Nvidia’s RTX 5090 is a uniquely capable GPU – the company’s highest-end consumer-grade GPU to date features a whopping 21,760 CUDA cores and 32GB of GDDR7 memory in its cut-down GB202 configuration.
It also happens to be one of the most expensive GPUs retail consumers can buy in 2026, going for as much as two times its stated MSRP of $2,000 for the Founders Edition at launch, leaving many cost-sensitive AI builders turning to older Nvidia GPUs equipped with 24GB of VRAM such as the RTX 4090, 3090Ti and 3090.
As AI models continue to advance in complexity (and memory requirements), a DRAM crisis is making it impossible for many to secure a capable GPU for their local AI needs, even as one option seems largely ignored: Intel’s Arc Pro B70.
A mid-range GPU with oodles of VRAM
Most consumers do not consider Intel a GPU maker, associating it instead with consumer- and enterprise-grade CPUs. This is despite the chipmaker offering some of the best value-for-money options in the mid-range segment it currently competes in.
Intel’s Arc Pro offerings, however, are unapologetically different; They focus exclusively on professional-grade, or rather, AI-centric workloads while skipping any pretense of catering to gaming consumers, even as they remain capable of running most titles one can throw at them.
The Intel Arc Pro B70 is its highest-end offering and commands a reference price of $950, with most retailers and OEM partners selling SKUs at around $1,000. With 32GB of GDDR6 memory in tow and a price that is effectively a quarter of most RTX 5090 SKUs on sale at the time of writing, it holds its own as a value-centric alternative to Nvidia’s Blackwell-based behemoth.
It leverages Intel’s BMG-G31 “Big Battlemage” chip that was originally slated to show on its since-canceled Arc B770 GPU, as it aims to address a vacuum in the local AI space that both AMD and Nvidia are not willing to address given their focus on maximizing profits at the highest end of the spectrum with enterprise consumers.
For those interested in looking at setting up 4 of the Arc Pro B70, Puget Systems has gone ahead and done the necessary testing, even drawing a direct comparison with the RTX 5090, which it finds 4-5x faster in decode tasks thanks to a significantly larger amount of memory bandwidth in play (1792 GB/s vs 608 GB/s)
The observations also pointed out a slightly obvious caveat in the comparisons, however: models that require more compute and bandwidth versus memory will lean toward the RTX 5090, which offers a significant advantage even versus multiple Arc Pro B70 GPUs, but those that require a large amount of memory for their parameters would find a B70 configuration a much more cost-efficient play with access to more memory than a corresponding RTX 5090-based offering.
Nvidia’s value proposition, however, runs deeper than the silicon itself, and it is where Intel struggles to find a willing buyer, thanks to better software support on various OSes, CUDA-based software stacks that the Arc Pro cannot emulate, and Nvidia-coded libraries that will not function on Intel’s hardware.
Intel’s own offerings (oneAPI, OpenVINO, and IPEX) are improving but are widely considered to be behind even AMD’s ROCm stack, which in turn lags behind Nvidia’s ecosystem. Despite this, the Arc Pro B70, as noted above, benefits from the lack of a real high-VRAM alternative and remains available close to its stated MSRP, making it a formidable and scalable alternative to Nvidia’s heavyweight GPU.
