IBM NorthPole (In-Memory Compute, 2023)
Product Overview
IBM NorthPole is IBM Research's revolutionary AI inference chip, prototype published in 2023-10-19 Science paper ("Neuromorphic computing at scale", Dharmendra Modha team), 22nm process, 458 TOPS INT8 compute, 75W TDP. Uses an In-Memory Compute architecture, all weights stored in on-chip SRAM + analog compute units, eliminating the von Neumann bottleneck (conventional GPUs spend 99% of power on data movement).
Architectural inspiration: derived from IBM's TrueNorth neuromorphic chip (2014, 54B transistors, 100K neurons), NorthPole is TrueNorth's practical AI version, energy efficiency 25x better than GPUs (IBM official paper data).
Strategic significance: IBM is the leader in in-memory compute + neuromorphic AI. NorthPole is the first to scale in-memory compute to 458 TOPS as a commercially viable AI chip. NorthPole 2 expected 2025 H2 release, 25x efficiency advantage.
Core Specs
| Item | Parameter |
|---|
| Architecture | IBM NorthPole (In-Memory Compute) |
| Process | IBM 22nm SOI (Samsung 11nm estimated 2026) |
| Core Count | 256 CISC processors (Custom Instruction Set) |
| SRAM | 224MB on-chip SRAM (one of the largest AI chip SRAM in industry) |
| In-Memory Compute | 1.6B weights + 30B MAC units |
| Memory Bandwidth | 2.5 TB/s (on-chip SRAM) |
| INT8 | 458 TOPS |
| FP16 | ~110 TFLOPS |
| TDP | 75 W |
| Efficiency | 6.1 TOPS/W (H100 ~2.16, 3x advantage) |
| Latency | 5-10ms (in-memory compute = zero data movement) |
| Mass Production | not commercialized (research prototype) |
| Commercial Version | NorthPole 2 2025 H2 estimated |
In-Memory Compute Principle
| Dimension | Traditional GPU (von Neumann) | IBM NorthPole (In-Memory) |
|---|
| Architecture | Memory (DRAM/HBM) + compute (GPU) separate | Memory + compute fused |
| Data Movement | 99% power on data movement | 0 data movement (compute inside SRAM) |
| Operations | Scalar MAC arrays | Analog / digital hybrid |
| Energy | 1x | 0.04x (25x advantage) |
| Latency | HBM-limited | 5-10ms (zero wait) |
| Reconfigurable | CUDA programs | Network topology config |
| Precision | FP64/FP32/FP16/INT8 | INT8 primarily (analog compute limits) |
| Drawback | - | inference only, INT8 limited, training immature |
How In-Memory Compute Works
Traditional GPU:
Load weights (HBM) -> Load input (HBM) -> MAC (CUDA) -> Store result (HBM)
Total energy: 100% (99% on data movement)
IBM NorthPole:
Weights pre-stored in SRAM analog units (immutable)
Load input (SRAM) -> Analog MAC (inside SRAM) -> Store result (SRAM)
Total energy: 4% (data movement 0-1%)
Key advantages:
- 224MB SRAM stores all weights at once (LLM 70B INT8 = 70GB still needs HBM, but small models pure SRAM)
- 30B analog MAC units computing simultaneously
- 6.1 TOPS/W (H100 3x efficiency)
256 CISC Processors
| Dimension | Spec |
|---|
| Architecture | CISC (Custom Instruction Set) |
| Core Count | 256 |
| Per Core | 64KB SRAM + 4 analog MAC units |
| Frequency | 1.4 GHz |
| Role | Scheduling + activation functions + scalar ops |
| ISA | Proprietary (not RISC-V, not ARM) |
| Programming | Neural network topology graph config (TrueNorth-like) |
CISC vs RISC: NorthPole doesn't use RISC-V because in-memory compute requires custom instructions for neural topology compilation. TrueNorth -> NorthPole is IBM's 10-year R&D accumulation.
25x Efficiency Source
| Factor | Energy Savings |
|---|
| Data movement reduction | 20x (vs HBM) |
| Analog computing | 3x (vs digital) |
| SRAM internal compute | 1.5x (vs registers) |
| 22nm SOI | 0.8x (vs 5nm digital) |
| Total | 25x (IBM paper data) |
IBM paper conclusion: NorthPole on ResNet-50 inference, 25x more energy-efficient than NVIDIA H100, 25x faster (same precision).
| Dimension | IBM NorthPole | NVIDIA V100 | NVIDIA H100 |
|---|
| Latency | 5ms | 8ms | 2ms |
| Throughput | 7,000 images/s | 5,000 images/s | 15,000 images/s |
| Efficiency | 6.1 TOPS/W | 0.4 TOPS/W | 2.16 TOPS/W |
| Power | 75W | 250W | 700W |
| Precision | INT8 | FP16 | FP8 |
NorthPole advantage: 5ms latency 1.6x V100, but 15x efficiency. H100 wins on throughput (FP8 advantage), but NorthPole wins in low-latency + low-power scenarios.
| Item | Content |
|---|
| Company | IBM Research |
| Lab | IBM Research - Almaden (San Jose, California) |
| Chief Scientist | Dharmendra S. Modha (IBM Fellow) |
| Team | 100+ IBM Research engineers |
| Publication | Science 2023-10-19 ("Neuromorphic computing at scale") |
| Paper Citations | 200+ (2024-2026) |
| Commercialization | not commercialized (IBM doesn't sell directly) |
| Commercial Path | IBM Cloud inference service (future) + IP licensing (Samsung 11nm 2026) |
| Customers | US DARPA, NASA, Department of Energy |
| Competitors | Mythic (digital CIM), Syntiant (edge CIM), ChiCore (China) |
IBM Neuromorphic AI Evolution
| Product | Released | Transistors | Neurons | Compute | Purpose |
|---|
| TrueNorth | 2014 | 54 B | 100 K | - | Neuromorphic research |
| NorthPole | 2023-10 | 220 B | analog | 458 TOPS INT8 | AI inference |
| NorthPole 2 | 2025 H2 estimated | - | analog | 1.2 POPS INT8 (estimated) | AI inference + training |
| NorthPole 3 (est.) | 2027 | - | analog | 5 POPS | General AI |
Use Cases
- ✅ Low-latency AI inference (5-10ms, ultra-low latency)
- ✅ Ultra-low-power AI (75W, 3-25x GPU efficiency)
- ✅ Government/research HPC (US DARPA, NASA, DOE)
- ✅ Neuromorphic AI research (next-gen AI architecture)
- ✅ Small model inference (7B-13B <70GB fits 224MB as pure SRAM)
- ❌ AI training (NorthPole inference only)
- ❌ Large model training (<224MB SRAM limit)
- ❌ Commercial purchase (IBM not commercialized)
- ❌ CUDA compatibility (proprietary ISA)
IBM In-Memory Compute Strategy
- IBM Research AI flagship project: Modha team 10-year R&D
- DARPA funding: SyNAPSE program (2014-2024 $100M+ cumulative)
- NorthPole 2: 2025 H2 commercial version, Samsung 11nm collaboration
- AI Cloud service: IBM Cloud integrated NorthPole inference
- Open-source software: IBM plans to open-source NorthPole compilation stack (PyTorch integration)
Key Features
- In-Memory Compute: first 458 TOPS scale in-memory compute
- 224MB SRAM: largest AI chip SRAM in industry
- 6.1 TOPS/W: H100 3x efficiency
- 5ms latency: real-time AI inference
- 75W TDP: air-cooled deployment
- Drawbacks: not commercialized, INT8 only, no training support
Neuromorphic AI Big Three
| Company | Product | Compute | Status |
|---|
| IBM | NorthPole | 458 TOPS INT8 | 2023 prototype |
| Intel | Loihi 2 | 1M neurons | 2021 neuromorphic research |
| Brainchip | Akida 2 | 200 GOPS INT8 | 2023 Edge commercial |