Core Technologies

Core Technologies

Our Products Trend

2025

  • Tranxform Anima Architecture
  • 20 TOPS / Core

2026

  • Tranxform Axion Architecture
  • 32 TOPS / Core

Scalable

  • 1W for a Core with 32 TOPS
  • 4 Cores in a Chip
  • 4 Cores with 128 TOPS
  • Power includes SRAM and DRAM
1X Performance

1X Performance

128 TOPS/sec

130 Watts (System)

2X Performance

2X Performance

256 TOPS/sec

160 Watts (System)

4X Performance

4X Performance

512 TOPS/sec

1120 Watts (System)

*System = Host CPU, power supply, DRAM, SRAM etc.

Applications with scalable chips

Accuracy and Error Rate

OPENAI Whisper (FP16) : Word Error Rate (the smaller, the better)

  • Nvidia Blackwell, Intel, ARM
2.8
2.7
2.6
2.5
2.4
2.66 Error Rate
 
Word Error Rate
2.8
2.7
2.6
2.5
2.4
2.51 Error Rate
 
Word Error Rate
  • Tranxform Axion Architecture

GPT-OSS-20B Performance Comparison

  • Nvidia RTX 5090 ($2799)
  • BlackWell Architecture (same as B200)
  • 221 tokens/sec
  • 575 watts
400
300
200
 
218 tokens/sec
 
performance
400
300
200
 
300 tokens/sec
 
performance
  • Tranxform SOC Chip (2)
  • High Dimensional Architecture
  • 320 tokens/sec
  • < 60 watts

GPT-OSS-20B comparison

  • Groq System
  • Cloud Based
  • 1200 tokens/sec
  • 2-4 K watts (8 Boards System)
4000
3000
2000
1000
 
3000 watts
 
power comparison
4000
3000
2000
1000
 
240 watts
 
power comparison
  • Tranxform Chiplets (8 Chips)
  • High Dimensional Architecture
  • 1200 tokens/sec
  • < 240 watts (a Board with 8 chips)

Low Power：Shorten wires to save energy

 

Coper wires on PCB

> 3.97 pJ / bit (HBM2)

COWOS

< 1 pJ / bit (3DIC)

3D Stacks

Comparison with Analog CIM

Power Comparison

Traditional Digital
Compute in Memory (Analog)
Compute in Memory (Digital)

Gain on Software and Hardware

Inside the AI SDK: Block Diagram and Workflow

Compiler-Frontend

Model parsing and model graph generation

Compiler-Middle-end/Back-end

Compilation and optimization of NPU execution

Simulator

Accuracy and functionality analysis

Profiler

Detailed op-level reports of inference performance