Intel unveils next-gen Xeon Phi chips for deep learning; Nvidia says Intel used old benchmarks in marketing; Xeon Phi chips seem to compete on price regardless
Lucian Armasu / Tom's Hardware :
Context & Ripple Effects
Intel's new Xeon Phi generation arrives with a track record behind it: the chips already anchor two Department of Energy supercomputers under the $200M DOE contract with Cray, so this launch is about pushing that credibility from HPC into commercial deep-learning clusters. The immediate friction is with Nvidia, which publicly accuses Intel of building its marketing on outdated benchmarks — a dispute over whose numbers define the performance baseline.
The telling detail is where the competition lands: with Nvidia holding the performance and software high ground, Xeon Phi is positioned primarily on price. That framing sets up the arc the related coverage traces — from Intel's later Ponte Vecchio GPU plans to head-to-head comparisons like the A100-era analysis — of Intel repeatedly attacking Nvidia's datacenter franchise from below.
First-order effects
- Deep-learning cluster buyers gain a lower-priced accelerator alternative to Nvidia's parts, with Intel betting that cost-per-training-run outweighs CUDA's ecosystem lock-in for some workloads.
- Nvidia is immediately put on defense over benchmark integrity, forcing it to contest Intel's marketing claims rather than just compete on specs.
Second-order effects
- Price-led competition pressures accelerator margins across the market, pushing Nvidia to justify premium pricing through software and performance rather than letting price-per-FLOPS become the default procurement metric.
- HPC vendors like Cray, already shipping Xeon Phi-based systems for the DOE, can extend those designs toward commercial AI customers, blurring the line between supercomputing and enterprise deep-learning hardware.
Third-order effects
- If price remains Intel's lever while performance stays Nvidia's, the pattern points toward Intel building dedicated datacenter GPUs — the direction its later Ponte Vecchio and oneAPI efforts confirm — rather than conceding the accelerator market.
- Sustained two-vendor rivalry in AI accelerators would push buyers toward multi-source procurement strategies, weakening any single vendor's ability to set both the benchmark methodology and the price floor.
The trend: Datacenter AI compute is splitting into a performance-led camp around Nvidia and a price-led challenger camp around Intel, with each successive product cycle re-fighting the same benchmark-and-pricing battle.