{"data":[{"id":"intel_b70","display_name":"Intel Arc Pro B70","vendor":"intel","kind":"gpu","memory_gb":32,"memory_bw_gbps":608,"fp8_dense_tflops":null,"confirm_sku":false,"reference_stack":"vLLM (XPU backend), OpenVINO","active":true,"notes":"R8b resolved \"B70\" as Intel Arc Pro B70 (Xe2/Battlemage, BMG-G31), confirm_sku=0 on exact-name-match + multiple corroborating vendor/retail/press sources. No native FP8 on this generation (fp8_dense_tflops NULL by design, not a gap) — best-available-precision dense is FP16 via XMX at 45.88 TFLOPS (community-derived, not on Intel's own spec page). XMX silicon supports BF16 @ 2048 ops/clock, but the current vLLM XPU backend only serves FP16 — a software gap, not a hardware ceiling (re-check closer to any future re-seed, may resolve). Launched 2026-03-25, $949 MSRP (not stored — no price column)."},{"id":"nvidia_dgx_spark","display_name":"NVIDIA DGX Spark (GB10 Grace Blackwell)","vendor":"nvidia","kind":"system","memory_gb":128,"memory_bw_gbps":273,"fp8_dense_tflops":null,"confirm_sku":false,"reference_stack":"llama.cpp (MXFP4), SGLang, vLLM (Grace-Blackwell build)","active":true,"notes":"fp8_dense_tflops intentionally NULL: NVIDIA publishes only FP4-sparse (1 PFLOP theoretical) on the official spec page, no vendor FP8 figure exists — do not fabricate one (R8b §1, Recommendation 2). Bandwidth-bound decode-mechanism finding (R8b §1, key takeaway): dense 70B FP8 ~2.7-3 tok/s single-stream (273GB/s / ~70GB weights ~= 3.9 tok/s theoretical ceiling, ~25-30% framework overhead observed) vs MoE gpt-oss-120B (~5B active/token) ~38-55 tok/s — MoE active-param-fraction must be a first-class input to the estimation ladder, not folded into a flat hardware-class multiplier. Price ~$3,999-4,699 (unconfirmed exact MSRP, not stored — no price column in schema)."},{"id":"nvidia_h100_sxm_80g","display_name":"NVIDIA H100 SXM5 80GB","vendor":"nvidia","kind":"gpu","memory_gb":80,"memory_bw_gbps":3350,"fp8_dense_tflops":1979,"confirm_sku":false,"reference_stack":"vLLM, TensorRT-LLM, SGLang","active":true,"notes":"1979 TFLOPS is DENSE FP8 (NOT the 3958 TFLOPS 2:4-structured-sparsity marketing figure — R8b §2). PCIe variant differs (HBM2e, ~2000 GB/s) — do not conflate with this SXM5 row. Calibration anchor: true batch=1 single-stream 70B-FP8 decode ~20-30 tok/s/user (R8b §2 takeaway); the ~400+ tok/s figures seen in vendor decks are aggregate-batched, not per-user."}]}