Explaining a Non-NVIDIA Accelerator Tool | Generated by AI

Home 2026.05 English 中文

---

Question: What is this card shown in the nvidia-smi-like output, and how should it be explained?

Answer:

What you’re looking at is not an NVIDIA GPU, even though the interface looks similar to nvidia-smi. This is a different accelerator ecosystem that mimics NVIDIA tooling for familiarity.

Let’s break it down clearly.

---

1. What is “PPU” here?

• PPU = Processing Unit (custom accelerator)
• It’s analogous to a GPU, but not from NVIDIA
• The tool ppu-smi (shown as PPU-SMI) is a drop-in replacement for nvidia-smi

👉 Think of it like:

• NVIDIA → GPU + nvidia-smi
• This system → PPU + ppu-smi

---

2. The specific card: PPU-ZW810E

This is the model of the accelerator:

• PPU-ZW810E
• Likely a datacenter AI training/inference chip

• Comparable to:

  • NVIDIA A100 / H100
  • or Huawei Ascend 910 (very similar ecosystem style)

---

3. Key specs from your screenshot

Memory

2MiB / 97920MiB

• Total memory ≈ 97 GB
• That’s very large → clearly AI training-class hardware

👉 Comparable to:

• A100: 40GB / 80GB
• H100: 80GB
• This: ~97GB

---

Power

62W / 400W

• Max power: 400W
• Current usage: 62W (idle)

---

Utilization

PPU-Util: 0%

• No workload running

---

Temperature

24°C

• Idle / cool

---

Processes

No running processes found

• Nothing is using the accelerator right now

---

4. What is “HGGC”?

HGGC Version: 13.0

This is the software stack / runtime, similar to:

• NVIDIA → CUDA
• This system → HGGC (custom compute platform)

👉 It likely includes:

• compiler
• runtime
• distributed training support

---

5. Why does it look like nvidia-smi?

Because many AI platforms intentionally copy NVIDIA’s interface:

• same columns:

  • memory usage
  • utilization
  • power

• same workflow:

  • check status
  • monitor jobs

👉 This lowers the learning curve for engineers.

---

6. What system is this likely from?

The “PAI DSW” in your terminal suggests:

• PAI = Platform for AI (Alibaba Cloud)
• DSW = Data Science Workshop

So this is probably:

👉 A cloud AI instance using non-NVIDIA accelerators

---

7. Simple mental model

You can think of this as:

PPU ≈ GPU
PPU-SMI ≈ nvidia-smi
HGGC ≈ CUDA
ZW810E ≈ A100/H100-class chip

---

8. Why 97GB memory matters

This is important for LLM work:

• Large models (like 70B+) need huge VRAM

• 97GB allows:

  • bigger batch sizes
  • larger context
  • fewer devices needed

---

9. Bottom line

This is:

👉 A high-end AI accelerator (non-NVIDIA) 👉 With ~97GB VRAM 👉 Running on a CUDA-like ecosystem (HGGC) 👉 Managed using a nvidia-smi-style tool

---

If you want, I can help you:

• compare this card vs A100 / H100 in real performance
• or show how to actually run LLMs on this PPU system
• or explain why OpenClaw might be consuming so many tokens on it

---

Back Donate