SambaNova SN40L
Overview
The SambaNova SN40L is the fourth-generation Reconfigurable Dataflow Unit and the chip behind the SambaRack system. Rather than a fixed matrix pipeline it is a coarse-grained reconfigurable array of 1,040 Pattern Compute Units and 1,040 Pattern Memory Units on a two-dimensional mesh, delivering 638 TFLOPS of BF16. Its defining feature is a three-tier memory system: 520 MB of on-chip distributed SRAM, 64 GB of on-package HBM at 1.8 TB/s, and 1.5 TB of off-package DDR at 200 GB/s. That third tier is the point of the design, letting a single socket hold many models in DDR and switch between them by moving weights into HBM in microseconds, which is how SambaNova runs a 150-expert, trillion-parameter Composition of Experts on eight sockets. The chip is built on TSMC 5nm as a dual-die socket using CoWoS packaging. Note that SambaNova publishes no specification table on its website and no datasheet; these figures come from Table II of its own architecture paper, and the memory tiers are independently confirmed by the SambaRack totals in SambaNova's product brief. SambaNova states the die size only as under 650 mm2 and the clock only as under 2 GHz, so neither is recorded here as a precise value, and no power figure is published anywhere. Neither the paper nor the brief labels the compute figure dense or sparse, and no sparse figure is published.
Performance Metrics
Peak theoretical throughput by precision type
| Precision | Bits | Peak TFLOPS | |
|---|---|---|---|
| BF16 | 16 | 638.0 |
Power Specifications
TDP
--
Max Power
--
Power Connector
PCIe Slot
Cooling
Air
Memory Specifications
Capacity
64 GB
Type
HBM
Bandwidth
1800 GB/s
Interface
--
Hardware & Design
Form Factor
--
Architecture
RDU
Process Node
5nm
Launch Year
--
Variant
Standard
Market Segment
Professional
Chiplets
2 (Dual-die socket packaged with TSMC CoWoS)
Full Specifications
| Compute Engine | |
|---|---|
| PCU | 1,040 |
| Chip Design | |
| Process Node | 5nm |
| Chiplets | 2 (Dual-die socket packaged with TSMC CoWoS) |
| Memory | |
| VRAM | 64 GB |
| Memory Type | HBM |
| Bandwidth | 1.8 TB/s |
| On-Die SRAM | 520 MB |
| Power & Thermal | |
| General | |
| Architecture | RDU |
Documentation & Resources
Official Datasheet
SambaNova SN40L technical specifications
SambaNova SN40L: Scaling the AI Memory Wall with Dataflow and Composition of Experts (Prabhakar et al., SambaNova Systems) -- Table II carries the chip parameters SambaNova publishes nowhere else
SambaNova · Latest version
Common Use Cases
The SambaNova SN40L is optimized for high-performance computing tasks with RDU architecture delivering high TFLOPS of compute power.
Where to Rent
Compare cloud providers offering on-demand GPU instances for AI training, inference, and HPC workloads.
Browse GPU Cloud ProvidersStay Updated on GPU Releases
Get notified when new GPUs are added or specifications are updated.
No spam, unsubscribe anytime.