192GB of VRAM in One PC… The Cheap Way
Alex Ziskind ·2026-07-15 ·2 min read
Summary written by us from the video's transcript. The video, and everything in it, is Alex Ziskind's work.
Learn how to achieve 192 GB of VRAM cheaply by comparing dual Intel B60 cards versus a single dual‑GPU card, with performance, power, and build considerations.
Takeaways
- Dual Intel B60 cards can be combined via tensor parallelism, but performance varies by model size and whether the workload is memory‑ or compute‑bound.
- A single dual‑GPU card (Mac Sun) requires PCIe bifurcation; once enabled, it delivers identical speed to two separate B60s with a small power advantage.
- Intel’s software stack trails Nvidia’s, so expect slightly lower efficiency and delayed support for the newest models.
- Higher VRAM density is achievable by stacking dual‑GPU cards: four such cards yield 192 GB of VRAM at a fraction of Nvidia’s cost.
- For most users, a single B70 offers simplicity; for large‑scale or multi‑user scenarios on a budget, consider two B60s or one dual‑GPU card if your motherboard supports bifurcation.
VRAM Options and Pricing
Running large language models locally hinges on VRAM. Nvidia’s RTX Pro 4000 offers 24 GB for $2,500, while Intel’s B60 provides the same memory for about $650. For higher capacities you can either use two B60 cards (48 GB) or a single card that houses two GPUs.
The Intel B70 costs around $1,000 and supplies 32 GB, whereas a dual‑GPU Mac Sun card (two B60s on one board) delivers 48 GB in the footprint of a single GPU.
Hardware Setup and Bifurcation
The dual‑GPU card requires PCIe bifurcation: splitting one slot into two logical lanes so each GPU is visible to the system. This setting is adjusted in the BIOS, not the OS.
If the motherboard doesn’t support bifurcation, the card appears as “no device discovered.” Consumer boards often allow X4/X4 splits, but verify the manual before buying.
Performance Benchmarks
Small 4 B FP8 model: B70 (92 t/s) outperforms two B60s in tensor parallelism; a single B60 is slower than the dual setup because of communication overhead.
Medium 27 B 4‑bit quantized model: Two B60s beat the B70 (28.3 t/s vs 24 t/s). One B60 cannot run it due to size (~30 GB) plus headroom requirements.
Large 30 B MoE integer model: B70 at 45 t/s, single B60 at 44.5 t/s, and two B60s slightly slower. Results show speed depends on whether the workload is memory‑bound or compute‑bound.
Software Maturity
Intel hardware performs well, but its software stack lags behind Nvidia’s mature CUDA ecosystem, affecting overall efficiency and model support.
The author uses LLM Scaler (a VLM fork) which trails the latest VLM releases by about a month; newer models may not be immediately available.
Power Consumption and Density
Under identical workloads, the dual‑GPU card consumes ~129 W versus 134 W for two separate B60 cards—a modest but consistent saving.
Two‑GPU cards enable higher VRAM density: four single‑slot B60s give 96 GB, four dual‑GPU cards provide 128 GB, and four Maxon dual‑GPU cards reach 192 GB, all far cheaper than Nvidia equivalents (e.g., $24k for comparable RTX Pro 5000 setup).
Choosing the Right Configuration
For small models and simplicity, a single B70 is easiest—fits any motherboard and needs no bifurcation.
When you need >27 B parameters or larger context on a budget, double‑up B60s or use a dual‑GPU card if slot space is limited. The dual‑GPU card offers similar speed with slightly lower power and higher VRAM density.