Hardware GA100 Silicon

The GA100 die in the CMP 170HX

What this page covers. The physical silicon: process, transistor count, die size and package; the GPC / TPC / SM hierarchy and how many SMs a 170HX actually has versus an A100; the floorsweeping fuses that decide that number and how they vary from die to die; the memory partition and cache hierarchy; compute capability and the instruction set that follows from it; and which fixed-function engines are present, fused off, or physically absent.

The headline: the CMP 170HX carries the same GA100 die as the NVIDIA A100, TSMC 7 nm, 54,200 million transistors, 826 mm². It is a harvested bin of that die, marked GA100-105F-A1 on the 8 GB card and GA100-105A-A1 on the 10 GB card, with 5 of 8 GPCs active: 35 TPC, 70 SM, 4480 FP32 lanes, compute capability 8.0. Both SKUs enumerate exactly 70 SMs. Nothing about the CMP restriction removes SMs, and nothing in the compute unlock adds any back: the mining-specific restriction is an issue-rate divider held in separate fuses, and the SM count you get is the ordinary silicon fuse floor for this bin.

Every 170HX reads PMC_BOOT_0 (0x00000000) = 0x170000a1, which is the GA100 chip-ID signature. A GA10x consumer control reads 0xb74000a1 at the same offset, so this one dword is a sufficient "is this really a GA100" test.

---

Die and package

Property	Value	Notes
Architecture	Ampere GA100	same die as A100, not a CMP-specific tape-out
Process	TSMC 7 nm (N7)
Transistors	54,200 million	65.6 M/mm² density
Die area	826 mm²	at or near the 7 nm reticle limit (~830 mm²)
Package	BGA-2743, roughly 55 × 55 mm	heatsink bolt pattern 57 × 68 mm centre-to-centre
ASIC marking, 8 GB	GA100-105F-A1	GPU part number 20C2-105-A1
ASIC marking, 10 GB	GA100-105A-A1	GPU part number 2082-105-A1
Retail A100 comparison marking	GA100-883AA-A1	photographed alongside a 170HX die
PMC_BOOT_0 0x00000000	0x170000a1	on every valid GA100
PCI device ID	10de:20c2 (8 GB) / 10de:2082 (10 GB)	see board and variants
Compute capability	8.0 (sm_80)	OpenCL 3.0, CUDA 11+
Base / boost SM clock	1140 MHz / 1410 MHz	1470 MHz observed at nvidia-smi -pl 300
TDP / maximum software power limit	250 W / 250 W on stock VBIOS	nvidia-smi -q reports Min 100 W, Max 250 W; 300 W only on cards carrying the NVIDIA OC mining VBIOS. Slot power limit in DevCap is 75 W

The PCB is the A100 40 GB PCIe reference design with parts deliberately deleted, so board-level absences (VRM phases, NVLink interface ICs, PCIe AC-coupling capacitors) are a separate topic from die-level absences. See physical mods and power delivery.

---

GPC / TPC / SM hierarchy

GA100's scale registers describe the full die independently of what is fused on. They read the same on a 170HX and on an A100-class reference part:

Scale register	Address	Value	Meaning
PTOP_SCAL_NUM_GPCS	0x00022430	0x8	8 GPCs on the full die
PTOP_SCAL_NUM_TPC_GPC	0x00022434	0x8	8 TPCs per GPC → 64 TPC → 128 SM full die
PTOP_SCAL_NUM_FBPS	0x00022438	0xc	12 FBPs
PTOP_SCAL_NUM_FBPAS	0x0002243c	0x18	24 FBPAs (a GA102 RTX 3090 reads 6)
PTOP_SCAL_NUM_LTCS	0x00022454	24	24 L2 cache slices
PTOP_SCAL_FBPA_PER_FBP	0x00022458	2	2 FBPAs per FBP
PTOP_SCAL_NUM_NVLINK	0x0002246c	12	12 NVLink units on the die
PTOP_FS_STATUS	0x00022470	0x0000003f	floorsweep status

A full GA100 would therefore be 128 SM / 8192 FP32 lanes. No product ever shipped that configuration: every A100 SKU ships 108 SM / 6912 shading units, and the CMP 170HX ships 70.

Part	GPCs active	TPC	SM	FP32 lanes
Full GA100 die	8	64	128	8192
A100 SXM4 / PCIe (all SKUs)	7 (FUSE_GPC_DISABLE = 0x08 / 0x08 / 0x80, one bit set on each of the three probed SKUs)	54	108	6912
DRIVE A100 (0x20bb, PG199)	6 (FUSE_GPC_DISABLE = 0x84)	48	96	6144
CMP 170HX, both SKUs	5 (RING_ENUM_GPC = 5)	35	70	4480

The 170HX figure is measured, not inferred. A PTX special-register dumper on an unlocked card reports SMs=70 warpsize=32, nsmid=70 nwarpid=64, with smid values spanning 0..69 with no gaps. An 8-card mixed rig enumerated cu: 70 on all eight devices while memory alternated between 9990 MiB and 7954 MiB, which is the cleanest single demonstration that SM count does not track SKU. OpenCL-Benchmark on an unlocked card likewise reports 70 compute units at 1410 MHz (4480 cores, 12.634 TFLOPs/s theoretical FP32).

Note that 35 TPC is an odd number: the five active GPCs do not all carry 8 TPCs, which is normal for a harvested part.

Note

Superseded

Two widely repeated SM figures are wrong. deviceQuery prints Total SPs: 8960 (70 MPs x 128 SPs/MP) and Compute throughput: 25267.20 GFlops; that is the tool applying the compute-capability 8.6 figure of 128 FP32 lanes per SM instead of GA100's (cc 8.0) 64. The arithmetic settles it: 4480 × 2 × 1410 MHz = 12.63 TFLOPS exactly, and every measured unlocked FP32 result lands at 12.28 to 12.99 TFLOPS, not ~25. Separately, one published specification page claims "8 GB variant: 56 SMs". The register accounting on an 8 GB card (OPT_GPC_DISABLE = 0x45, RING_ENUM_GPC = 5, 35 active TPC) gives 70 SM, the same as the 10 GB card.

---

Floorsweeping and per-die binning

Floorsweeping is the disabling of defective or surplus units by blowing OTP fuses at test time. On the 170HX it works entirely through the fuse and STATUS path: every CTRL_OPT override register reads zero, and the topology report's own summary line is held back by CTRL_OPT: 0 TPC = 0 SM. That means the 70 SM figure already is the fuse floor for these dies, and there is no software override layer sitting on top of it that could be relaxed.

Register	Address	Role	170HX reading
OPT_GPC_DISABLE	0x00820350	GPC disable mask (fuse)	per-die, 3 bits set
STATUS_OPT_GPC	0x00820c1c	effective GPC mask	always mirrors the fuse
OPT_GPC_DEFECTIVE	0x008205c4	which GPCs are genuinely bad	0x00 on some cards
RING_ENUM_GPC	0x00120078	GPCs enumerated on the ring	5
CTRL_OPT_GPC	0x0082081c	software floorsweep override	0x00000000
CTRL_OPT_FBIO / _FBPA / _FBP	0x00820814 / 0x00820818 / 0x00820938	as above, memory side	0x00000000
CTRL_OPT_PERLINK / _PCIE_LANE / _NVLINK	0x00820820 / 0x0082082c / 0x008209b8	as above	0x00000000
FUSE_EN_SW_OVERRIDE	0x00820040	enables the CTRL_OPT layer at all	0x00000000
gpcMask	0x00408970	GR-side GPC mask	0xdc, re-asserts if written
OPT_PCIE_DEVIDA	0x008204d8	SKU identity fuse	0x20c2 reported on an 8 GB card (see note below)
OPT_SLT_REV	0x008204bc	slot/test revision	per-die

The mask varies per individual card, not per SKU, and not with driver version. Seven distinct OPT_GPC_DISABLE values are on record across three separate surveys, four of them from four cards read in a single afternoon, all with three GPCs disabled and all totalling 70 SM:

OPT_GPC_DISABLE	GPCs disabled	Card
0x85	0, 2, 7	10 GB
0x45	0, 2, 6	8 GB
0x13	0, 1, 4	8 GB
0xa8	3, 5, 7	10 GB
0xd0	4, 6, 7	fuse-table card
0x23	0, 1, 5	fuse-table card
0x15	0, 2, 4	10 GB

Some of the disabled GPCs are marked physically good. On the 8 GB card used for the high-security write experiments, OPT_GPC_DEFECTIVE read 0x00000000 while OPT_GPC_DISABLE had three bits set, so all three disabled GPCs are healthy silicon fused off to hit a product spec. On one 10 GB card OPT_GPC_DEFECTIVE = 0x81 (GPCs 0 and 7 genuinely defective) while GPC 2 was disabled but not defective.

The restriction / binning split

A full 120-register diff of two physical 170HX cards found 107 registers identical and 13 different, and every one of the 13 is a binning value. This is the single most actionable tooling result on the die:

• Product-line constants, identical on every 170HX, safe to hard-code in a recipe: the nine speed-select fuses (FUSE_SS_DP = 0x1, the other eight = 0x5), FUSE_PCIE_GEN23_DIS = 0x1, FUSE_PCIE_GEN3_DIS = 0x1, FUSE_NVLINK_DIS = 0x7, FBPA_CFG1_BROADCAST = 0x02449000, FUSE_PCIE_DEVIDB = 0x20c2, FUSE_ECC_EN = 0x0, FUSE_EN_SW_OVERRIDE = 0x0. (Both cards in the 120-register diff were 10 GB units, so any register that varies by SKU rather than by die looks like a constant in that diff. Two do: FUSE_SKU_ID (0x00821060) reads 0x68 on a 10 GB card and 0x80 on an 8 GB card, and FUSE_PCIE_DEVIDA (0x008204d8) reads 0x2082 on a 10 GB card and 0x20c2 on an 8 GB card, while FUSE_PCIE_DEVIDB is 0x20c2 on both. Neither DEVIDA nor SKU_ID is safe to hard-code, and lspci -nn remains the simplest SKU test.)
• Per-die values that must never be hard-coded: all floorsweep masks and their STATUS mirrors, FEAT_OVR_SM_SPD (0x0082381c), FEAT_OVR_SM_SPD_1 (0x00823820), FEAT_OVR_QUADRO (0x00823808), I1500_DATA, I1500_SHADOW_WDR, and every per-FBPA readback.

That split is why the compute unlock can ship two fixed magic constants and still be correct on every card, and why any tool that compares your card against "the" stock SS0 value is comparing against noise.

Surveying your own card

The read-only survey tool ga100_topology_report.py (v1 4848 bytes; v2 8128 bytes, adding an InfoROM dump) reads PMC_BOOT_0, OPT_GPC_DISABLE, STATUS_OPT_GPC, OPT_GPC_DEFECTIVE, RING_ENUM_GPC, PTOP_SCAL_NUM_GPCS, PBUS_SW_SCRATCH(1) (0x00001404), 0x00118f78, OPT_PCIE_DEVIDA, OPT_SLT_REV, and the per-GPC OPT_DISABLE / RECONFIG / CTRL_OPT / STATUS / RECONF_OVR set. At least four independent people have run it on both SKUs with self-agreeing output.

Note

Open problem: the 38 missing SMs

Going from 70 SM to the A100's 108, or the die's 128, would be the single largest gain available on this card, and on cards where OPT_GPC_DEFECTIVE = 0 the disabled GPCs are known good silicon. Every write path found so far is latched. FUSE_CTRL_OPT_TPC_GPC is remove-only (an OR-test on an active TPC did not even drop the count); high-security writes to OPT_GPC_DISABLE, STATUS_OPT_GPC, OPT_TPC_GPC2 (0x00820768) and DIS_SW_OVR (0x00820084) all read back unchanged, in an experiment that carried two positive controls proving the write primitive was live; and forcing gpcMask three separate ways (RM struct, host MMIO write to 0x00408970, patching the GSP firmware's andi to li a4,255) made the software stack report 8 GPC / 112 SM while 0x00408970 read back 0xdc every time and cuInit segfaulted. Untried candidates: a GSP-RPC path via the static floorsweeping-mask queries (classes 0x2080122a / 0x2080122b), a GR-shadow write, or porting the write primitive to PMU / GSP / FECS / GPCCS. See dead ends.

Warning

Experimental

One card in the wild was reported CTRL_OPT-swept to 56 SM rather than the fuse floor of 70, and 6 SM were clawed back to reach 62, with the remaining TPCs genuinely failing when enabled. This is described as the first compute-swept card seen, and no before/after register dump was published. Every other surveyed card is already at its fuse floor, where CTRL_OPT costs nothing. Treat a below-70 SM count as rare, not normal.

---

Memory partitions and the cache hierarchy

The framebuffer side is floorswept independently of the graphics side. GA100 has 12 FBPs, each with 2 FBPAs, so 24 FBPAs at 256 bits each. Each 1024-bit HBM interface is really four 256-bit channels, which is why partial-stack enablement is a real hardware state and FBPA masks show partial rather than clean whole-stack failures.

Quantity	8 GB SKU (0x20c2)	10 GB SKU (0x2082)
Active FBPAs	16 (of 24)	20 (of 24)
Active FBPs	8 (of 12)	10 (of 12)
Memory bus width	4096-bit	5120-bit
Stock capacity	8192 MiB	10240 MiB
Unlocked capacity	65536 MiB	40960 MiB

The floorsweep masks again vary per die. Two 10 GB cards read FUSE_FBPA_DISABLE (0x00820368), FUSE_FBIO_DISABLE (0x0082036c), STATUS_FBPA (0x00820c18) and STATUS_OPT_FBIO (0x00820c14) all at 0x0003c000 on one unit (FBPAs 14 to 17 off) and 0x000000c3 on the other (FBPAs 0, 1, 6, 7 off), both leaving 20 active. Disabled FBPAs return a 0xbadf20xx sentinel from their CSTATUS registers rather than a value, and the sentinel tracks which FBPAs are off: 0xbadf2010 and 0xbadf2013 on the card with FBPAs 0, 1, 6, 7 disabled, 0xbadf2017 and 0xbadf2018 on the card with FBPAs 14 to 17 disabled. A dump from an 8 GB card enumerated its 12 FBP half-stack regions individually: FBP 1 and 4 disabled, FBP 6 and 11 defective, the other eight active, giving 8 × 8 GB = 64 GB, which is exactly the capacity the unlock reaches. Full detail is on memory subsystem and memory geometry.

Cache level	Size	Basis
L1 / shared, per SM	192 KB	GA100 architectural figure
L2	32 MB (32768 KB)	CUDA deviceQuery plus an independent latency-spike microbenchmark
L2, full A100	40 MB	for comparison
OpenCL global cache / local memory	1960 KB / 48 KB	OpenCL-Benchmark on an unlocked card

Note

Disputed

TechPowerUp lists the 170HX with 8 MB of L2. The runtime deviceQuery figure of 32768 KB and an independent pointer-chase latency measurement both say 32 MB, and this wiki uses 32 MB. No source in the corpus reconciles the two; a published latency/bandwidth curve showing where the working set falls off a cliff would close it. Note also that the correct TechPowerUp entry is gpu-specs/cmp-170hx-8-gb.c3830; the older c3824 URL now redirects to an AMD card.

HBM bandwidth is not restricted on this part and is unchanged by the unlock (a same-card A/B measured 1592 GB/s stock versus 1599 GB/s modded at a 256 MB working set, a ratio of 1.0x, in the same table where FP32 moved 30.7x). Theoretical peak is 1555.2 GB/s (= 1448.4 GiB/s) from 1215 MHz DDR on 5120-bit. Measured figures span 1305.86 to 1600 GB/s depending on tool and access pattern, and no single canonical number exists; see performance.

---

Compute capability and instruction set

Compute capability 8.0 (sm_80) fixes what the die can and cannot do, independently of any fuse:

• 64 FP32 lanes per SM, 64 warps per SM, warp size 32. FP64 at the GA100 1:2 ratio, non-tensor FP16 at 4:1 versus FP32 (architecturally unusual, and the reason locked cards were already usable for LLM token generation).
• 280 third-generation Tensor Cores (4 per SM), 280 TMUs, 128 ROPs. The tensor cores are present and functional, not fused off: unlocked cards measure 158.7 to 190 TFLOPS FP16 tensor and 164.4 to 192.7 TFLOPS BF16 tensor.
• No FP8 and no NVFP4 hardware path. Those need sm_89+ / sm_120. A tool enumerating supported MMA shapes on this card lists mma_mxf8mxf8f32_16_8_32 and mma_f8f8f16/f32_16_8_32 as unsupported, which is expected for Ampere. INT8, INT4 and INT1 are native in hardware, though INT1 shares the INT8 XNOR-popcount path with no dedicated unit.
• Practical consequence for inference: FP8 KV cache is unsupported on sm_80, so KV must be BF16.

---

Engines present, fused off, and absent

Engine / feature	Status on the 170HX	Evidence
CUDA cores, Tensor Cores	Present, functional, issue-rate throttled at stock	see compute throttle
FP64 units	Present, restored by the unlock	measured at the 1:2 ratio
NVENC	Absent. The GA100 die generally carries no video encoder	reported in-channel and consistent with the die's feature set
NVDEC	Die-level: five instances of 4th-generation NVDEC per the specification database. Driver-level: not exposed on this SKU. A probe of the NVDEC falcon mailbox 0x00830040 returned 0xbadf1100, i.e. blocked or read-only	specification database, medium confidence; falcon probe
Display engine / outputs	Absent. No display outputs of any kind, Slot Width: IGP, no DirectX / Vulkan / OpenGL exposure	specification database and every lspci capture, which names the device a 3D controller
NVLink	Fused off. FUSE_NVLINK_DIS / STATUS_OPT_NVLINK (0x00820db8) = 0x7; the board is also missing its NVLink interface ICs. No NVLink register appears in the 0x00823800-0x0082382c block and no branch contains NVLink code	fuse read plus teardown; see NVLink hardware
ECC	Fused off. FUSE_ECC_EN = 0x0, no telemetry, no known lever. The high ECC-status nibbles of FEAT_READOUT_0 read zero, consistent with the fuse	see ECC frontier
P2P	Absent. No P2P code in any shipping or branch tree; the sole MIG profile reports P2P: No	see P2P frontier
MIG	Hardware supports it (enable bit 0 of 0x00820840), but only one profile, 1g.64gb, is exposed, so the GPU cannot actually be partitioned	community finding, not in shipping code
Resizable BAR	Present but limited to 64 MiB	lspci capability [bb0]
FLR	Present (FLReset+ in DevCap), and load-bearing for every unlock harness	lspci -vvv

Note

Open problem: is NVENC fused off or simply not built?

The in-channel statement is "nvenc is disabled ... idr if it's fused off or if it's fuse gated but it's not available by default if it has the hardware." Nobody has reported an NVENC session working and no unlock path has been proposed. The obvious next step is the same differential method that cracked the compute throttle: read the NVENC-related OPT_*_DISABLE fuses in the 0x00820xxx block on a 170HX and on an A100 or DRIVE A100 control and diff them. Until that is done, assume the card has no hardware encoder.

---

Why the die matters to the unlock

Three properties of this specific silicon are what make the whole project possible:

1. The master kill fuse is unblown. OPT_FEATURE_FUSES_OVERRIDE_DISABLE (0x008203f0) reads 0x00000000. Had it been blown, every feature override would be permanently locked and no software path would exist.
2. The restriction fuses are a product-line constant, not per-die binning. One recipe works on every card.
3. The DRIVE A100 (0x20bb, PG199, GA100-550F-A1) is a clean negative control. Two physical boards share the 170HX's NVLink kill and its EN_SW_OVERRIDE = 0 / DIS_SW_OVR = 1 state, yet read all nine speed-select fuses at 0x00000000, FEATURE_READOUT_1 = 0x00000000 and have full compute at 96 SM. That single-variable isolation is what proved the OPT_SM_SPEED_SELECT fuse block, and nothing else, is the compute restriction.

See fuses and OTP for the full fuse map, and compute throttle for how the override is reached.