Accelerated Computing
FIELD GUIDE
§ 09
The wire

News

Significant moves in accelerated computing — flagship hardware launches, notable benchmarks, and architecture announcements. Curated from vendor sources.

Hardware

Intel details Crescent Island data center GPU with 480 GB of LPDDR5XIntel's Xe3P inference GPU carries up to 480 GB of LPDDR5X in a 350W air-cooled PCIe card and adds native FP4 and MXFP4 support, with sampling in H2 2026.

Intel · 2 days ago4 min →Edge

NVIDIA unveils RTX Spark, a Grace-plus-Blackwell superchip with 128 GB unified memoryNVIDIA's RTX Spark pairs a 20-core Grace CPU with a 6,144-core Blackwell GPU over NVLink-C2C, claiming 1 petaflop of AI compute for on-device agents.

NVIDIA · 3 days ago4 min →Hardware

Phoronix benchmarks NVIDIA Vera CPU against Intel and AMD x86 flagshipsFirst independent Vera review shows a 1.5x geomean lead over a 128-core x86 processor and a 1.6x gain over Grace, with 1.2 TB/s of LPDDR5X bandwidth.

NVIDIA · 1 week ago4 min →Hardware

AMD launches Instinct MI350P PCIe with 144GB HBM3EThe MI350P drops AMD's MI350-class compute into a dual-slot PCIe card aimed at on-prem inference, claiming up to 4,600 TFLOPS at MXFP4.

AMD · 4 weeks ago4 min →Hardware

AMD previews Instinct MI430X with 200+ FP64 TFLOPSAMD projects its upcoming HPC accelerator at over 200 FP64 TFLOPS, more than six times its claim for NVIDIA Rubin.

AMD · 4 weeks ago4 min →Hardware

Tenstorrent ships Galaxy Blackhole, its 32-chip AI serverEach 6U Galaxy node packs 32 Blackhole accelerators, 1 TB of GDDR6, and 23 PFLOPS of FP8, listing at $110,000 per system.

Tenstorrent · 1 month ago4 min →Hardware

Google reveals TPU 8t and TPU 8i, splitting training and inference at 2nmGoogle's eighth-generation TPUs ship as two chips: a training part that scales to 9,600 in a superpod, and an inference part with 3x the on-chip SRAM.

Google Cloud · 1 month ago5 min →Hardware

AMD Instinct MI355X reaches NVIDIA B200 parity in MLPerf Inference v6.0MI355X tied B200 on Llama 2 70B Offline, beat it 19% on Interactive, and crossed one million tokens per second at cluster scale.

AMD · 2 months ago4 min →Hardware

NVIDIA puts the Vera Rubin platform into production with seven new chipsVR200 GPU pairs 288 GB of HBM4 with 50 PFLOPS of FP4 compute, and the Vera Rubin NVL72 rack claims a 3.3x throughput jump over Blackwell Ultra B300.

NVIDIA · 3 months ago5 min →Hardware

NVIDIA announces Space-1 Vera Rubin Module for orbital data centersSpace-1 packages Rubin-class silicon into a flight-rated module and pairs it with IGX Thor and Jetson Orin for AI inference on orbit.

NVIDIA · 3 months ago4 min →Hardware

Apple debuts M5 Pro and M5 Max with multi-die Fusion ArchitectureNew MacBook Pro silicon splits CPU and GPU across two 3 nm dies and adds a Neural Accelerator to every GPU core, claiming 4x AI performance over M4.

Apple · 3 months ago4 min →