The A2O core was a follow-on to A2I, written in Verilog, and supported a lower thread count than A2I, but higher performance per thread, using out-of-order execution (register renaming, reservation stations, completion buffer) and a store queue
You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 
openpowerwtf 6f37e8f136 init 2 years ago
dev init 2 years ago
rel add source 3 years ago
.gitattributes init 3 years ago
LICENSE init 3 years ago
README.md add dev 2 years ago

README.md

A2O

Current Status

  • for now, /rel is the original and /dev has updates:

    • compiles with verilator, iverilog, yosys
    • runs original simple boot code up to test invocation with cocotb (A2L2 interface partially implemented in Python)

To Do

  • continue with cocotb testing
  • add A2Node bridge to WB, and Litex wrapper
  • experiment with parameters to create smaller version(s) for dOpenLane

Original Release

The Project

This is the release of the A2O POWER processor core RTL and associated FPGA implementation (using ADM-PCIE-9V3 FPGA).

See Project Info for details.

The Core

The A2O core was created to optimize single-thread performance, and targeted 3+ GHz in 45nm technology.

It is a 27 FO4 implementation, with an out-of-order pipeline supporting 1 or 2 threads. It fully supports Power ISA 2.07 using Book III-E. The core was also designed to support pluggable implementations of MMU and AXU logic macros. This includes elimination of the MMU and using ERAT-only mode for translation/protection.

The History

The A2O design was a follow-on to A2I, written in Verilog, and supported a lower thread count than A2I, but higher performance per thread, using out-of-order execution (register renaming, reservation stations, completion buffer) and a store queue.

The A2L2 external interface is largely the same for the two cores.

FPGA Implementation Notes

  1. There are lots of knobs available for tweaking generation parameters. Very little experimentation was done to test whether they work, or the effects on area, etc.
  2. Only single-thread generation has been done so far. The FPGA in use has very high utilization with one thread.
  3. A2I used clk_1x and clk_2x (for some of the special arrays), but A2O also uses clk_4x. This (and possibly along with the area congestion) led to changing the clk_1x to 50MHz to lessen timing pressure (both setup and hold misses).

Technology Scaling

A comparison of the design in original technology and scaled to 7nm (SMT2, fixed-point, no MMU):

Freq Pwr Freq Sort Pwr Sort Area Vdd
45nm 2.30 GHz 1.49 W 4.90 mm2 0.97 V
7nm 3.90 GHz 0.79 W 4.17 GHz 0.85 W 0.31 mm2 1.1 V
7nm 3.75 GHz 0.63 W 4.03 GHz 0.67 W 0.31 mm2 1.0 V
7nm 3.55 GHz 0.49 W 3.87 GHz 0.52 W 0.31 mm2 0.9 V
7nm 3.07 GHz 0.32 W 3.60 GHz 0.38 W 0.31 mm2 0.8 V
7nm 2.40 GHz 0.20 W 3.00 GHz 0.25 W 0.31 mm2 0.7 V

These estimates are based on a semicustom design in representative foundry processes (IBM 45nm/Samsung 7nm).

Compliancy

The A2O core is compliant to Power ISA 2.07 and will need updates to be compliant with either version 3.0c or 3.1. Power ISA 3.0c and 3.1 are the two Power ISA versions contributed to OpenPOWER Foundation by IBM. Changes will include:

  • radix translation
  • op updates, to eliminate noncompliant ones and add missing ones required for a given compliancy level
  • various 'mode' and other changes to meet the open specification targeted compliancy level (III-E needs to be changed to III)

Miscellaneous

  1. A2O was not released as a product; the documentation was derived from A2I but is much less complete than the A2I version. The documentation has been edited and updated where possible, but undoubtedly, there remain errors vis a vis the RTL (especially likely in implementation-specific SPRs).

    Please use 'issues' to report errors.

Errata

  1. There is a problem that is being circumvented by setting LSUCR0.DFWD=1, AND limiting the store queue size (currently at 4). While it appears to be directly related to forwarding (L1 DC hit returns 0's instead of data), the store queue size also had to be limited.

    Not debugged at this time; could be related to:

    1. bad generation parm
    2. bad edit for source updates related to compiling for Vivado
    3. ???