Intel Data Center GPU Flex 170

Intel ARK Flex 170 spec table, retrieved 2026-10-04: 'Memory Size 16 GB' | 'Memory Type GDDR6' | 'Graphics Memory Interface 256 bit' | 'Graphics Memory Bandwidth 576 GB/s'. Bandwidth is ARK's own figure, not derived.

COLUMN-MIXING ERROR, CORRECTED 2026-10-05. This record previously carried tdp_w 75, memory_bandwidth_gbps 384 and memory_bus_bit 128. All three are the Flex 140's figures; this record is the Flex 170. Intel's own ARK page for the 170 prints 'Memory Size 16 GB' | 'Memory Type GDDR6' | 'Graphics Memory Interface 256 bit' | 'Graphics Memory Bandwidth 576 GB/s' | 'TBP 150 W'. The 384 GB/s and 128-bit were never printed by Intel for either part - 384 is the 140's bandwidth and the 128-bit was an INFERENCE from GDDR6, which ARK contradicts at 256 bit. STATUS CONTESTED because the flops rows are still not reconcilable with any Intel publication. ARK publishes no compute figure for this part at all - no TFLOPS row, no TOPS row. The stored fp32 20 / fp16 80 / int8 160 match neither the 170 nor the 140. The Flex Series product brief gives the 170 as '16 TFLOPS (FP32) / 250 TOPS (INT8)', which is a different pair again. Rather than pick a figure no source supports, the rows are kept and the disagreement is stated. A reader should not plan on these three numbers. SPARSITY: ARK carries no sparsity footnote and no row titled for one on this part, so every published figure is dense as printed and must not be halved - Intel's convention in its strongest form, where there is no sparse row to find. CITATION WITHDRAWN 2026-10-05, THE MOST IMPORTANT CORRECTION IN THIS RECORD: every flops row here cited xpu-onednn-github-readme, and that document contains NONE of these figures. Verified by reading the raw README rather than the rendered GitHub page - the rendered page is what made a citation of this kind look reasonable in the first place. The README's GPU section is an unnumbered bulleted list of supported architectures ('Intel Data Center GPU Flex Series (formerly Arctic Sound)', 'Intel Data Center GPU Max Series (formerly Ponte Vecchio)'). It is arguably the single most plausible-looking citation an Intel GPU's compute figure could carry: same vendor ecosystem, same product-family names, machine-readable, and already present in the record's sources list. That is what makes it the worst kind of error in this repo - a citation that does not support the number, disguised as a first-party Intel document. The rows now carry basis 'unverified', vendor_claim false and source_id null, and the values stay in place under status contested so the disagreement remains open and visible rather than being deleted into a silence that looks like a settled question. WHAT WOULD SETTLE IT: Intel ARK. Prior passes established that ARK publishes no compute row for either part (w4h-intel-dc-gpu-flex-170-ark-spec-table, w4h-intel-dc-gpu-max-1550-ark-spec-table), and the Flex Series product brief gives the 170 as '16 TFLOPS (FP32) / 250 TOPS (INT8)' - a third pair, matching neither. So as of this pass there is NO Intel-published compute figure for this part and a reader should not plan on these numbers. See w5a-part2-onednn-readme-contains-no-tflops-figure.