This repo contains the bsg_manycore source code with contributions from the Bespoke Silicon Group and others.
The tile based architecture is designed for computing efficiency, scalability and generality. The two main components are:
- Computing Node: Purpose-designed high-performance core that runs at 1.4GHz@16nm, but nodes also can be any other accelerators.
- Mesh or Ruche Network : Dimension ordered, single flit network with inter-nodes synchronization primitives (mutex, barrier etc.)
Without any custom circuits, a 16nm prototype chip with 16x31 tiles on a 4.5x3.4 mm^2 die space achieves 812,350 aggregated CoreMark score, a world record. Many improvements have been made since this previous version.
- Chip gallery, publications, and artworks:
- See this website: http://bjump.org/manycore/
- Bleeding edge features and proceedings:
- BaseJump Manycore Accelerator Network
- Version tag: tile_group_org_master
- The mesh network architecture, protocols, constrains and guidelines.
- HammerBlade Manycore Technical Reference Manual
- Version tag: tile_group_org_master
- A more comprehensive document including programming model, FPGA emulation and applications (TVM) of manycore.
- BaseJump Manycore Accelerator Network
To install most dependencies, execute the following command:
sudo apt install build-essential gawk texinfo bison flex libgmp-dev libmpfr-dev libmpc-dev libz-dev device-tree-compiler cmake
Check the version of your gcc or g++ by the following command:
gcc --version # or g++ --version
If your gcc/g++ version is 12 or above, you need to downgrade your gcc/g++ or install older version and set alternative version, you may follow this link for more information.
NOTE: If you do not have bsg_cadenv, then you will have to add IGNORE_CADENV=1 to your make commands. To use commerical tools in this case, you may need to set some platform-dependent VCS or Xcelium variables. If you do not have access to commerical CAD tools, then we suggest using the free and open-source Verilator flow described in the section below.
In a scratch directory:
git clone bsg_manycore
git clone basejump_stl
make -C basejump_stl/imports DRAMSim3
# If a BSG group member
git clone bsg_cadenv
This should result in your directory looking like the following:
bsg_manycore/
basejump_stl/
bsg_cadenv/ (if BSG member)
In bsg_manycore:
make checkout_submodules
: To update all submodules inimports/
.make tools
: To install software toolchain required running programs on BSG Manycore. (This build uses 12-16 threads by default.)make machines
: Compile simulation executables inmachines/
.- Edit
BSG_MACHINE_PATH
insoftware/mk/Makefile.paths
to choose the machine to run spmd programs on. - go into
software/spmd/bsg_barrier
and typemake
to run a test!
BSG Manycore has preliminary support for simulating with the open-source Verilator toolchain!
To test this feature, set BSG_PLATFORM=verilator in machines/platform.mk and then follow the above instructions to run tests normally. This platform only currently supports the machine pod_1x1_4X2Y due to excessive compilation times for larger machines. Most likely, future work can enable larger machines with hierarchical Verilation. verilator must be on your path (or override the VERILATOR variable in machines/Makefile.verilator).
On CentOS, you may need to use a modern GCC installation with scl enable devtoolset-8 -- bash
or
by putting source scl_source enable devtoolset-8
in your .bashrc.
BSG Manycore has preliminary support for parsing with the open-source Surelog toolchain!
To test this feature, run make -C machines parse
. Parse-only module is supported, which verifies
that each file in bsg_manycore can be parsed as an individual compilation unit, in order to provide
the greatest tool compatibility. Future work will enable full UHDM generation.
If you're developing on a branch called mybranch
, please pull a branch called ci_mybranch
based
on mybranch
to run CI and mybranch
. It's advised to keep working on mybranch
for incremental
updates and rebase ci_mybranch
on mybranch
when it's ready for another CI run.
Comming Soon!