T Tenstorrent Last verified

Tenstorrent Wormhole n300

PCIE In production Released 2024 tenstorrent-wormhole-gen2
BF16
TFLOP/s
233 厂商声称
FP8
TFLOP/s
466 厂商声称
FP4
TFLOP/s
unsupported
Memory
GB
24 厂商声称
Mem BW
GB/s
576 厂商声称
TDP
W
300 厂商声称

Full specs

Compute

FP4 TFLOPS
unsupported
FP8 TFLOPS
466
BF16 TFLOPS
233
FP16 TFLOPS
233
INT8 TOPS
466

Memory

Capacity
24 GB
Bandwidth
576 GB/s
Type
GDDR6

Die architecture 🟢 vendor floorplan

Tensix count
128
Process
12 nm
PCIe
Gen 4 ×16

Scale-Up (intra-node)

Protocol
Galaxy-Mesh
Per-link BW
800 GB/s
World size
32
Topology
2D-torus
Switch
—

Scale-Out (inter-node)

Per-card NIC
200 Gbps
Protocol
Ethernet
NIC
—

Topology

拓扑结构 · Topology
32 卡 scale-up domain
芯片内部 / Die-level architecture
GDDR6 24 GB @ 0.6 TB/s Tenstorrent Wormhole n300 L2 / shared cache · NoC L1$ / register file (per Tensix) 128 Tensixs · darker block = tensor / matrix engine 233 TFLOPS BF16 · 466 FP8 · 24 GB GDDR6 @ 0.6 TB/s · 300 W TDP

⚠ illustrative / 示意性版图: compute-unit and HBM-stack count are inferred from public BF16 / memory specs. architecture field not populated for this card yet. Contribute floorplan data →

集群拓扑 / Cluster topology · Galaxy-Mesh @ 800 GB/s
ToR · Galaxy-Mesh Node 1 Node 2 Node 3 Node 4 4 节点 × 8 卡 = 32 卡 · 节点内 800 GB/s · 节点间 200 Gbps RoCE/IB
Scale-Up · 域内
Galaxy-Mesh
800 GB/s · 拓扑: 2D-torus
world_size = 32
Scale-Out · 跨域
Ethernet
200 Gbps/卡 NIC

Which models can it run?

Quick estimates · decode tok/s/card 上界

TP=8 · FP8 · batch=16 · prefill=1024 · decode=256 · 已应用 efficiency 校准

在计算器中调整 →
模型 参数 (active) Decode tok/s/card 瓶颈
DeepSeek V4 Pro
deepseek
 49B — 显存不足
DeepSeek V4 Flash
deepseek
 13B — 显存不足
Mistral Small 4
mistral
 22B 7 内存带宽
GLM-5 Reasoning
zhipu
 32B 6 内存带宽
GLM-5.1
zhipu
 32B — 显存不足
Qwen3.6 Plus
alibaba
 35B — 显存不足
Kimi K2.6
moonshot
 32B — 显存不足
MiniMax M2.7
minimax
 46B — 显存不足

Operator-level fit · per-model bottleneck + upper bound

算子级 fit · operator-level fit (per-token roofline)

基于每个模型 operator_decomposition + 本卡 BF16 233 TFLOPS / 576 GB/s 计算 · ridge point ≈ 405 FLOPs/byte

上界 = min(计算屋顶, 内存带宽屋顶) · efficiency 未应用
模型 domain 主导算子 AI · F/B 瓶颈 tok/s 上界
DeepSeek V4 Pro llm matmul 245.5 💾 内存带宽 24k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 1063
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 319
GPT-OSS llm matmul 0.7 💾 内存带宽 47
Gemma 4 26B llm matmul 0.7 💾 内存带宽 35
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 33
Mistral Small 4 llm matmul 0.6 💾 内存带宽 15
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 15
需要 efficiency 校准 + concurrency 扫描 + TCO 估算 → 在计算器中评估 →

Operator support & optimization headroom

算子支持 & 优化空间 / Operator support & headroom

Per-operator support derived from software_support.engines + scale-up topology. Optimization headroom from measured efficiency factor.

Optimization headroom
+50 pp
moderate

No cases yet — using default 0.5 efficiency. Real headroom unknown until first measurement lands.

Communication (collective)
All-to-All 🟢 mature
all-to-all via Galaxy-Mesh world_size=32
AllReduce 🟢 mature
Galaxy-Mesh ring all-reduce
Attention
Multi-Head Attention 🟢 mature
paged-attention via vLLM/SGLang/MindIE
FlashAttention-3 🟢 mature
FA-3 on modern engine + tensor cores
Matrix multiply (GEMM)
Matrix Multiplication 🟢 mature
GEMM supported on all inference engines
MoE routing
MoE Routing 🟢 mature
MoE gating supported via vLLM ≥0.4 / SGLang
Normalization
RMSNorm 🟢 mature
fused into engine kernels
Embedding
Rotary Position Embedding 🟢 mature
fused into engine kernels
Activation
SiLU / Swish 🟢 mature
fused into engine kernels
Softmax 🟢 mature
fused into engine kernels

最接近的替代卡 (按规格相似度)

基于 BF16 算力 / 显存 / 显存带宽 / FP8 加权欧氏距离。供选型决策参考。

AWS Inferentia 2
AWS Annapurna Labs
100%
相似度
BF16
190
Mem
32
BW
0.8
壁仞 BR104
壁仞科技
100%
相似度
BF16
128
Mem
32
BW
1.1
⇄ 国产 替代
天数智芯 天垓 100
天数智芯
100%
相似度
BF16
95
Mem
32
BW
1.2
⇄ 国产 替代
平头哥 含光 800
平头哥
100%
相似度
BF16
—
Mem
16
BW
0.3
⇄ 国产 替代

Software-stack support

Engine Status BF16FP16FP4FP8 E4M3FP8 E5M2INT4 AWQ
HanGuangAI unconfirmed — — — — — —
LMDeploy unconfirmed — — — — — —
MindIE unconfirmed — — — — — —
MoRI unconfirmed — — — — — —
SGLang unconfirmed — — — — — —
TensorRT-LLM (Dynamo) unconfirmed — — — — — —
vLLM community ✓ ✓ — ✓ — —

Existing deployment cases (0)

No measured cases yet for this card. Be the first contributor?

Citations

  1. [1] Tenstorrent Wormhole n300 — 128 Tensix cores, 24GB GDDR6, 12nm Global Foundries, RISC-V based, fully open-source software stack (TT-Metal/TT-NN). Galaxy interconnect for 32-card 2D-torus topology. — https://tenstorrent.com/hardware/wormhole · accessed 2026-04-29 厂商声称
⚠ Tenstorrent emphasizes open-source software and RISC-V instruction set — distinct from CUDA/ROCm/CANN walled gardens.
⚠ Tile-based architecture: each Tensix core has 5 RISC-V CPUs + 1 matrix engine + 32 KB compute SRAM.