wu-arch/chipyard
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Compare commits

merge into: wu-arch:pre-wu-blackwell
Branches Tags
wu-arch:wu-architecture wu-arch:pre-wu-blackwell wu-arch:wu-blackwell-fp8-e4m3 wu-arch:wu-tmem-bank-scalar-softmax wu-arch:nvidia-style-arch wu-arch:feature/scalar-tmem-softmax-vector wu-arch:wu-blackwell-4lane-tc wu-arch:asplos-ae
...
pull from: wu-arch:nvidia-style-arch
Branches Tags
wu-arch:pre-wu-blackwell wu-arch:wu-blackwell-fp8-e4m3 wu-arch:wu-tmem-bank-scalar-softmax wu-arch:nvidia-style-arch wu-arch:feature/scalar-tmem-softmax-vector wu-arch:wu-blackwell-4lane-tc wu-arch:wu-architecture wu-arch:asplos-ae

3 Commits
pre-wu-bla ... nvidia-sty

Author SHA1 Message Date
Zhongdi LUO
2f84a144a9 config: reduce Radiance shared memory to 32 KiB 2026-07-13 07:59:42 +00:00
Zhongdi LUO
739edee677 feat: configure 1-core Virgo systolic architecture 2026-07-13 07:48:14 +00:00
Zhongdi LUO
9c9029fb9f feat: add NVIDIA-style 1c4l4w configurations 2026-07-13 07:20:56 +00:00
2 changed files with 13 additions and 11 deletions
Expand all files Collapse all files

22
generators/chipyard/src/main/scala/config/RadianceConfigs.scala
View File

@@ -25,6 +25,8 @@ class WithRadBootROM(address: BigInt = 0x10000, size: Int = 0x10000, hang: BigIn
// aliases for virgo
class VirgoConfig extends RadianceClusterConfig
class VirgoFP16Config extends RadianceFP16ClusterConfig
class VirgoVoltaConfig extends RadianceFP16ClusterConfig
class VirgoAmpereConfig extends RadianceFP16ClusterConfig
class VirgoHopperConfig extends Radiance4CFP16ClusterConfig
class VirgoBlackwellConfig extends RadianceBlackwellClusterConfig
class VirgoFlashConfig extends RadianceClusterConfig
@@ -56,8 +58,8 @@ class RadianceBaseConfig extends Config(
class RadianceFP16ClusterConfig extends Config(
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
new radiance.subsystem.WithRadianceCores(8, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, useVxCache = false) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 16) ++
new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, useVxCache = false) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 16) ++
new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
new radiance.subsystem.WithRadianceCluster(0) ++
@@ -66,7 +68,7 @@ class RadianceFP16ClusterConfig extends Config(
class Radiance8B8WFP16ClusterConfig extends Config(
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
new radiance.subsystem.WithRadianceCores(8, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, useVxCache = false) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 8, numWords = 8) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 8, numWords = 8) ++
new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
new radiance.subsystem.WithRadianceCluster(0) ++
@@ -74,13 +76,13 @@ class Radiance8B8WFP16ClusterConfig extends Config(
class Radiance4CFP16ClusterConfig extends Config(
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
new radiance.subsystem.WithRadianceCores(4, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = true, useVxCache = false) ++
// new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 16,
new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = true, useVxCache = false) ++
// new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 16,
// memType = radiance.subsystem.TwoReadOneWrite,
// serializeUnaligned = radiance.subsystem.CoreSerialized) ++
// NOTE: Hopper Tensor Core does not work with 16-word config due to the
// address alignment requirement
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 8) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 8) ++
new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
new radiance.subsystem.WithRadianceCluster(0) ++
@@ -89,7 +91,7 @@ class Radiance4CFP16ClusterConfig extends Config(
class RadianceBlackwellClusterConfig extends Config(
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, tensorCoreBlackwell = true, startupAddress = BigInt("80000000", 16), useVxCache = false) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 8) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 8) ++
new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
new radiance.subsystem.WithRadianceCluster(0) ++
@@ -98,9 +100,9 @@ class RadianceBlackwellClusterConfig extends Config(
class RadianceClusterConfig extends Config(
// important to keep gemmini tile before RadianceCores to ensure radiance tile id is 0-indexed
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 8, accSizeInKB = 16, tileSize = 8) ++
new radiance.subsystem.WithRadianceCores(4, location = InCluster(0), tensorCoreFP16 = false, tensorCoreDecoupled = false, useVxCache = false) ++
new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = false, tensorCoreDecoupled = false, useVxCache = false) ++
// new radiance.subsystem.WithRadianceFrameBuffer(x"ff018000", 16, 0x8000, x"ff011000", "fb0") ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 256 << 10/*KBytes*/, numBanks = 8, numWords = 8,
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10/*KBytes*/, numBanks = 8, numWords = 8,
// memType = radiance.subsystem.TwoReadOneWrite,
serializeUnaligned = radiance.subsystem.CoreSerialized) ++
new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
@@ -133,7 +135,7 @@ class RadianceBigLittleClusterConfig extends Config(
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 4, accSizeInKB = 16, tileSize = 16) ++
new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 8, accSizeInKB = 16, tileSize = 8) ++
new radiance.subsystem.WithRadianceCores(2, location = InCluster(0), useVxCache = false) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 64 << 10, numBanks = 4, numWords = 8) ++
new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 8) ++
new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
new radiance.subsystem.WithVortexL1Banks(nBanks = 8)++
new radiance.subsystem.WithRadianceCluster(0) ++

2
generators/radiance

Submodule generators/radiance updated: bb1459a209...2d6bf7dd45