inference/research
contestedlow · 0.60updated 2026-10-16

Xilinx Alveo U250/U280 (used FPGA accelerator card)

amd FPGA (UltraScale+ / Alveo) 2018

Fields

id
accelerators/xilinx-alveo-u250-u280-used-fpga-accel-card
Vendor
amd
Architecture
FPGA (UltraScale+ / Alveo)
Release year
2018
Process (nm)
16
Form factors
pcie
VRAM (GB)
8
Memory type
HBM2 (8 GB) + DDR4 RDIMM (32 GB, two 16 GB 2400 MT/s DIMMs)
Memory bandwidth (GB/s)
460
Bandwidth basis
DS963 v1.2 Table 1 for the production U280: HBM2 total bandwidth 460 GB/s over 8 GB, plus a separate DDR total bandwidth of 38 GB/s over 32 GB DDR4. These are two pools on different SLRs, not one address space, so the record reports the HBM2 figure and the DDR4 figure lives in notes. The previous value (77 GB/s, 64 GB DDR4) does not appear in DS963 and is dropped.
TDP (W)
225
Unified memory
no
Consumer part
no
notes
Phase-A correction. DS963 v1.2 (2019-11-20) specifies the production U280 as 8 GB HBM2 at 460 GB/s plus 32 GB DDR4 at 38 GB/s, 225 W, PCIe Gen3 x16 or Gen4 x8, full-height dual-slot, active and passive SKUs. An earlier ingest recorded 64 GB DDR4 at 77 GB/s, which the datasheet does not support; status is contested because the U250 (a different card, no HBM) was lumped into this same record and has not been checked against its own datasheet. The LLM verdict is unchanged and comes from the survey, not the datasheet: used U250/U280 cards are cheap on eBay (~$200-2000 per research-exotic.md) but there is no maintained load-a-GGUF-and-go FPGA LLM stack. Academic flows (SECDA-LLM, custom HLS) exist but require building the compiler yourself. Cheap used Alveo is a paperweight for LLM inference unless the goal is FPGA development.

Notes

Phase-A correction. DS963 v1.2 (2019-11-20) specifies the production U280 as 8 GB HBM2 at 460 GB/s plus 32 GB DDR4 at 38 GB/s, 225 W, PCIe Gen3 x16 or Gen4 x8, full-height dual-slot, active and passive SKUs. An earlier ingest recorded 64 GB DDR4 at 77 GB/s, which the datasheet does not support; status is contested because the U250 (a different card, no HBM) was lumped into this same record and has not been checked against its own datasheet. The LLM verdict is unchanged and comes from the survey, not the datasheet: used U250/U280 cards are cheap on eBay (~$200-2000 per research-exotic.md) but there is no maintained load-a-GGUF-and-go FPGA LLM stack. Academic flows (SECDA-LLM, custom HLS) exist but require building the compiler yourself. Cheap used Alveo is a paperweight for LLM inference unless the goal is FPGA development.

Cross-references

Interconnect (1)