Performance Analysis of Kunlun Chip Product Iterations

According to the latest information, the performance improvement of Kunlun’s next-generation chips is quite significant:

• Process Architecture
  : XPU-R architecture, based on 7nm process
• Computing Power Indicators
  : INT8 computing power of 256 TOPS, FP16 computing power of 128 TFLOPS
• Performance Improvement
  :
  2-3x
  improvement over the first-generation product [1]

• Process Architecture
  : Self-developed XPU-P architecture
• Computing Power Indicators
  : FP16 computing power reaches
  345 TFLOPS
• Function Upgrade
  : Supports 10,000-card cluster deployment
• Actual Performance
  : Throughput reaches
  2437 tokens/s
  in a single-machine 8-card configuration [1]

• M100
  : Optimized for large-scale inference scenarios, scheduled to launch in early 2026, with significant performance improvements in MoE model inference [2]
• M300
  : For ultra-large-scale multi-modal model training and inference, scheduled to launch in early 2027 [2]

Baidu’s concurrently released Tianchi Super Node series also shows significant performance improvements:

• Tianchi 256 Super Node
  (First half of 2026): Total inter-card interconnect bandwidth increased by
  4x
  , performance improved by
  over 50%
  , and single-card token throughput for mainstream large model inference tasks increased by
  3.5x
  [2]
• Tianchi 512 Super Node
  (Second half of 2026): Compared to the Tianchi 256 Super Node, total inter-card interconnect bandwidth increased by another
  1x
  , and a single node can complete training of trillion-parameter models [2]

From Kunlun Generation 1 to Generation 2, there was a

performance leap, and by the P800, FP16 computing power reached an advanced level of 345 TFLOPS. The upcoming M100 and M300 will further enhance inference and training capabilities. Combined with the new generation of super node products, inter-card interconnect bandwidth and overall system performance will achieve several-fold improvements.

---

[1] ESM China - “Kunlun Chip Spin-off Listing? Baidu’s Latest Response” (https://www.esmchina.com/news/13732.html)
[2] EETimes China - “Behind Baidu’s Response to ‘Kunlun Chip Listing’: Traditional Business Under Pressure, Need for New Stories” (https://www.eet-china.com/mp/a458286.html)