Offload economics

1Three currencies: cores, watts, dollars

An offload case has exactly three quantities, and each must be sourced. Cores freed on the host: cores_freed = hosts x cores_per_host x infra_share x offloadable_share, where infra_share is the fraction of host CPU consumed today by infrastructure work (measure it per process on the customer’s fleet) and offloadable_share is the fraction of that work an engine on the DPU can actually take, which M6.2 taught you to answer engine by engine: on BlueField-3, SHA is zero because “NVIDIA BlueField-3 does not support this library because it has no SHA acceleration engine”, and deflate compression is zero because the Compress guide lists only decompress operations for BlueField-3.^[8][9] Watts: watts_net = cpu_watts_saved - dpu_watts + displaced_nic_watts; the DPU’s own draw is a real line item, and Dell rates the B3220 and B3240 at 150 W in the R7725 and R770 manuals, with the B3140H SuperNIC at 75 W.^[5][6] Dollars: annual_savings = cores_freed x L + watts_net x 8.76 x P, where L is the customer’s per-core annual software or consolidation value, 8.76 converts one watt for a year into kWh, and P is their electricity price; payback_years = dpu_premium / annual_savings.

What can be measured versus asserted matters more than the arithmetic. DOCA Bench’s --record-cpu-usage “reports CPU statistics for any CPU cores DOCA Bench is executing on. This provides guidance on how much CPU time is returned, and thus available to other processes or threads”, and --sweep core-count,1,8,*2 with --csv-output-file produces one row per core count — the CSVs you collected in M6.1 through M6.3 are that evidence.^[7] The libraries index tells you which engines are GA, Beta or Alpha in 3.5.0, which decides whether a saving is bankable this quarter.^[10] The datasheet’s workload buckets — cloud overlay and SDN acceleration, NVMe-oF and elastic storage, encryption, data deduplication, decompression and erasure coding/RAID, distributed next-generation firewall and IDS/IPS, cloud-native supercomputing — are the list of infrastructure work a DPU can host, not a list of savings; note that the storage bucket says decompression and the features list a “Decompression engine”, which is the same asymmetry M6.2 measured.^[11]

2The public numbers and where they come from

Every widely repeated offload figure has a provenance, a hardware generation and a workload. The table below keeps them together; quote a number only with its whole row.

Three things stand out. Every CPU-saving figure is BlueField-2 era; none was re-measured on BlueField-3 in a fetched source, so the BF-3 case borrows BF-2 proportions and must say so.^[1][2] The BlueField-3 blog gives capability multipliers (bandwidth, compute, memory) but no CPU-saving measurement of its own; “nearly half” is an industry-wide framing, not a per-server result.^[3] And the Red Hat test explicitly omits core counts and CPU models, so a 70% utilization drop cannot be converted to cores without the customer’s own baseline.^[2]

3DPU mode vs NIC mode vs ConnectX: the break-even

The modes guide is the whole argument in two sentences. In DPU mode “the embedded Arm system runs services that manage the NIC resources and data path”; in NIC mode on BlueField-3 “the Arm cores of BlueField are inactive, and the device functions as an NVIDIA ConnectX network adapter”, and “Operating in NIC Mode on BlueField-3 reduces power consumption, improves network performance, and minimizes the host memory footprint”.^[4] So DPU mode has a standing cost — card power, host memory footprint, a second OS to patch — that is only repaid by services running on the Arm cores: OVS-DOCA or HBN for networking, SNAP for storage emulation, App Shield and Argus for security, DPF for fleet control (M4 and M5).^[4][11] If none of those will run, the break-even never arrives.

The product ladder follows. BlueField-3 DPU SKUs (B3220, B3240) are 150 W cards with the 8-pin auxiliary connector; the B3140H SuperNIC is a 75 W slot-powered card shipped in NIC mode; ConnectX-8 has no Arm cores at all and is positioned for hyperscale AI east-west traffic at 800 Gb/s.^[5][13][12] Spectrum-X pairs SuperNICs, not DPUs, with its switches for the GPU fabric.^[14] From M6.3 you know the DPA exists on ConnectX-8 too, so PCC and data-path kernels do not require Arm cores; from M6.4 you know UROM does.^[16][12]

Break-even reasoning, per host and per year: choose DPU mode when cores_freed_per_host x L + isolation_value > (dpu_watts - supernic_watts) x 8.76 x P + dpu_premium / horizon, where isolation_value is what the customer assigns to running infrastructure outside the tenant’s trust domain (Zero-Trust mode) — often the real reason, and often unpriced.^[4] Choose NIC mode or ConnectX when the left side is near zero: GPU-to-GPU RDMA, HPC MPI, or any node where infrastructure already lives elsewhere. Choose a DPU for a subset of nodes — the north-south and storage gateways — and SuperNICs for the rest when the fleet is mixed; that is how a training pod is usually laid out, with the exact per-node BOM confirmed from the customer’s configuration rather than assumed.^[14]

4Compute one scenario by hand, then defend two