iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Tenstorrent’s Grayskull e75 and e150 are PCIe inference accelerator cards for 64-bit x86 systems. Their Tensix cores combine RISC-V-based process cores with specialized tensor, SIMD, and network/compression hardware; the cards are not standalone RISC-V computers. The e150 offers higher published throughput, but uses more power, occupies two slots, and needs additional power and cooling provisions.
What Tenstorrent launched
The Grayskull e75 and e150 are inference-only PCIe cards intended to run in a host computer. Both support Tenstorrent’s TT-Buda and TT-Metalium software stacks. Each Tensix core combines five process cores built around the open RISC-V instruction-set architecture with tensor, SIMD, and network/compression hardware.
That makes “RISC-V-powered” an incomplete description if it suggests a general-purpose RISC-V PC: the cards are accelerators that rely on a separate x86_64 host, with RISC-V-based cores as part of their compute architecture.
e75 vs. e150: published specifications
The following figures are from Tenstorrent’s published specifications. Throughput figures are stated as published; no workload-specific performance comparison is established by these specifications alone.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
| Specification | Grayskull e75 | Grayskull e150 |
|---|---|---|
| Tensix cores | 96 | 120 |
| AI clock | 1 GHz | 1.2 GHz |
| On-chip SRAM | 96 MB | 120 MB |
| External memory | 8 GB LPDDR4 | 8 GB LPDDR4 |
| Memory bandwidth | 102 GB/s | 118 GB/s |
| FP8 throughput | 221 teraFLOPs | 332 teraFLOPs |
| FP16/BFP8 throughput | 55 teraFLOPs | 83 teraFLOPs |
| Total board power | 75 W | 200 W |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 |
| Board width | Single-slot | Dual-slot |
| Cooling | Active blower included | Passive; active cooling kit needed when airflow is insufficient |
| Power connectors | One PCIe 6-pin | One PCIe 6+2-pin and one PCIe 6-pin |
Which card makes more sense for inference?
Choose the e75 when fit and power are priorities
The e75 has fewer Tensix cores and lower published throughput than the e150, but its 75 W total board power, single-slot width, and included active blower make it the less demanding installation. It still requires a PCIe 6-pin power connector and a compatible host.
Choose the e150 when you need its additional compute resources
The e150 has more Tensix cores, a higher AI clock, more on-chip SRAM, greater memory bandwidth, and higher published FP8 and FP16/BFP8 throughput. It retains the same 8 GB of external LPDDR4 as the e75, so it does not provide more external-memory capacity for model fit. Its 200 W total board power, dual-slot width, two power connectors, and passive cooling make host clearance, power delivery, and airflow important constraints.
Rank #2
- Designed exclusively for Coral M.2 Accelerator with Dual Edge TPU modules to maximize AI inference performance.
- Fits standard M.2 2280 B-key or M-key slots (PCIe protocol only - not compatible with SATA M.2).
- Bidirectional Gen2 bandwidth: Upstream: ×1 PCIe Gen2 (5Gbps) Downstream: Dual ×1 PCIe Gen2 lanes
- Includes stainless steel mounting screw for vibration-resistant PCB fixation.
- Explicitly incompatible with Raspberry Pi CM4/USB enclosures - prevents buyer errors.
Peak published throughput is not enough to predict which card will be faster for a particular model. The figures do not establish end-to-end performance for a specific model, precision, software version, or workload.
Free tools Windows power users keep installed
One-click scans. No signup required.
Check the host, power, and cooling before installation
Tenstorrent lists the following host requirements for both cards:
Rank #3
- 900-2G193-0000-000
- An x86_64 host and a PCIe 4.0 x16 slot.
- 64 GB of system RAM and at least 100 GB of storage; 2 TB is recommended.
- Ubuntu 20.04 and an internet connection for driver and software-stack installation.
Match the power supply to the specific card’s connectors, not just its stated board-power figure. The e75 needs one PCIe 6-pin connector. The e150 needs one 6+2-pin and one 6-pin connector. Tenstorrent warns that the e150’s passive cooling may require an active cooling kit in systems without sufficient forced airflow; inadequate cooling can reduce performance and risk card damage. Check that the case has room for a dual-slot card and can provide the airflow the e150 requires.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.TT-Buda or TT-Metalium: which software route?
TT-Buda for existing model workflows
Tenstorrent presents TT-Buda as the higher-level path for running existing PyTorch and TensorFlow models. It is the more natural starting point for users who want to bring a model into the supported stack rather than develop directly against lower-level hardware controls.
Rank #4
- High-Performance ML Accelerator: Integrates Edge TPU, delivering 4 TOPS (int8) peak performance for machine learning inference tasks.
- Strong Compatibility: Supports M.2 A+E key interface for easy integration into existing systems.
- Low Power Design: Provides 2 TOPS per watt, ideal for embedded and energy-efficient applications.
- Wide OS Support: Compatible with Linux (Debian 10/Ubuntu 16.04+) and Windows 10 (64-bit).
- Industrial-Grade Reliability: Operating temperature range of -20°C to +85°C, suitable for harsh environments.
TT-Metalium for lower-level development
TT-Metalium is aimed at developers building or tuning workloads and experimenting beyond standard machine-learning flows. It offers a lower-level route than TT-Buda, so it is more relevant when workload control and optimization matter than a higher-level model-execution path.
Recommended Free Tools
Launch pricing and availability
At launch, the e75 was listed at $599 with limited-availability wording, and the e150 at $799 through Tenstorrent’s store. These are historical launch prices, not confirmation of current pricing or stock; check the store for current availability before making a purchase.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
