Researched 2026-06-18, cited. Prices are June-2026 street estimates (volatile — GDDR shortage + AI demand). Standard (non-tensor) throughput unless noted.
| GPU | Mem BW (GB/s) | VRAM | FP64 std (TF) | FP32 (TF) | ~Price $ | GB/s per $ | VRAM-GB per $ |
|---|---|---|---|---|---|---|---|
| Intel Arc Pro B70 | 608 (spec; 530 meas. sustained) | 32 GB | ~1.3 (meas.) | 22.9 | ~950–1083 | ~0.59 | ~0.031 ★ |
| NVIDIA RTX 4090 | 1008 | 24 GB | 1.29 (1:64) | 82.6 | ~2400 | ~0.42 | ~0.010 |
| NVIDIA RTX 5090 | 1792 | 32 GB | 1.64 (1:64) | 104.8 | ~4000 | ~0.45 | ~0.008 |
| NVIDIA L40S | 864 | 48 GB | ~1.4 (1:64) | 91.6 | ~8000+ | ~0.10 | ~0.006 |
| NVIDIA A100 80GB | 1935–2039 | 80 GB | 9.7 | 19.5 | ~15k (used less) | ~0.12 | ~0.005 |
| NVIDIA H100 SXM | 3350 | 80 GB | 34 | 67 | ~30–40k | ~0.10 | ~0.002 |
| AMD RX 7900 XTX | 960 | 24 GB | 1.92 (1:32) | ~46 | ~800–1300 | ~0.85 | ~0.022 |
| AMD Radeon Pro W7900 | 864 | 48 GB | ~1.9 | ~46 | ~3400 | ~0.24 | ~0.013 |
| AMD Instinct MI210 | 1638 | 64 GB | 22.6 | 22.6 | ~6–10k | ~0.20 | ~0.008 |
| AMD Instinct MI300X | 5300 | 192 GB | 81.7 | 163.4 | ~15–18k | ~0.32 | ~0.012 |
| Intel Max 1100 | 1229 | 48 GB | 22.2 | 21.8 | ~3–4k | ~0.35 | ~0.013 |
| Intel Max 1550 | 3277 | 128 GB | 52.4 | 51.4 | ~8–10k | ~0.39 | ~0.014 |
- B70 is the clear winner on VRAM-per-dollar among 32GB+ cards (~2.5–4× better than RTX 5090 / W7900 / MI300X / Max 1550). The cheapest way to get 32 GB at decent bandwidth, with native oneAPI/SYCL fit. Sweet spot for large-memory, bandwidth-efficient FP32/FP16 explicit CFD (LBM/FluidX3D) on a budget.
- Bandwidth (608 GB/s spec) is mid-pack consumer-class (≈L40S, below 7900XTX/ 4090); 3.3× behind MI300X, 5.5× behind H100 SXM. GB/s-per-$ competitive (~0.59).
- FP64 is weak (~1.3 TF, like all consumer cards) — strong FP64 is data-center only (MI300X 82, Max 1550 52, H100 34, A100 9.7). But our diagnostics show FP64 is NOT the bottleneck (bandwidth + the SYCL preconditioner software are).
- LBM efficiency: B70 hits 85% of its bandwidth (FluidX3D) — ties RTX 4090, beats Intel's own Max 1100 (47%). For FP32/FP16 LBM the B70 is genuinely good.
- No-compromise upgrade = AMD MI300X (3.3× bandwidth, 192 GB, full FP64) at ~16–19× the price; A100 80GB = meaningful bandwidth+FP64 step-up at sometimes-comparable used price but 2.5× less VRAM/$ and no SYCL-native fit.
Sources: TechPowerUp, Intel ARK (B70), FluidX3D README, Lehmann et al. PRE 2022 (arXiv:2112.08926), NVIDIA/AMD datasheets, NVIDIA HPCG blog.
Note: B70 bandwidth spec is 608 GB/s (256-bit × 19 Gbps); our 530 GB/s is measured sustained (87%). FP64 is vendor-unpublished; ~1.3 TF = our clpeak.