← sources

MATX IR snapshot

source: ir (ir:MATX)published 2026-09-11 00:51ingested 2026-09-11 00:52

extract

High-throughput chips for LLMs Our goal is to make the best chips physically possible for the large model needs of frontier labs. The [MatX One](/research/series_b) chip delivers higher throughput than any announced product while also matching the best latencies of any products. For training and prefill, it excels on FLOPS; for decode and RL it excels on latency, FLOPS, and long-context support. What we offer - The highest FLOPS/mm2. - Weights are typically in SRAM, for low latency. Allows >2000 output tokens/second for large 100-layer MoE models. - KVs are typically in HBM, to support long context well. - The most scale-up interconnect of any product. - Excellent scale-out interconnect, supporting clusters with hundreds of thousands of chips. - A programming model that gives you direct control over the hardware. Target workloads - Training, RL, inference prefill, inference decode. - Large MoE models work well, as do large dense models. There is no upper limit on model size. - No small models, no convolutions, no recommenders. Investors Jane Street, Situational Awareness LP, Spark Capital, Nat Friedman and Daniel Gross's fund, Triatomic Capital, and [many others](/research/series_b).

retained text

High-throughput chips for LLMs
Our goal is to make the best chips physically possible for the large model needs of frontier labs.
The [MatX One](/research/series_b) chip delivers higher throughput than any announced product while also matching the best latencies of any products. For training and prefill, it excels on FLOPS; for decode and RL it excels on latency, FLOPS, and long-context support.
What we offer
- The highest FLOPS/mm2.
- Weights are typically in SRAM, for low latency. Allows >2000 output tokens/second for large 100-layer MoE models.
- KVs are typically in HBM, to support long context well.
- The most scale-up interconnect of any product.
- Excellent scale-out interconnect, supporting clusters with hundreds of thousands of chips.
- A programming model that gives you direct control over the hardware.
Target workloads
- Training, RL, inference prefill, inference decode.
- Large MoE models work well, as do large dense models. There is no upper limit on model size.
- No small models, no convolutions, no recommenders.
Investors
Jane Street, Situational Awareness LP, Spark Capital, Nat Friedman and Daniel Gross's fund, Triatomic Capital, and [many others](/research/series_b).

document fetched 2026-11-17 02:57