SambaNova SambaRack SN40L-16
Sixteen SN40L Reconfigurable Dataflow Units in a single air-cooled rack, which SambaNova positions as able to run models the size of DeepSeek R1 671B and Llama 4 Maverick without liquid cooling. The rack draws an average of 10 kW. Compute is 10.2 PFLOPS of BF16 and memory is three-tier: 8.3 GB of on-chip SRAM, 1 TB of HBM at 28.8 TB/s, and 24 TB of high-capacity DDR. That third tier is the point of the architecture, letting a rack hold many models in DDR and move the active one into HBM in microseconds. Note that the 24 TB of DDR is deliberately not recorded as the system memory figure here, which carries the 1 TB of HBM instead, because the two tiers are different things and conflating them would suggest 24 TB of high-bandwidth memory. SambaNova publishes no rack-level compute figure; the 10.2 PFLOPS is sixteen times the per-socket figure from its own architecture paper, which is safe because its two published rack memory figures of 1 TB and 24 TB are likewise exactly sixteen times the socket. Neither source labels the compute dense or sparse. SambaNova also publishes per-user token rates for this configuration, shown separately below.
We will point you at suppliers who have it. Free, and no signup.
System Details
GPU Configuration
System Specifications
Precision Performance Breakdown
| Precision | System Performance | Per GPU | Efficiency |
|---|---|---|---|
| 10,208 TFLOPS | 638 TFLOPS | — |
Where a vendor states a basis, the figure shown is the dense one, and any figure it publishes is listed separately. Vendors commonly headline the sparse number instead: for NVIDIA's 2:4 structured sparsity that is exactly twice the dense figure, though other vendors' sparse modes do not all follow that ratio.
Published throughput
| Model | Tokens/sec | Basis | Conditions |
|---|---|---|---|
| Llama 3.1 8B | 1,042 | Per user | BF16, 8192 context |
| Llama 3.1 70B | 457 | Per user | BF16, 8192 context |
| Llama 3.1 405B | 129 | Per user | BF16, 8192 context |
Tokens per second is a benchmark result, not a hardware specification. It depends on the model, the precision and the latency target it was measured under, so figures are only comparable when every condition matches. Vendor claims unless stated otherwise.
Powered by SambaNova SN40L
This system utilizes 16 × SambaNova SN40L GPUs, each delivering exceptional performance for AI and HPC workloads.
Per GPU TDP
—
Per GPU Memory
64 GB
Process Node
5nm
Architecture
RDU
Documentation & Resources
Flopper Spec Sheet
SambaNova SambaRack SN40L-16 specifications, generated from our database. Printable.
Vendor product page
SambaNova SambaRack SN40L-16 documentation
Introducing the SN50 RDU, purpose-built for agentic inference (SambaNova blog)
SambaNova • 2026-02-24
Typical Use Cases
The SambaNova SambaRack SN40L-16 runs 16× SN40L GPUs over Peer-to-peer RDU interconnect with hardware collective offload; PCIe host attach per socket, delivering 10,208 TFLOPS BF16 dense.