Nvidia Tesla M60 GPU Overview

Nvidia Tesla M60 GPU Overview | Generated by AI

Home 2026.07

Question: Intro to NVIDIA Tesla M60 16GB GPU (used card, ¥288)

Answer:

This is NVIDIA Tesla M60, a data-center GPU from the Maxwell generation (2015). It is a special card because it is basically two GPUs on one PCB.

Basic specs

Item	Tesla M60
Architecture	Maxwell (GM204)
Release	2015
GPU cores	2 × GM204 GPUs
VRAM	16GB total (2 × 8GB GDDR5)
Memory bus	2 × 256-bit
PCIe	PCIe 3.0 x16
TDP	~300W
FP32 compute	~9 TFLOPS
Form factor	Dual-slot, passive server card originally

Important: 16GB does NOT mean one GPU has 16GB.

It is:

Tesla M60
 ├── GPU #1
 │    └── 8GB VRAM
 │
 └── GPU #2
      └── 8GB VRAM

The two GPUs do not combine memory automatically.

What was Tesla M60 designed for?

Main purpose:

VDI (Virtual Desktop Infrastructure)
Cloud gaming
Virtual workstation
NVIDIA GRID

For example:

One M60
 |
 +-- VM 1 → GPU slice
 |
 +-- VM 2 → GPU slice
 |
 +-- VM 3 → GPU slice

It was popular in cloud providers around 2015-2020.

AI / LLM usage today

For AI training/inference:

Good

✅ CUDA works ✅ PyTorch can run ✅ 16GB total VRAM (if using both GPUs separately) ✅ Very cheap

Bad

❌ Old Maxwell architecture ❌ No Tensor Cores ❌ No BF16 ❌ No FP16 acceleration like modern RTX cards ❌ Very slow compared with RTX 3060/3090/4070

Performance comparison:

GPU	FP32
Tesla M60	~9 TFLOPS
RTX 3060 12GB	~13 TFLOPS
RTX 4070	~29 TFLOPS
RTX 3090	~36 TFLOPS

For LLM:

M60 16GB
≈ two slow 8GB GPUs

RTX 3090 24GB
≈ modern AI workstation card

A single RTX 3090 is massively faster.

Compatibility issues

This card has some traps.

1. Power

It needs:

PCIe slot
+
2 × 8-pin power

Your listing says it includes power cable, so likely OK.

2. Cooling

Many M60 cards are passive:

server airflow required

They expect a high-pressure server fan.

A normal desktop case may overheat.

The seller includes a fan, which is important.

3. Driver support

Modern NVIDIA drivers still support Maxwell, but:

CUDA 12 dropped some old architectures
newer PyTorch builds may not support it well

You may need:

CUDA 11.x
PyTorch older versions

At ¥288 CNY, is it worth it?

Depends on purpose.

Good buy

✅ CUDA experiment ✅ learn multi-GPU programming ✅ cheap GPU cluster experiments ✅ CUDA kernel development ✅ old workstation

Not recommended

❌ LLM training ❌ fine-tuning Qwen/Llama models ❌ replacing RTX 4070 ❌ building AI workstation

For comparison:

Tesla M60 ¥288
↓
16GB VRAM

RTX 3060 12GB used ~¥1200-1500
↓
much faster

RTX 3090 24GB used ~¥2500-3500
↓
AI monster

For your AI experiments (nanoGPT, Qwen, agents), I would rank:

RTX 3090 24GB
    >
RTX 4090
    >
RTX 4070 12GB
    >
RTX 3060 12GB
    >
Tesla M60

The M60 is more like a GPU museum piece / CUDA playground, not an AI training card.

References:

NVIDIA Tesla M60 product overview: https://www.nvidia.com/en-us/data-center/tesla-m60/
NVIDIA Tesla M60 datasheet: https://images.nvidia.com/content/pdf/tesla/184457-Tesla-M60-datasheet.pdf

Back Donate