BMG Hardware Peak and MFU Reference

This note documents the Intel Arc Pro B60 peak throughput assumptions used by the ARK/Sage benchmark notes, including the source links and formulas behind the FP16, BF16, INT8, and FP32 denominators.

INT8 XMX Peak

196.6 TOPS

Matches Intel B60 197 peak TOPS spec after rounding.

FP16/BF16 XMX Peak

98.3 TFLOP/s

Half of INT8 because XMX FP16/BF16 ops per cycle are half of INT8.

FP32 Vector Peak

12.28 TFLOP/s

Matches Intel B60 FP32 throughput spec after rounding.

For the 75.6k non-causal attention benchmark, use 98.3 TFLOP/s as the FP16/BF16 XMX denominator and 196.6 TOPS as the INT8 XMX denominator. The old shorthand FP16 = INT8 / 2 is correct for XMX peak, but the comment should say that 196.6 is the INT8 dense XMX peak, not the FP16 peak.

Source Facts

Fact	Value	Source
B60 GPU peak TOPS	197 TOPS, INT8 dense XMX workload	Intel Arc Pro B60 datasheet
B60 FP32 throughput	Up to 12.28 TFLOP/s	Intel Arc Pro B60 datasheet
B60 compute resources	20 Xe-cores, 160 XMX engines	Intel Arc Pro B-series overview
Per Xe-core matrix throughput	4096 INT8 ops/cycle, 2048 FP16/BF16 ops/cycle	Intel Xe GPU architecture guide
Per Xe-core vector throughput	256 FP32 ops/cycle	Intel Xe GPU architecture guide

Formulas

The B60 spec implies a maximum graphics clock around 2.4 GHz because the official INT8 and FP32 peak numbers both solve to that frequency.

clock_GHz = official_INT8_TOPS / (Xe_cores * INT8_ops_per_Xe_core_per_cycle) = 197 / (20 * 4096) = 0.002404 TOPS per ops/cycle = 2.404 GHz clock_GHz = official_FP32_TFLOPS / (Xe_cores * FP32_ops_per_Xe_core_per_cycle) = 12.28 / (20 * 256) = 0.002398 TFLOP/s per ops/cycle = 2.398 GHz

Using 2.4 GHz for simple benchmark denominators:

INT8_XMX_peak_TOPS = 20 Xe-cores * 4096 INT8 ops/cycle/Xe-core * 2.4 GHz = 196,608 GOPS = 196.608 TOPS ~= 196.6 TOPS FP16_XMX_peak_TFLOPS = 20 Xe-cores * 2048 FP16 ops/cycle/Xe-core * 2.4 GHz = 98,304 GFLOP/s = 98.304 TFLOP/s ~= 98.3 TFLOP/s BF16_XMX_peak_TFLOPS = 20 Xe-cores * 2048 BF16 ops/cycle/Xe-core * 2.4 GHz = 98.304 TFLOP/s FP32_vector_peak_TFLOPS = 20 Xe-cores * 256 FP32 ops/cycle/Xe-core * 2.4 GHz = 12.288 TFLOP/s ~= 12.28 TFLOP/s

MFU Denominator

Model FLOP utilization is a ratio between the benchmark's counted attention FLOP/s and the device's theoretical peak for the datatype and execution unit used by that kernel.

MFU = achieved_TFLOP/s / theoretical_peak_TFLOP/s

Benchmark path	Measured throughput	Peak denominator	MFU formula	MFU
ARK FP16 SDPA, BMG 75.6k non-causal	83.02 TFLOP/s	98.3 TFLOP/s FP16 XMX peak	`83.02 / 98.3`	84.45%
ARK Sage v1 INT8, BMG 75.6k non-causal	98.38 TFLOP/s	196.6 TOPS INT8 XMX peak	`98.38 / 196.6`	50.04%

Do not compare FP16 MFU and INT8 MFU without checking the denominator. INT8 can have better latency and higher absolute achieved throughput while reporting lower MFU because its theoretical XMX peak is twice the FP16/BF16 XMX peak on this architecture.

Benchmark Interpretation

The standalone sycl-tla BKC result around 82.7 TFLOP/s is plausible: it is about 82.7 / 98.3 = 84.1% of B60 FP16/BF16 XMX peak.
The PyTorch XPU SDPA result around 55.9 TFLOP/s is about 55.9 / 98.3 = 56.9% of B60 FP16/BF16 XMX peak, so the gap is a backend/kernel efficiency issue rather than a hardware peak mismatch.
The Sage v1 INT8 result around 98.4 TOPS is about half of INT8 dense XMX peak, but it is already roughly equal to the full FP16/BF16 XMX theoretical peak. That is why raw TFLOP/s and MFU tell different stories.