Skip to main content

Tenstorrent Blackhole (RISC-V 2nd Gen, 2024)

Product Overview

Tenstorrent Blackhole is Tenstorrent's second-generation RISC-V AI chip, launched H1 2024, 6nm process (TSMC), 120 Tensix cores (1.5× Wormhole's 80 cores), 8 GB SRAM (one of the largest on-chip SRAMs in the industry), 16 BF16 PFLOPS cluster compute (8-card interconnect). The architecture is led by legendary CPU designer Jim Keller (former Apple A14/M1, AMD Zen, Tenstorrent CTO).

Key positioning: Tenstorrent is the only AI startup with RISC-V + custom ASIC + top-tier architect combined. Blackhole is the company's 2024 flagship, customers: Foxconn (manufacturing), LG AI Research, RIKEN (Japan), Bosch, Mercedes-Benz.

Core Specifications

ItemSpec
ArchitectureTenstorrent Tensix + RISC-V
ProcessTSMC 6nm (vs. Wormhole 12nm)
Tensix Cores120 (vs. Wormhole 80)
Per Tensix Core5 small RISC-V cores + 1 NoC + 1 MB SRAM
On-chip SRAM8 GB (120 Tensix × 1 MB = 120 MB, likely with shared L2)
LPDDR4X24 GB (possibly 16 GB / 32 GB variants)
Memory Bandwidth307 GB/s (LPDDR4X)
BF161.2 PFLOPS per card (8 cards = 16 PFLOPS)
INT84.8 POPS per card (speculative)
TDP~300 W
Form FactorPCIe Gen5 ×16
InterconnectEthernet (standard, open)
ProductionH2 2024
Unit Price~$1,500–3,000

Tensix Core Architecture

DimensionSpec
Per Core5× RISC-V Baby + 1× NoC core
Baby RISC-V1× 32-bit scalar + 2× 32-bit SIMD + 2× 32-bit DSP
NoC CoreNetwork-on-Chip scheduling
SRAM per Core1 MB
BF16 per Core~10 TFLOPS
BF16 per Card120 × 10 = 1.2 PFLOPS
Inside a Tensix core:
┌──────────────────────────────────────┐
│ RISC-V 1 │ RISC-V 2 │ RISC-V 3 │ 32-bit scalar/SIMD/DSP
│ │ │ │
│ RISC-V 4 │ RISC-V 5 │ NoC core │
│ │ │ │
│ 1 MB SRAM (dual-port) │
│ Tensix Compute: Matrix + Vector + Scalar │
└──────────────────────────────────────┘

Tensix innovation: 5 RISC-V + 1 NoC sharing 1 MB SRAM, more flexible than traditional GPU's "CUDA core + Tensor core + shared memory" model, can independently run 6 RISC-V threads (vs. GPU's 32 threads/warp).

8 GB On-chip SRAM Advantage

DimensionAdvantage
LLM Inference7B FP16 = 14 GB slightly exceeds; 7B INT4 = 3.5 GB fits entirely in SRAM
13B LLMINT4 = 6.5 GB fits entirely in SRAM
70B LLMINT4 = 35 GB requires LPDDR4X, but weights loaded to SRAM eliminate HBM wait
LatencyMatrix multiply inside SRAM, latency < 1 ms
Power EfficiencySRAM uses 10× less power than HBM

Blackhole 8 GB SRAM advantage: 7B/13B LLM fully on-chip inference, less power-hungry than H100's 80 GB, with higher energy efficiency (pending benchmarks).

8-Card Cluster 16 PFLOPS

ItemConfig
Blackhole Cards8
BF16 Compute16 PFLOPS (8 × 1.2 PF)
Total SRAM64 GB
Total LPDDR4X192 GB
InterconnectStandard Ethernet (100G/200G open)
Rack TDP~2.4 kW
Rack Price~$30K

Tenstorrent 4U Server (standardized):

  • Grayskull (1st gen) — 2021, 120 W, 32 GB
  • Wormhole (2nd gen) — 2023, 200 W, 80 Tensix
  • Blackhole (3rd gen) — 2024, 300 W, 120 Tensix

Comparison with NVIDIA H100

MetricTenstorrent Blackhole 8-cardNVIDIA H100 singleDifference
BF16 Compute16 PF1.5 PF (FP8 sparse)Blackhole 10×
TDP2400 W700 WBlackhole 3.4×
Energy Efficiency6.7 TOPS/W2.16 TOPS/WBlackhole 3×
Memory192 GB LPDDR4X80 GB HBM3Blackhole 2.4×
Bandwidth2.5 TB/s3.35 TB/sH100 1.3×
SoftwareTT-MetaliumCUDAH100 more mature
Price~$30K (8 cards)~$25–30KComparable

Blackhole 8-card cluster vs. H100 single: 10× cluster compute, 3.4× power (3× energy efficiency), ideal for ultra-large LLM inference (Llama 3 405B split across 8 cards).

Manufacturer Info

ItemDetail
CompanyTenstorrent Inc.
CTOJim Keller (former Apple A14/M1, AMD Zen, Intel, Tesla AI chip)
CEOLjubisa Bajic (former AMD)
Founded2016
HQSan Jose, CA + Toronto, Canada
Funding$700M+ (Series B Q1 2024, led by Bezos Expeditions + Samsung Securities)
Valuation (2025)$3B+ (unicorn)
2024 Revenue~$80M
Employees~500
FabTSMC 6nm + Samsung 4nm (roadmap)
Strategic PartnersFoxconn (manufacturing + servers), LG AI Research, RIKEN Japan, Bosch, Mercedes-Benz, LG, Rapidus (Japan 2nm)
StatusPrivate (considering 2026–2027 IPO)

Tenstorrent Product Line

ProductReleasedTensix CoresProcessBF16Customers
Grayskull202112012nm368 TFEarly customers
Wormhole20238012nm600 TFLG / RIKEN
BlackholeH1 20241206nm1.2 PFFoxconn / LG / Bosch
Quasar (speculative)H2 20252005nm3 PFRoadmap
Grendel (speculative)20262564nm6 PFLong-term

Software Stack: TT-Metalium / TT-Forge

LayerToolDescription
AI FrameworkTT-ForgePyTorch 1:1 compatible (auto-maps to Tensix)
TT-MetaliumLow-level C++ programming (direct Tensix core control)
JAX / TensorFlowCompatible (experimental)
CompilerTT-Forge CompilerModel → Tensix binary
RuntimeTT-RuntimeMulti-card orchestration (standard Ethernet)
Open SourceFully open-source (GitHub 10K+ stars)Opposite of CUDA proprietary

Tenstorrent's killer feature: fully open-source software stack (vs. CUDA 18-year proprietary), 6 RISC-V threads / Tensix (vs. GPU 32 threads/warp black box), standard Ethernet interconnect (vs. NVLink proprietary).

Use Cases

  • RISC-V software ecosystem (fully open-source + heterogeneous RISC-V)
  • Large enterprise LLM inference (Jim Keller brand)
  • Automotive AI (Bosch, Mercedes-Benz customers)
  • Government / National Lab HPC (RIKEN Japan, LG Korea)
  • Manufacturing customers (Foxconn production line deployment)
  • Budget-sensitive (~$1,500/card, far below H100's $25K)
  • AI training focus (Blackhole has weak training ecosystem)
  • CUDA-proprietary workloads (requires TT-Forge porting)
  • Latency-critical (HBM bandwidth advantage)

Key Features

  • 120 Tensix cores + 5 RISC-V/core: the highest RISC-V core count in the industry (600 RISC-V cores)
  • 8 GB SRAM: one of the largest on-chip SRAMs among AI chips
  • Fully open-source software: vs. proprietary CUDA
  • Standard Ethernet interconnect: vs. proprietary NVLink
  • Jim Keller architecture: legendary designer (Apple A14, AMD Zen)
  • Drawbacks: slow LPDDR4X, weak training ecosystem, only 1 year in production

Jim Keller Career Timeline

CompanyRoleContribution
DEC AlphaArchitectAlpha 21264
AMD K8Chief ArchitectAthlon 64
AMD K8/K10Lead ArchitectBarcelona
AppleChip ArchitectApple A4/A5
AMD ZenLead ArchitectZen / Zen 2 (Ryzen 1000–3000)
TeslaVP of HardwareCustom AI chip (unreleased)
IntelSenior VPBrief tenure
TenstorrentCTOGrayskull / Wormhole / Blackhole

Jim Keller at Tenstorrent is the technical + brand core of the company; every chip generation is led by his design.

Big Four US AI Chip Startups

CompanyArchitecture2024 FlagshipFundingStatus
SambaNovaDataflowSN40L$1.1B+Commercialization leader
CerebrasWafer-scaleWSE-3$1.5B+2026 IPO
GroqLPULPU v2$1B+2026 NVIDIA acquisition
TenstorrentRISC-VBlackhole$700M+2026–2027 IPO