Intel Announces Arc Pro B70 and Arc Pro B65 GPUs
Intel’s new Arc Pro B70 and B65 workstation GPUs, each with 32 GB of VRAM and around 600 GB/s memory bandwidth, are drawing interest as relatively affordable alternatives to Nvidia and AMD cards for AI inference, virtualization, and media workloads. Commenters highlight strong value for local LLMs, SR-IOV support, and good Linux integration, while noting weaker software maturity, limited VR and gaming optimizations, and uncertainty about Intel’s long‑term commitment to discrete GPUs. Many see the cards as attractive for exploratory or multi-GPU setups, but question whether 32 GB is sufficient for future large models and whether driver support will keep pace.
Pricing, Specs, and Positioning
- Both Arc Pro B70 and B65 offer 32 GB VRAM and ~600 GB/s bandwidth, seen as strong on memory for the price.
- Reported pricing: B70 around $950–$1,000, B65 around $650, with some comparing them favorably to 32 GB NVIDIA cards starting near $4,000 and AMD’s ~32 GB alternatives at ~$1,400.
- Some characterize B70 as roughly a “70-class” card with extra VRAM, not a top-tier compute monster.
- Debate over performance: one estimate claimed ~1/10th of a 5090’s INT8 TOPS, later corrected to closer to ~1/2 after accounting for sparsity and FP4 differences and more apples-to-apples comparisons.
AI, Inference, and VRAM Needs
- Many see 32 GB VRAM at this price as compelling for local LLMs, exploratory work, and long-context models.
- Others dismiss 32 GB as already inadequate and fantasize about 128–512 GB or even 1 TB VRAM cards, though commenters argue such capacities are currently impractical and would be astronomically expensive.
- Some users report successful vLLM/intel-xpu setups on Arc Pro for FP8/FP16 models with large contexts, albeit with rough edges and missing features (e.g., FP8 KV cache).
- Interest in multi-GPU setups and cheap platforms with enough PCIe lanes; suggestions include used server platforms (e.g., EPYC/Threadripper) due to lane count limits on consumer CPUs.
Virtualization and Workstation Use
- SR-IOV support is a major attraction for VM and virtual workstation workloads.
- Several see these as viable replacements for older Tesla cards or as cheap inference/virtualization accelerators.
- Questions arise about mixing Intel and NVIDIA cards in one system; no clear consensus is given.
Linux vs macOS / Apple Silicon
- Some compare the B70 to Apple’s M-series (e.g., M4 Max) and argue a Mac Studio might be a better value.
- Counterpoints emphasize Linux as a better server OS, easier hardware modularity, and stronger support for virtualization, containers, and standard PCIe GPUs.
- macOS is criticized as a server platform: power-management quirks, security restrictions, weaker ecosystem for drivers and filesystems.
Drivers, Stability, and Intel’s GPU Future
- Multiple users report generally solid Linux support for Arc (especially for media transcoding), with some bugs for LLM workloads.
- Concern and confusion about reports that Intel laid off much of its Arc and Linux driver teams; some claim future dGPU support is effectively dead, while others point out these are rumor-heavy, recurring claims and note Intel continues to release new GPUs.
- Some criticize Intel’s timing and past “paper launches,” calling these cards “too little, too late” if availability and long-term support lag.