inference/research
verifiedhigh · 0.95updated 2026-10-03

Intel Arc B580

intel battlemage (Xe2, BMG) 2024

Fields

id
accelerators/intel-arc-b580
Vendor
intel
Architecture
battlemage (Xe2, BMG)
Release year
2024
Process (nm)
5
Form factors
pcie
VRAM (GB)
12
Memory type
gddr6
Memory bus (bit)
192
Memory bandwidth (GB/s)
456
Bandwidth basis
Intel ARK spec table: 'Graphics Memory Bandwidth 456 GB/s' together with 'Graphics Memory Speed 19 Gbps' and 'Graphics Memory Interface 192 bit' (19 Gbps x 192-bit / 8 = 456 GB/s, consistent)
Peak FLOPS
precisionint8tflops233denseyesvendor_claimyessource_idconsumer-intel-arc-b580-ark-specificationsbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int8 for Intel Arc B580 (source of record: 'Intel Arc B580 Graphics ARK product specifications (SKU 241598)', kind: spec-sheet). Recorded dense:true as printed.
TDP (W)
190
Unified memory
no
Consumer part
yes
notes
Consumer Battlemage part: 20 Xe-cores, 160 Xe Matrix Extensions (XMX) engines, 20 RT units, 2670 MHz graphics clock, TSMC N5, PCIe 4.0 x8 (note x8, NOT x16), 190 W TBP, device ID 0xE20B, $249 RCP, launched Q4'24. PRECISION CAVEAT - and Intel's labelling is better than NVIDIA's: the only throughput figure in the ARK table is 'GPU Peak TOPS (Int8) 233'. Intel's own Arc Pro B60 datasheet defines the equivalent figure as 'the peak throughput when running XMX workloads with INT8 datatype and DENSE models'. So 233 TOPS should be read as DENSE INT8 - an integer matmul figure, NOT an FP16 number and NOT an INT4 number. No FP16/BF16/FP8 throughput is published for this part in ARK, so none is recorded here. Inference: 12 GB at 456 GB/s. Capacity is the binding constraint by a wide margin. Also note the x8 PCIe interface: half the Gen4 x16 link rate matters for weight streaming from system RAM and for multi-GPU host transfers, which is a real penalty for offload-based workflows. INTERCONNECT FILLED 2026-10-04: interconnect was empty; it now reads ['pcie-gen4'], from the vendor's own spec table (source Intel ARK's 'PCI Express Configurations' row). SCALE-UP IS ABSENT AND THAT IS THE POINT: this part has no NVLink and no other GPU-to-GPU fabric, so the host link is the ONLY interconnect a multi-GPU deployment has. Do not infer a fabric from the presence of a PCIe value, and do not read the Gen number as a multi-GPU capability.

Notes

Consumer Battlemage part: 20 Xe-cores, 160 Xe Matrix Extensions (XMX) engines, 20 RT units, 2670 MHz graphics clock, TSMC N5, PCIe 4.0 x8 (note x8, NOT x16), 190 W TBP, device ID 0xE20B, $249 RCP, launched Q4'24.

PRECISION CAVEAT - and Intel's labelling is better than NVIDIA's: the only throughput figure in the ARK table is 'GPU Peak TOPS (Int8) 233'. Intel's own Arc Pro B60 datasheet defines the equivalent figure as 'the peak throughput when running XMX workloads with INT8 datatype and DENSE models'. So 233 TOPS should be read as DENSE INT8 - an integer matmul figure, NOT an FP16 number and NOT an INT4 number. No FP16/BF16/FP8 throughput is published for this part in ARK, so none is recorded here.

Inference: 12 GB at 456 GB/s. Capacity is the binding constraint by a wide margin. Also note the x8 PCIe interface: half the Gen4 x16 link rate matters for weight streaming from system RAM and for multi-GPU host transfers, which is a real penalty for offload-based workflows. INTERCONNECT FILLED 2026-10-04: interconnect was empty; it now reads ['pcie-gen4'], from the vendor's own spec table (source Intel ARK's 'PCI Express Configurations' row). SCALE-UP IS ABSENT AND THAT IS THE POINT: this part has no NVLink and no other GPU-to-GPU fabric, so the host link is the ONLY interconnect a multi-GPU deployment has. Do not infer a fabric from the presence of a PCIe value, and do not read the Gen number as a multi-GPU capability.

Cross-references

Interconnect (1)