One I2C header for QSFP cable and retimer management.
    • One I2C header for power regulation components, PCIe clock generation, EEPROM, temperature sensor, etc.
    • One I2C header for Mellanox NIC ASIC management and LED status.
    • One header for JTAG connection using a standard Altera Blaster.
    • One header for FAN connector (Only present in PCIe variant).
    • One header with three GPIO pins wired to FPGA (Only present in PCIe variant).
    • One header with USB connection to FT232H onboard blaster (Muxed with external USB ports).
  • External USB ports for FT232H on-board blaster.
    • Micro USB in the OCP variant.
    • USB type B in the PCIe variant.
  • One unknown header with 2 pins.
  • Miscellaneous discrete components (resistors, capacitors, level shifters, etc.).

Current status

At this moment the following tasks are accomplished:

  • Identification of FPGA clock pins and frequencies (Tested).
  • Identification of FPGA LEDs pins (Tested).
  • Identification of three I2C buses.
    • Two buses for FPGA with identification of all the components attached to both buses (Tested).
    • One bus for Mellanox NIC ASIC external management (Untested).
  • Identification of several GPIO/Unknown purpose pins in FPGA (Untested).
  • Identification of PCI Express interfaces.
    • First 8x interface for FPGA (Tested PCIe communication with computer using a simple design in both boards).
    • Second 8x interface for FPGA (Tested PCIe communication with computer using a simple design in both boards).
    • 8x interface for Mellanox NIC ASIC (Tested PCIe communication with PC in the OCP variant).
  • Identification of network interface pinouts.
    • QSFP cage pins to retimer (Tested).
    • Retimer to FPGA transceiver pins (Tested).
  • Identification of two 72bit DDR4 interfaces (Tested both channels in both boads).
  • Identification of FPGA transceiver pins for communication with Mellanox NIC ASIC (Untested).

The FPGA_Pinouts spreadsheet contains the FPGA pinouts and other findings.
The Header_Pinouts spreadsheet contains the pinouts for the external headers.
The BOM spreadsheet contains a Bill of Materials for some ICs.
The GoldenTop project contains a simple Quartus design with a top level entity exposing all found FPGA pins.

TODO List

The following tasks are pending:

  • Test/verify the untested tasks in previous point.
  • Stablish communication between the FPGA and the Mellanox NIC ASIC.
  • Incomplete or inaccurate constraints for already identified pins.
  • Find some remaining unknown/GPIO pins in FPGA.
  • The OCP card variant mounting the Arria 10 SOC isn't documented in this project.
  • Find the purpose of following components.
    • U20 (Unknown, not found datasheet for this component).
    • U55 (I2C multiplexer). Able to identify and manipulate this component through I2C bus, but the purpose of some some IO pins is unknown.
    • U22 (Level shifter with programmable I/O directions), unknown purpose of some pins, seems to communicate the FPGA with the Mellanox NIC ASIC and with the U55 IOs. This component interconnects several FPGA pins with Mellanox NIC ASIC and U55 pins.
  • Find the datasheet for some small components (maybe level shifters or voltage regulators?).
  • Create a complete documentation (PDF document, etc.).
  • Create a software interface to manage the devices attached to I2C bus.

Some notes

About the contents in this repo

There are the following folders in this repo:

  • Documents: Contains several subfolders with pictures, datasheets, spreadsheets, etc.
  • Projects: Contains the Quartus projects organized in subfolders for each category.
    Some projects have a 'Software' subfolder containing the NIOS II EDS projects with the apps and BSP to be used in conjunction with the hardware project.

About the hardware components

  • The FPGA P/N 10AXF40GAA isn't listed in the Arria 10 datasheets and seems to use a non-standard nomenclature pattern.

    According to this @rombik_su twitter thread, this device is identical to 10AX115N4F40E3SG, also the Quartus device properties window reports the same characteristics.
    However, there are some pictures for the OCP card mounting this P/N 10AX090N3F40E2SG.

    Also, in this picture, some components such as Y4 oscillator is missing, this clock is supposed that drives the DDR4 interface on the top side, maybe this is another variant of these cards.
    Anyways, Quartus is able to generate the bitstreams for the 10AXF40GAA and this strange Arria 10 device seems to accept all of them (10AXF40GAA, 10AX115N4F40E3SG or 10AX090N3F40E2SG)
  • It seems that the PCIe version have a better Mellanox NIC ASIC hardware: ConnectX 3Pro vs ConnectX 4Lx.
    ConnectX 3Pro is 40GbE capable and ConnectX 4Lx is 50GbE.
  • Although the amount of RAM is 5GB only 4.5GB are usable. One of the chips of each channel has only 8 bits wired to the FPGA.
  • Unable to find any documentation for the PCIe side connector J8 in OCP card, only the SAMTEC bottom connector appears in the OCP documents, this one seems to be a proprietary connection. It carries 16 pairs of lanes with differential signals for PCIe (8Rx + 8Tx), one PCIe reset signal (PERST_N) and one cable detection pin.
  • FPGA oscillator Y6 is 644.53125 MHz for PCIe card and 156.250 MHz for OCP card. This oscillator is supposed that clocks the FPGA communication with the Mellanox NIC ASIC. This makes sense that ConnectX 3Pro has a lower speed for interconnect with FPGA (maybe 40GbE or 10GbE?).
  • The dedicated JTAG header (J5) can't be used with some cheap USB Blaster clones.

About the use of Quartus programmer and tools with the onboard USB programmer

When connected to a computer via USB, these boards are recognized as standard FTDI COM ports.
Even though some tools such TopJTAG Probe or OpenOCD allows to use the FTDI ports as JTAG programmers, the Quartus tools doesn't recognize these devices as valid programming hardware.
The following options are reliable to make Quartus tools work:

  • The jtag-mpsse-blaster project allows to use these boards with Quartus tools in Windows and Linux systems.
  • Jan Marjanovič's project https://github.com/j-marjanovic/jtag-quartus-ft232h, allows using the FT232H as a JTAG interface for use with Quartus tools in Linux environments.
  • Applying a small patch to the Arrow USB Programmer lib/dll allows any FTDI chip that implements the MPSSE protocol such as FT2232D, FT2232H, FT4232H or FT232H to be used as USB Blaster by Quartus tools in Windows or Linux environments.

Anyways, a standard Altera USB Blaster connected on the J5 Header works like a charm. Note that some cheap USB Blaster clones doesn't work in these boards.

Configuring the retimer (DS250DF8) for 40/50Gbps Ethernet connectivity

According to the datasheet, on power up, the retimer is configured by default to operate at a rate of 25 Gbps per channel.
The Arria 10 GX transceivers can operate at a maximum rate of 17.4 or 12.5 Gbps depending on speed grade.
If the retimer can't get CDR lock, the output is muted and connection can't be done.
To establish a 40/50 Gbps link, the retimer must be programmed with a rate of 10.3125 Gbps or 12.5 Gbps for each lane. The DS250DF810 programming guide isn't publicly available but there are some TI forum threads talking about quick programming the bitrate for this device.
The retimer is present at address 0x22 and can be programmed via I2C using a soft NIOS processor instantiated in the FPGA or using an external I2C device connected to the J10 header.

The following two steps allows to configure the eight retimer channels to operate at the desired rate.

  1. Enable access to channel registers and configure it to broadcast changes to all channels (this allows to set the bitrate for the eight retimer lanes in a single instruction)
    Writing a value 0x3 in the register 0xFF

  2. Set the rate
    Writing one of the following values in the register 0x2F with mask 0xF0

Single lane rate QSFP port aggregated rate Value for reg 0x2F Write mask
10.3125 Gbps 41.25 Gbps 0x00 0xF0
10.9375 Gbps 43.7428 Gbps 0x10 0xF0
12.5 Gbps 50.0 Gbps 0x20 0xF0

It's possible to configure the retimer to work in non-retimed mode, this allow to bypass CDR lock and output raw data for use with custom bitrates.
Writing the value 0x0 in the register 0x1E with mask 0xE0 enables the bypass mode, the value 0xE0 disables the bypass mode

Fine tunning the retimer requires the study of the device datasheet and the programming guide, but the previous steps allows to test the QSFP port connectivity for 40/50Gbps Ethernet or other network protocols.

The project SuperliteII_V4_QSFP contains a port of one of the IntelFPGA Wiki examples for High Speed Transceiver Demo Designs - Arria 10 Series.
This project uses a custom protocol called Superlite II to test the connectivity in the QSFP port, it only works when two SmartNICs running this project are connected with a QSFP cable. This protocol is incompatible with Ethernet, IB or other network protocols.
This project is basically a copy/paste version of the original one adapted to work in these Catapult V3 SmartNICs, only the following new features where added.

  • FPGA pin assignments and some other minor changes to adapt it for Catapult V3 boards.
  • New option to change the bitrate between 10.3125, 10.9375 or 12.5 Gbps.
  • New option to show QSFP cable information.
  • Removed some unused options to simplify the code.
  • Status information mapped to LEDs.

Connectivity between the FPGA and the Mellanox NIC

Although the NIC has PCIe connectivity and it's detected as a valid network adapter, it's impossible to establish a link with the FPGA.
At this point several posible things can happen.

  • Some IO pins from the NIC remains unidentified and they are used to bring up the chip.
  • Hardware fail in Mellanox NIC side, FPGA side or both.

Some pictures

Dragontails Peak top view

Dragontails Peak bottom view

Longs Peak top view

Longs Peak bottom view

Some FPGA info

Quartus JTAG scan and EPCQ FPGA configuration memory dump

Quartus PIN Planner

Quartus EMIF Toolkit checking the top side DDR4 interface

I2C Bus scan

Testing the three PCI Express interfaces in the Dragontails Peak board

PCIe connection using an USB 3.0 cable and 1x riser adapter for each interface.

Basic PCI Express project with two PCIe interfaces

The custom PCIe devices detected by Windows device manager using the Altera PCI API Driver.

RW Everything tool writing a 01010101 pattern on BAR1 base address of the first PCIe device, at this memory address is mapped an Avalon PIO core wired to the LEDs.

The green LEDs displaying the 01010101 pattern.

Testing the QSFP link at 50 Gbps between two Dragontails Peak boards

QSFP connection done using a Mellanox MC2207130-00A cable.

Superlite II v4 summary screen.

Superlite II v4 ODI eye measure.

External resources

There are some online resources mentioning these boards or the previous generation boards (Catapult v2)

Disclaimer

All trademarks mentioned in this project are the property of their respective owners.
This project has no commercial purpose.

About

Documenting the Catapult v3 SmartNIC FPGA boards (Dragontails Peak & Longs Peak)

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published