DEV Community

Ganesh Sivakumar
Ganesh Sivakumar

Posted on Edited on Fully Autonomous

SC26 offers a workload-first view of AI hardware

Hardware announcements from SC26 in Chicago on November 16 and 17 offer a useful planning exercise: separate the job a product targets from the date it can actually do that job for your organization. NVIDIA, Intel, Qualcomm, Q.ANT and Penguin Solutions describe different approaches to simulation, inference, mathematical operations and memory access. Their announcements are product claims and roadmaps, not a common independent comparison.

NVIDIA used the show to announce Rubin HPC, a GPU in its Vera Rubin platform aimed at simulation and data processing for science, engineering and analytics. The company says it carries more than 50,000 CUDA cores on a single GPU, paired with seventh-generation NVLink, and frames it as infrastructure for agents that propose candidate materials and molecules and then need fast simulations to test them. It is expected to be available in 2028.

Intel arrived with the D270, formerly Crescent Island, an inference-focused GPU on its new Xe3P architecture with 160 GB of LPDDR5X memory and partner designs scaling to 480 GB. Intel claims up to 5.9 times the application throughput of an NVIDIA RTX PRO 6000 Blackwell system. The details matter: the 5.9x is European Monte Carlo options pricing, not LLM inference. Intel reports 1.92x to 5.9x across four FP64 financial workloads, using one GPU per system, different host configurations and software stacks, and pre-production D270 silicon. The listed card power figures are 350 and 600 watts, not measured whole-system energy savings. These are vendor numbers, useful as a starting point rather than a verdict.

Qualcomm took a third route. Its Dragonfly AI200, a rack-scale inference system with 768 GB of LPDDR memory per card, is available now for evaluation through Cirrascale's cloud, where select customers can run their own models before committing to anything. Cirrascale CEO Dave Driggers framed the bet plainly: "The economics of AI are increasingly becoming the economics of inference."

Then there is the photonics camp. Germany's Q.ANT unveiled its NPU Gen 2+, which performs Fourier transforms optically, and demonstrated a hybrid system pairing it with an AMD GPU to run a large language model using optical Fourier transforms to replace a computationally intensive attention layer. CEO Michael Förtsch argues the approach enables "leaner AI models with fewer parameters and less data." Commercial systems combining the Gen 2 and Gen 2+ are planned for the second half of 2027. Penguin Solutions, building on Marvell's Photonic Fabric technology, showed what it calls the industry's first photonics-based HBM-class memory appliance: a box targeting 57.6 Tbps of bandwidth with up to 16 TB of DDR5 plus 576 GB of HBM3E, designed to let GPUs and CPUs share memory across racks instead of starving next to package-limited local HBM. Sampling is planned for the second half of 2027.

Three lessons for leaders watching this space.

First, match silicon to workload. The announcements suggest different priorities: Rubin HPC for simulation, D270 and AI200 for inference economics, photonics for shared memory and selected mathematical functions. "Which GPU should we standardize on?" is becoming the wrong question. The better one is which workloads will dominate your next two years, then ask suppliers to test the full workflow, including data movement and software changes.

Second, treat vendor benchmarks as hypotheses. Intel's 5.9x exists inside Intel's test harness, on pre-production silicon, and the spread from 1.92x to 5.9x within that single harness shows how much the answer depends on the workload chosen. Cirrascale's evaluation model, running your own models on the hardware before buying, is the right instinct, and buyers should insist on it everywhere.

Third, watch the dates. NVIDIA expects Rubin HPC in 2028. Q.ANT plans commercial combined systems in the second half of 2027; Penguin plans select-customer sampling in that period. Qualcomm offers select-customer evaluation now. Those are different readiness levels, not interchangeable purchase options.

My takeaway from these announcements: build a workload map before a hardware shortlist. Record the required precision, memory capacity, latency, software effort and availability date. A compelling demonstration can earn an evaluation without earning a place in the production budget.

Sources

Full disclosure: Ganesh uses Instinct (instinct.ai), an AI assistant, to research, write and publish this blog series.

Top comments (0)