SiFive licences its own AI accelerator, Arm-style, as chip materials squeeze supply

SiFive built its business on RISC-V CPU cores for other people's AI chips. Now it wants to sell the accelerator itself. The Intelligence XM series, announced this week, is a licensable cluster that customers can drop into a processor or system-on-chip and send to a fab, The Register reported on 19 September.

The base cluster is built around four of SiFive's Intelligence X RISC-V CPU cores wired to an in-house matrix math engine. Each cluster supports up to 1TB/sec of memory bandwidth through a coherent hub interface and is rated at up to 16 TOPS of INT8 or 8 teraFLOPS of BF16 per gigahertz. SiFive expects most chips built on the design to run at around 1GHz.

That per-gigahertz figure needs context. This is not a complete chip.

Real performance will depend on how many clusters a customer places, how they are wired internally, what else sits on the die, and the power and cooling available. John Ronco, SVP and GM of SiFive UK, told The Register that most chips based on the design will use between four and eight clusters. At 1GHz, that works out to between 4 and 8TB/sec of peak memory bandwidth and up to 32 to 64 teraFLOPS of BF16. SiFive's product slide deck suggests 512 clusters is within the bounds of possibility, though Ronco was hesitant to say how far the design scales. At 512 clusters and 1GHz, the design would claim roughly four petaFLOPS of BF16 matrix compute. Nvidia's top-specced Blackwell GPUs claim 2.5 petaFLOPS.

Ronco expects the clusters will not be used much for AI training. FLOPS are not everything, especially for bandwidth-constrained inference work, and price, power and process node all matter.

Analysts quoted by Tom's Hardware point to a different constraint on AI accelerator supply: T-glass, a low-CTE glass-fibre cloth used in the organic core of IC substrates. Nittobo, the Japanese company that controls roughly 90% of global supply, is tripling capacity at its Fukushima plant. Tom's Hardware reported on 9 March 2026 that the new supply will not reach the market until mid-2027.

Prices for T-glass have risen 20 to 30%, and lead times for downstream materials such as copper-clad laminates have stretched from a normal 8 to 10 weeks to beyond 20. "With T-glass supply even more constrained now, suppliers are no longer providing lead times," said Bill Ho, an analyst at Yuanta.

Bilal Hachemi, an analyst at Yole Group who tracks the IC substrate supply chain, told Tom's Hardware Premium that replacing the material is not easy. T-glass "has specific dielectric and CTE values that work better for the AI chips, especially for the organic core." He added that the IC substrate industry has historically run on thin margins, so "any increase in demand for build-up materials, ABF material, or T-glass can cause potential shortage, because it's against the basics of this industry."

Nvidia data cited by Tom's Hardware shows interposer sizes growing from 814mm² for Hopper to 1,700mm² for Blackwell, a 109% increase, with Rubin and Feynman scaling further. Bank of America estimates Nittobo's electronic materials segment will nearly double sales from ¥40.9 billion ($266 million) in 2025 to ¥87.7 billion by March 2028, with operating margins approaching 48%. Hachemi said Nvidia reaching out directly to an upstream material supplier was unprecedented.

Elsewhere, TSMC has reportedly cut off shipments to Chinese chip designer Sophgo over allegations it tried to supply components to Huawei in violation of US sanctions, The Register reported on 28 October. The move followed TSMC alerting US officials about a customer order resembling Huawei's Ascend 910B accelerator. Reuters, citing two people familiar with the matter, named Sophgo. The company denied the reports in a Sunday statement, saying it had "never engaged in any direct or indirect business relationship with Huawei" and had submitted a detailed investigation report to TSMC. Bitmain, a Chinese crypto equipment firm associated with Sophgo, also denied involvement, calling the allegations "false and baseless."

Huawei's Ascend 910B is rated at 320 teraFLOPS of FP16 or 640 teraFLOPS of Int8, roughly on par with Nvidia's A100. TSMC stopped doing business with Huawei in 2020 under US Entity List rules.

On the architecture side, Rebellions showed its REBEL-Quad accelerator running at Hot Chips 2025, according to ServeTheHome. The package combines four compute ASICs, four HBM3E sites for 144GB of memory and four integrated silicon capacitors, built on Samsung SF4X and CoWoS-S, and uses UCIe-A as its chiplet interconnect. ServeTheHome reported a live Llama 3.3 70B demo running at 35.5 msec average per output token on a development board.

d-Matrix used Hot Chips 2026 to present Raptor, an accelerator that stacks a TSMC N4 logic die on top of a 3D DRAM die using 36-micron face-to-face stacking. ServeTheHome reported the company's claim that a 1-Hi logic-on-top stack at no more than 0.5 W/mm² can be liquid cooled while keeping DRAM under 100 C. At 32GB per card, with 4-bit weights and an 8-bit KV cache, d-Matrix says a 72-card scale-up can host a frontier model such as Kimi K3 at 1M context.

On the client side, Draw Things said its Metal FlashAttention v2.5 with Neural Accelerators delivers up to 4.6 times better performance on Apple's M5 than on M4, and that an M5 iPad with 16GiB of RAM can generate 480p video in five seconds with Wan 2.2 A14B models. The release is a preview: the company said BF16 support was initially turned off due to unsolved bugs, and that neural-accelerator shaders take 10 seconds or more to specialise on first generation. A later edit noted that build 1.20251117.1 added BF16 and a binary artefacts cache.