[libreriscv.git] / shakti / m_class.mdwn

# Shakti M-Class Libre SoC

This SoC is a propsed libre design that draws in expertise from mass-volume
SoCs of the past six years and beyond, and is being designed to cover just
as wide a range of target embedded / low-power / industrial markets as those
SoCs.  Pincount is to be kept low in order to reduce cost as well as increase
yields.

* See <http://rise.cse.iitm.ac.in/shakti.html> M-Class for top-level
* See [[pinouts]] for auto-generated table of pinouts (including mux)
* See [[peripheralschematics]] for example Reference Layouts
* See [[ramanalysis]] for a comprehensive analysis of why DDR3 is to be used.

## Rough specification.

Quad-core 28nm RISC-V 64-bit (RISCV64GC core with Vector SIMD Media / 3D
extensions), 300-pin 15x15mm BGA 0.8mm pitch, 32-bit DDR3/DDR3L/LPDDR3
memory interface and libre / open interfaces and accelerated hardware
functions suitable for the higher-end, low-power, embedded, industrial
and mobile space.

A 0.8mm pitch BGA allows relatively large (low-cost) VIA drill sizes
to be used (8-10mil) and 4-5mil tracks with 4mil clearance.  For
details see
<http://processors.wiki.ti.com/index.php/General_hardware_design/BGA_PCB_design>

## Targetting full Libre Licensing to the bedrock.

The only barrier to being able to replicate the masks from scratch
is the proprietary cells (e.g. memory cells) designed by the Foundries:
there is a potential long-term strategy in place to deal with that issue.

The only proprietary interface utilised in the entire SoC is the DDR3
PHY plus Controller, which will be replaced in a future revision, making
the entire SoC exclusively designed and made from fully libre-licensed
BSD and LGPL openly and freely accessible VLSI and VHDL source.

In addition, no proprietary firmware whatsoever will be required to
operate or boot the device right from the bedrock: the entire software
stack will also be libre-licensed (even for programming the initial
proprietary DDR3 PHY+Controller)

# Inspiration from several sources

The design of this SoC is drawn from at least the following SoCs, which
have significant multiplexing for pinouts, reducing pincount whilst at
the same time permitting the SoC to be utilised across a very wide range
of markets:

* A10/A20 EVB <http://hands.com/~lkcl/eoma/A10-EVB-V1-2-20110726.pdf>
* RK3288 T-Firefly <http://www.t-firefly.com/download/firefly-rk3288/hardware/FR_RK3288_0930.pdf>
* Ingenic JZ4760B <ftp://ftp.ingenic.cn/SOC/JZ4760B/JZ4760B_DS_REVISION.PDF>
  LEPUS Board <ftp://ftp.ingenic.cn/DevSupport/Hardware/RD4760B_LEPUS/RD4760B_LEPUS_V1.3.2.PDF>
* GPL-violating CT-PC89e <http://hands.com/~lkcl/seatron/>, 
  and <http://lkcl.net/arm_systems/CT-PC89E/> this was an 8.9in netbook
  weighing only 0.72kg and having a 3 HOUR battery life on a single 2100mAh 
  cell, its casework alone inspired a decade of copycat china clone
  netbooks as it was slowly morphed from its original 8.9in up to (currently)
  an 11in form-factor almost a decade later in 2017.
* A64 Reference Designs for example this: <http://linux-sunxi.org/images/3/32/Banana_pi_BPI-M64-V1_1-Release_201609.pdf>

TI Boards such as the BeagleXXXX Series, or the Freescale iMX6
WandBoard etc., are, whilst interesting, have a different kind of focus
and "feel" about them, as they are typically designed by Western firms
with less access or knowledge of the kinds of low-cost tricks deployed
to ingenious and successful effect by Chinese Design Houses.  Not only
that but they typically know the best components to buy.  Western-designed
PCBs typically source exclusively from Digikey, AVNet, Mouser etc. and
the prices are often two to **TEN** times more costly as a result.

The TI and Freescale (now NXP) series SoCs themselves are also just as
interesting to study, but again have a subtly different focus: cost of
manufacture of PCBs utilising them not being one of those primary focii.
Freescale's iMX6 is well-known for its awesome intended lifespan and support:
**ninteen** years.  That does however have some unintended knock-on effects
on its pricing.

Instead, the primary input is taken from Chinese-designed SoCs, where cost
and ease of production, manufacturing and design of a PCB using the planned
SoC, as well as support for high-volume mass-produced peripherals is
firmly a priority focus.

# Target Markets

* EOMA68 Computer Card form-factor (general-purpose, eco-conscious)
* Smartphone / Tablet (basically the same thing, different LCD/CTP size)
* Low-end (ChromeOS style) laptop
* Industrial uses when augmented by a suitable MCU (for ADC/DAC/CAN etc.)

## Common Peripherals to majority of target markets

* SPI or 8080 or RGB/TTL or LVDS LCD display.  SPI: 320x240.  LVDS: 1440x900.
* LCD Backlight, requires GPIO power-control plus PWM for brightness control
* USB-OTG Port (OTG-Host, OTG Client, Charging capability)
* Baseband Modem (GSM / GPRS / 3G / LTE) requiring USB, UART, and PCM audio
* Bluetooth, requires either full UART or SD/MMC or USB, plus control GPIO
* WIFI, requires either USB (but with power penalties) or better SD/MMC
* SD/MMC for external MicroSD
* SD/MMC for on-PCB eMMC (care needed on power/boot sequence)
* NAND Flash (not recommended), requires 8080/ATI-style Bus with dedicated CS#
* Optional 4-wire SPI NAND/NOR for boot (XIP - Execute In-place - recommended).
* Audio over I2S (5-pin: 4 for output, 1 for input), fall-back to USB Audio
* Some additional SPI peripherals, e.g. connection to low-power MCU.
* GPIO (EINT-capable, with wakeup) for buttons, power, volume etc.
* Camera(s) either by CSI-1 (parallel CSI) or better by USB
* I2C sensors: accelerometer, compass, etc.  Each requires EINT and RST GPIO.
* Capacitive Touchpanel (I2C and also requiring EINT and RST GPIO)
* Real-time Clock (usually an I2C device but may be on-board a support MCU)

## Peripherals unique to laptop market

* Keyboard (USB or keyboard-matrix managed by MCU)
* USB, I2C or SPI Mouse-trackpad (plus button GPIO, EINT capable)

## Peripherals common to laptop and Industrial Market

* Ethernet (RGMII or better 8080-style XT/AT/ATI MCU bus)

## Augmentation by an embedded MCU

Some functions, particularly analog, are particularly tricky to implement
in an early SoC.  In addition, CAN is still patented.  For unusual, patented
or analog functionality such as CAN, RTC, ADC, DAC, SPDIF, One-wire Bus
and so on it is easier and simpler to deploy an ultra-low-cost low-speed
companion Micro-Controller such as the crystal-less STMS8003 ($0.24) or
the crystal-less STM32F072 or other suitable MCU, depending on requirements.
For high-speed interconnect it may be wired up as an SPI device, and for
lower-speed communication UART would be the simplest and easiest means of
two-way communication.

This technique can be deployed in all scenarios (phone, tablet, laptop,
industrial), and is an extremely low-cost way of getting RTC functionality
for example.  The cost of, for example, dedicated I2C sensors that provide
RTC functionality, or ADC or DAC or "Digipot", are actually incredibly
high, relatively speaking.  Some very simple software and a general-purpose
MCU does the exact same job.  In particularly cost-sensitive applications,
DAC may be substituted by a PWM, an RC circuit, and an optional feedback
loop into an ADC pin to monitor situations where changing load on the RC
circuit alters the output voltage.  All done entirely in the MCU's software.

An MCU may even be used to emulate SPI "XIP" (Execute in-place) NAND
memory, such that there is no longer a need to deploy a dedicated SPI
NOR bootloader IC (which are really quite expensive).  By emulating
an SPI XIP device the SoC may boot from the NAND Flash storage built-in
to the embedded MCU, or may even feed the SoC data from a USB-OTG
or other interface.  This makes for an extremely flexible bootloader
capability, without the need for totally redoing the SoC masks just to
add extra BOOTROM functions.

## Common Internal (on-board) acceleration and hardware functions

* 2D accelerated display
* 3D accelerated graphics
* Video encode / decode
* Image encode / decode
* Crypto functions (SHA, Rijndael, DES, etc., Diffie-Hellman, RSA)
* Cryptographically-secure PRNG (hard to get right)

### 2D acceleration

The ORSOC GPU contains basic primitives for 2D: rectangles, sprites,
image acceleration, scalable fonts, and Z-buffering and much more.

<https://opencores.org/project,orsoc_graphics_accelerator>

### 3D acceleration

* MIAOW: ATI-compatible shader engine <http://miaowgpu.org/>
* ORSOC GPU contains some primitives that can be used
* SIMD RISC-V extensions can obviate the need for a "full" separate GPU

### Video encode / decode

* video primitives <https://opencores.org/project,video_systems>
* MPEG decoder <https://opencores.org/project,mpeg2fpga>
* Google make free VP8 and VP9 hard macros available for production use only

### Image encode / decode

partially covered by the ORSOC GPU

### Crypto functions

TBD

### Cryptographically-secure PRNG

TBD

# Proposed Interfaces

* RGB/TTL up to 1440x900 @ 60fps, 24-bit colour
* 2x 1-lane SPI
* 1x 4-lane (quad) SPI
* 4x SD/MMC (1x 1/2/4/8-bit, 3x 1/2/4-bit)
* 2x full UART incl. CTS/RTS
* 3x UART (TX/RX only)
* 3x [[I2C]] (in case of address clashes between peripherals)
* 8080-style AT/XT/ATI MCU Bus Interface, with multiple (8x CS#) lines
* 3x PWM-capable GPIO
* 32x EINT-cable GPIO with full edge-triggered and low/high IRQ capability
* 1x I2S audio with 4-wire output and 1-wire input.
* 3x USB2 (ULPI for reduced pincount) each capable of USB-OTG support
* DDR3/DDR3L/LPDDR3 32-bit-wide memory controller

Some interfaces at:

* <https://github.com/sifive/sifive-blocks/tree/master/src/main/scala/devices/>
  includes GPIO, SPI, UART, JTAG, I2C, PinCtrl, UART and PWM.  Also included
  is a Watchdog Timer and others.
* <https://github.com/sifive/freedom/blob/master/src/main/scala/everywhere/e300artydevkit/Platform.scala>
  Pinmux ("IOF") for multiplexing several I/O functions onto a single pin

## I2C

At its own page [[I2C]]

## I2S

<https://github.com/skristiansson/i2s>

## FlexBus

FlexBus is capable of emulating the 8080-style / ATI MCU Bus, as well as
providing support for access to SRAM.  It is extremely likely that it will
provide access to MCU-style Ethernet PHY ICs such as the DM9000, the
AX88180 (gigabit ethernet but an enormous number of pins), the AX88796A
(8/16-bit 80186 or MC68k).

## RGB/TTL interface

<https://opencores.org/project,vga_lcd> full linux kernel driver also available

## SPI

* APB to SPI <https://opencores.org/project,apb2spi>
* ASIC-proven <https://opencores.org/project,spi_master_slave>
* Wishbone-compliant <https://opencores.org/project,simple_spi>

## SD/MMC (including eMMC)

* <https://opencores.org/project,sd_mmc_emulator>
* (needs work) <https://opencores.org/project,sdcard_mass_storage_controller>

# Pin Multiplexing

Complex!  Covered in [[pinouts]].  The general idea is to target several
distinct applications and, by trial-and-error, create a pinmux table that
successfully covers all the target scenarios by providing absolutely all
required functions for each and every target.  A few general rules:

* Different functions (SPI, I2C) which overlap on the same pins on one
 bank should also be duplicated on completely different banks, both from
 each other and also the bank on which they overlap.  With each bank having
 separate Power Domains this strategy increases the chances of being able
 to place low-power and high-power peripherals and sensors on separate
 GPIO banks without needing external level-shifters.
* Functions which have optional bus-widths (eMMC: 1/2/4/8) may have more
 functions overlapping them than would otherwise normally be considered.
* Then the same overlapped high-order bus pins can also be mapped onto
 other pins.  This particularly applies to the very large buses, such
 as FlexBus (over 50 pins).  However if the overlapped pins are on a
 different bank it becomes necessary to have both banks run in the same
 GPIO Power Domain.
* All functions should really be pin-muxed at least twice, preferably
 three times.  Four or more times on average makes it pointless to
 even have four-way pinmuxing at all, so this should be avoided.
 The only exceptions (functions which have not been pinmuxed multiple
 times) are the RGB/TTL LCD channel, and both ULPI interfaces.  

## GPIO Pinmux Power Domains

Of particular importance is the Power Domains for the GPIO.  Realistically
it has to be flexible (simplest option: recommended to be between
1.8v and 3.3v) as the majority of low-cost mass-produced sensors and
peripherals on I2C, SPI, UART and SD/MMC are at or are compatible with
this voltage range.  Long-tail (older / stable / low-cost / mass-produced)
peripherals in particular tend to be 3.3v, whereas newer ones with a
particular focus on Mobile tend to be 1.2v to 1.8v.

A large percentage of sensors and peripherals have separate IO voltage
domains from their main supply voltage: a good example is the SN75LVDS83b
which has one power domain for the RGB/TTL I/O, one for the LVDS output,
and one for the internal logic controller (typical deployments tend not
to notice the different power-domain capability, as they usually supply all
three voltages at 3.3v).

Relying on this capability, however, by selecting a fixed voltage for
the entire SoC's GPIO domain, is simply not a good idea: all sensors
and peripherals which do not have a variable (VREF) capability for the
logic side, or coincidentally are not at the exact same fixed voltage,
will simply not be compatible if they are high-speed CMOS-level push-push
driven.  Open-Drain on the other hand can be handled with a MOSFET for
two-way or even a diode for one-way depending on the levels, but this means
significant numbers of external components if the number of lines is large.

So, selecting a fixed voltage (such as 1.8v or 3.3v) results in a bit of a
problem: external level-shifting is required on pretty much absolutely every
single pin, particularly the high-speed (CMOS) push-push I/O.  An example: the
DM9000 is best run at 3.3v.  A fixed 1.8v FlexBus would
require a whopping 18 pins (possibly even 24 for a 16-bit-wide bus)
worth of level-shifting, which is not just costly
but also a huge amount of PCB space: bear in mind that for level-shifting, an
IC with **double** the number of pins being level-shifted is required.

Given that level-shifting is an unavoidable necessity, and external
level-shifting has such high cost(s), the workable solution is to
actually include GPIO-group level-shifting actually on the SoC die,
after the pin-muxer at the front-end (on the I/O pads of the die),
on a per-bank basis.  This is an extremely common technique that is
deployed across a very wide range of mass-volume SoCs.

One very useful side-effect for example of a variable Power Domain voltage
on a GPIO bank containing SD/MMC functionality is to be able to change the
bank's voltage from 3.3v to 1.8v, to match an SD Card's capabilities, as
permitted under the SD/MMC Specification.  The alternative is to be forced to
deploy an external level-shifter IC (if PCB space and BOM target allows) or to
fix the voltage at 3.3v and thus lose access to the low-power and higher-speed
capabilities of modern SD Cards.

In summary: putting level shifters right at the I/O pads of the SoC, after
the pin-mux (so that the core logic remains at the core voltage) is a
cost-effective solution that can have additional unintended side-benefits
and cost savings beyond simply saving on external level-shifting components
and board space.

# Items requiring clarification, or proposals TBD

## Core Voltage Domains from the PMIC

See [[peripheralschematics]] - what default (start-up) voltage can the
core of the proposed 28nm SoC cope with for short durations?  The AXP209
PMIC defaults to a 1.25v CPU core voltage, and 1.2v for the logic.  It
can be changed by the SoC by communicating over I2C but the start-up
voltage of the PMIC may not be changed.  What is the maximum voltage
that the SoC can run at, for short durations at a greatly-reduced clock rate?

## 3.3v tolerance

Can the GPIO be made at least 3.3v tolerant?

## Shakti Flexbus implementation: 32-bit word-aligned access

The FlexBus implementation may only make accesses onto the back-end
AXI bus on 32-bit word-aligned boundaries.  How this affects FlexBus
memory accesses (read and write) on 8-bit and 16-bit boundaries is
yet to be determined.  It is particularly relevant e.g. for 24-bit
pixel accesses on 8080 (MCU) style LCD controllers that have their
own on-board SRAM.

## Confirmation of GPIO Power Domains

The proposed plan is to stick with a fixed 1.8v GPIO level across all
GPIO banks.  However as outlined in the section above, this has some
distinct disadvantages, particularly for e.g. SRAM access over FlexBus:
that would often require a 50-way bi-directional level-shifter Bus IC,
with over 100 pins!

## Proposal / Concept to include "Minion Cores" on a 7-way pinmux

The lowRISC team first came up with the idea, instead of having a pinmux,
to effectively bit-bang pretty much all GPIO using **multiple** 32-bit
RISC-V non-SMP integer-only cores each with a tiny instruction and data
cache (or, simpler, access to their own independent on-die SRAM).
The reasoning behind this is: if it's a dedicated core, it's not really
bit-banging any more.  The technique is very commonly deployed, typically
using an 8051 MCU engine, as it means that a mass-produced peripheral may
be firmware-updated in the field for example if a Standard has unanticipated
flaws or otherwise requires updating.

The proposal here is to add four extra pin-mux selectors (an extra bit
to what is currently a 2-bit mux per pin), and for each GPIO bank to map to
one of four such ultra-small "Minion Cores".  For each pin, Pin-mux 4 would
select the first Minion core, Pin-mux 5 would select the second and so on.
The sizes of the GPIO banks are as follows:

* Bank A: 16
* Bank B: 28
* Bank C: 24
* Bank D: 24
* Bank E: 24
* Bank F: 10

Therefore, it is proposed that each Minion Core have 28 EINT-capable
GPIOs, and that all but Bank A and F map their GPIO number (minus the
Bank Designation letter) direct to the Minion Core GPIOs.  For Banks
A and F, the numbering is proposed to be concatenated, so that A0 through
A15 maps to a Minion Core's GPIO 0 to 15, and F0 to F10 map to a Minion
Core's GPIO 16 to 25 (another alternative idea would be to split Banks
A and F to complete B through E, taking them up to 32 I/O per Minion core).

With careful selection from different banks it should be possible to map
unused spare pins to a complete, contiguous, sequential set of any given
Minion Core, such that the Minion Core could then bit-bang anything up to
a 28-bit-wide Bus.  Theoretically this could make up a second RGB/TTL
LCD interface with up to 24 bits per pixel.

For low-speed interfaces, particularly those with an independent clock
that the interface takes into account that the clock changes on a different
time-cycle from the data, this should work perfectly fine.  Whether the
idea is practical for higher-speed interfaces or or not will critically
depend on whether the Minion Core can do mask-spread atomic
reads/writes from a register to/from memory-addressed GPIO or not,
and faster I/O streams will almost certainly require some form of
serialiser/de-serialiser hardware-assist, and definitely each their
own DMA Engine.

If the idea proves successful it would be extremely nice to have a
future version that has direct access to generic LVDS lines, plus
S8/10 ECC hardware-assist engines.  If the voltage may be set externally
and accurate PLL clock timing provided, it may become possible to bit-bang
and software-emulate high-speed interfaces such as SATA, HDMI, PCIe and
many more.

# Research (to investigate)

* <https://level42.ca/projects/ultra64/Documentation/man/pro-man/pro25/index25.1.html>
* <http://n64devkit.square7.ch/qa/graphics/ucode.htm>
* <https://dac.com/media-center/exhibitor-news/synopsys%E2%80%99-designware-universal-ddr-memory-controller-delivers-30-percent> 110nm DDR3 PHY
[[!tag cpus]]