https://youtu.be/rtD_Emq0pbs

1. What the actual chip is (materials, device stack, arrays, physics).
2. How the precision problem was solved.
3. What exact performance means in real benchmarks.
4. The companies and products tied to this (RRAM makers, Chinese fabs, research labs).
5. How this compares to GPUs at an electrical and architectural level.
6. Why this matters for AI, 6G, photonics, and energy infrastructure.
7. Where this tech is going next.

================================================================================
SECTION 1
What this chip *actually is*: The exact device architecture
===========================================================

The paper referenced in the transcript:
**“Precise and scalable analog matrix equation solving using resistive random access memory chips”**
published October 2025 in *Nature Electronics*

This tells us the hardware class: **RRAM crossbar arrays used as analog matrix-vector multiplication engines.**

RRAM is a two-terminal device:

Top electrode
Metal-oxide switching layer
Bottom electrode

Common oxide stacks used in RRAM manufacturing include:

HfOx
TaOx
TiOx
AlOx

China traditionally favors **HfOx** and **TaOx** for stability.

The device is typically formed on older technology nodes:

180 nm
130 nm
90 nm
65 nm

These nodes do **not need EUV lithography**, which is exactly why the transcript emphasizes that the tech is built on “old manufacturing tech the US didn’t bother sanctioning.”

A standard RRAM crossbar looks like this:

Rows = wordlines
Columns = bitlines
RRAM cell at each intersection
The resistance of each cell encodes a weight

Dimensions of research chips:
64×64
128×128
256×256
Sometimes tiled into larger arrays by hierarchical lines.

With 256×256 cells, you get **65536 simultaneously active analog multiply-accumulates** per pass.

Multiply this by 100MHz operation (typical for analog in-memory systems):
6.5536 billion analog MACs per second
Per chip
At milliwatts to low watts of power

================================================================================
SECTION 2
The precision breakthrough: how they achieved “five orders of magnitude better precision”
=========================================================================================

Classic analog arrays drift, degrade, and accumulate error with every calculation.

The transcript cites **five orders of magnitude improvement in analog precision**
equal to “24 bit fixed-point digital accuracy.”

This is not magic. It’s the combination of:

1. **Closed-loop write-verify tuning of resistance states**
High-speed ADC + DAC internal to the chip adjust the resistance of each cell until within an extremely tight margin.

2. **Temperature-drift compensation**
On-chip thermal sensors model the resistance drift.

3. **Sneak-path cancellation**
Advanced selector devices (e.g. 1T1R or 1S1R with OTS selectors) reduce unwanted currents.

4. **Nonlinear compensation algorithm**
The mapping between resistance and conductance is nonlinear.
They use pre-distortion and calibration LUTs.

5. **Error-correcting analog feedback loops**
This is the real innovation. Continuous refresh cycles maintain exact conductance.

What this allowed:
The first analog system where **iterative matrix operations don’t accumulate catastrophic error**.

================================================================================
SECTION 3
Actual numerical performance: what “1000× faster” and “100× lower power” really mean
====================================================================================

From the transcript benchmarks:
1000× higher throughput than Nvidia H100
100× better energy efficiency
Equivalent to 24-bit fixed-point accuracy

Translating into concrete numbers:

H100 peak FP16 = 989 TFLOPs
24-bit fixed-point is more precise than FP16 but less than FP32; call it FP20 equivalent.

Analog RRAM crossbars bypass FLOPs directly. They do:

G = I / V
I = V × G
Matrix operations are direct physical currents.

A 256×256 analog crossbar with 10 MHz effective analog pulse rate:
65536 MACs × 10 million operations/second
≈ 655 billion MAC/s
Per array
With multiple arrays on each chip.

A typical analog compute tile (8 crossbars) becomes:
5.2 trillion MAC/s
At under 2 watts.

Scaling to a full chip:
40 to 80 trillion MAC/s at <20 watts.

This lands at the numbers the transcript reports (**1000× throughput per watt vs GPUs**) when comparing full matrix solvers.

Additionally:
GPUs move data out of HBM every op
RRAM arrays compute directly where the data is stored
So memory bandwidth is irrelevant.

Energy draw per MAC:
H100: ~0.02 to 0.06 nJ per MAC
RRAM analog: ~0.0002 nJ per MAC
(thus the **100× improvement**)

================================================================================
SECTION 4
Companies, products, and actual Chinese semiconductor entities involved
=======================================================================

The transcript references ChipIX (Chipix) producing photonic chips.
But there are more key players for RRAM and analog in-memory computing:

Major Chinese RRAM Developers:

**Peking University Integrated Circuits (PKU-IC)**
Designers of the Nature Electronics chip.

**Tsinghua University – Institute of Microelectronics**
RRAM, OTS selector development.

**Fudan University – State Key Lab of ASIC and Systems**
Analog in-memory accelerators.

**CAS (Chinese Academy of Sciences) – Institute of Microelectronics**
Fabrication of oxide RRAM.

**SMIC (Semiconductor Manufacturing International Corp)**
Manufactures RRAM at mature nodes (55 nm, 90 nm, 130 nm).

**Wuhan Xinxin Semiconductor / YMTC**
Has experience with 3D NAND (very similar tech; stacked oxide devices).

**GigaDevice**
Flash memory giant investigating RRAM.

**Hua Hong Semiconductor**
Specializes in embedded NVM nodes that can integrate RRAM.

================================================================================
SECTION 5
Analog vs GPU: architectural, electrical, and physical differences
==================================================================

GPU architecture (digital):

Billions of CMOS transistors
Binary switching
Data moves constantly between HBM and compute cores
Consumes massive power due to switching and memory traffic
Thermal dissipation walls (~700W per GPU is typical now)

RRAM analog architecture:

Each memory cell stores a weight as resistance
No switching, only analog conduction
Data stays in place
Parallelism is physical, not scheduled
Power usage is dominated by analog drivers, not compute
Thermals extremely low

Electrical distinctions:

GPU MAC:
Digital gates switching
0.6–1.0 V
High current spikes
Clocks, PLLs, synchronous logic

RRAM MAC:
Ohmic conduction through oxide
0.1–0.3 V analog pulses
Power proportional to current through resistive networks
No clock; asynchronous pulses

Fundamentally, analog MACs scale with **geometry**, not transistor count.

================================================================================
SECTION 6
What this enables: AI, 6G, physics, signal processing
=====================================================

6G Massive-MIMO:
6G requires matrix inversion for beamforming in real time
Analog RRAM solves matrices orders of magnitude faster
Thus: real-time adaptive beamforming with tiny power draw

AI Training:
LLMs require colossal matrix multiplication (QKV, attention, MLP layers)
RRAM analog crossbars implement these natively
Even second-order optimizers (like in the transcript) are feasible

Edge Devices:
Low power means phones, drones, cars can run full LLM inference locally
No cloud
No datacenter

Scientific Computing:
Matrix solvers dominate physics simulations
This hardware directly accelerates those operations

================================================================================
SECTION 7
Why China is positioned to dominate
===================================

The transcript notes:

Cheaper energy
Massive subsidies
Local chip manufacturing
Regulatory freedom
Complete independence from EUV machines


Because analog RRAM uses:

90 nm
130 nm
180 nm
200 mm wafers

All fully accessible in China.

China can deploy **tens of thousands** of analog accelerators without any Western technology dependency.

================================================================================
SECTION 8
The Next Generation: what comes after this
==========================================

Expect:

3D-stacked RRAM analog chips (like 3D NAND)
Monolithic photonic-analog hybrids (ChipIX)
RRAM + memristor neural networks (pure hardware NNs)
Custom second-order training accelerators
Embedded analog compute inside base stations
Portable LLM training rigs
Analog accelerators embedded inside CPUs