ALVEO U250 Platform
Video AI on a single card.
Every video pipeline runs on the AMD ALVEO U250 data center accelerator. Consolidating AI models and ASIC encoding logic onto one card delivers throughput and power efficiency well beyond CPU or GPU.
2026 · REV. 1
ALVEO U250 Platform at a glance
Data center-grade hardware
UltraScale+ XCU250 FPGA with 1,728K LUTs and 12,288 DSP slices.
64GB ECC memory
Four 16GB DDR4 ECC DIMMs at 77GB/s, enough for 8K frame buffers.
Single-card integration
AI preprocessing and ASIC encoding finish on one card — no PCIe round trip.
Linear scaling
Multiple cards per server and rack-level scaling grow channel count directly.
How it works
The path data takes, from host to accelerator.
Host interface
PCIe Gen3 ×16 connects to the host, with XDMA moving frame buffers.
SLR partitioning
AI kernels and encoding ASIC logic are placed across four SLRs to reduce contention.
DDR4 bank separation
Per-channel memory banks guarantee bandwidth under concurrent multi-channel load.
Key specifications
Technical specifications
Card specifications per the AMD datasheet (DS962), plus the Lemonflex implementation.
- Device
- UltraScale+ XCU250
- Look-up tables
- 1,728K
- Registers
- 3,456K
- DSP slices
- 12,288
- UltraRAM
- 1,280
- SLRs
- 4 (SSI technology)
- DDR4 capacity
- 64GB (4×16GB ECC)
- DDR4 data rate
- 2,400 MT/s
- DDR4 bandwidth
- 77 GB/s
- Host interface
- PCIe Gen3 ×16
- Network
- 2× QSFP28 (100GbE)
- Form factor
- FHFL · dual slot
- Maximum total power
- 225W
- Slot power
- 65W
- AUX power
- 8-pin PCIe (150W)
- Cooling
- Active or passive
- Monitoring
- SMBus voltage, current, temperature
- Concurrent 4K channels
- Up to 32ch
- Maximum resolution
- 8K (multi-card)
- Supported codecs
- H.264 · HEVC · AV1
- Encoding latency
- Sub-frame (< 8ms)
- Driver
- XRT (Linux)
Frequently asked questions
Anything not covered here, we answer in a technical meeting.
Why an FPGA-based accelerator rather than a GPU?
Encoding is a workload of repeated fixed-function operations. Implementing the encoding pipeline directly in circuit on an FPGA removes GPU kernel launch overhead and memory round trips, bringing latency below frame time. Within the same power budget, the number of channels you can process also rises.
Can we just add a card to an existing server?
If the server has FHFL dual-slot clearance, an 8-pin PCIe AUX power connector (150W), and a PCIe Gen3 ×16 slot, then yes. Send us your server specification and we will confirm compatibility.
How do we increase channel count?
Adding cards per server or adding nodes grows channel count close to linearly. There is no shared state between cards, so scaling out costs almost no performance.
What happens if a card fails?
In multi-card configurations we support failover that redistributes channels across remaining cards. SMBus temperature and power monitoring detects problems in advance, and dedicated SLA agreements guarantee replacement response times.
Your video infrastructure,
one level up.
Start with a single card and scale as needed. We design the demo, the rollout, and the technical review with you.