Google Coral Edge TPU
google Edge TPU (vision ASIC) 2019
Fields
id
accelerators/google-coral-edge-tpuVendor
google
Architecture
Edge TPU (vision ASIC)
Release year
2019
Form factors
pciem.2usb
Bandwidth basis
Edge TPU is an INT8 vision accelerator with no LLM-relevant memory-bandwidth figure published in a form comparable to GPU VRAM bandwidth; it runs LiteRT/EdgeTPU-compiled CNNs, not transformer decode.
Peak FLOPS
precisionint8tflops4denseyesvendor_claimyessource_idgoogle-coral-products-pagebasisvendor-spec-sheetbasis_detailGoogle Coral Edge TPU: 4 TOPS INT8, vendor-published. This is a vision/camera inference figure, not an LLM path.unit—
Unified memory
no
Consumer part
yes
notes
PAPERWEIGHT for LLM. Coral is an INT8-only VISION accelerator (~4 TOPS). The whole stack is LiteRT/EdgeTPU-compiled CNNs - classification, detection, pose, segmentation. There is NO LLM serving path, no transformer-decode support, no int4/GGUF story; 4 TOPS INT8 cannot hold a 7B weight set let alone stream one. Used units trade at $50-130 and are often bundled cheap with other gear. Fine for a camera/detection project; do not buy it expecting language models.
Notes
PAPERWEIGHT for LLM. Coral is an INT8-only VISION accelerator (~4 TOPS). The whole stack is LiteRT/EdgeTPU-compiled CNNs - classification, detection, pose, segmentation. There is NO LLM serving path, no transformer-decode support, no int4/GGUF story; 4 TOPS INT8 cannot hold a 7B weight set let alone stream one. Used units trade at $50-130 and are often bundled cheap with other gear. Fine for a camera/detection project; do not buy it expecting language models.