<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb" -->

---
title: Z2RAM Layers Promise Denser SRAM Memory for AI | daily.dev
description: AI inference is increasingly limited by power, driving a wave of optical-computing startups aiming to speed up or shrink the energy cost of matrix-vector...
canonical: https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Z2RAM Layers Promise Denser SRAM Memory for AI | daily.dev
og:description: AI inference is increasingly limited by power, driving a wave of optical-computing startups aiming to speed up or shrink the energy cost of matrix-vector...
og:url: https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb
og:image: https://api.daily.dev/og/posts/kg2zXmHEb.png
og:image:alt: Z2RAM Layers Promise Denser SRAM Memory for AI
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Z2RAM Layers Promise Denser SRAM Memory for AI

**[IEEE Spectrum](https://daily.dev/sources/ieeespectrum)** · 18 min read · 0 upvotes · 0 comments

## Summary

AI inference is increasingly limited by power, driving a wave of optical-computing startups aiming to speed up or shrink the energy cost of matrix-vector multiplication. Companies include Lumai, Neurophos, Q.ANT, Lightelligence, LightSolver, and Optalysys, each taking different approaches: free-space optics, on-chip metamaterials, integrated photonics, or optical processing inside network paths. Neurophos is targeting 50 times Blackwell B200's speed and 35 times its energy efficiency, with a roadmap to 250 times speed and 90 times efficiency. Lumai's upcoming Iris Aura server aims to cut prefill energy to one-tenth of a GPU's. Researchers caution that Amdahl's law and data-encoding costs limit real-world gains, with optical-computing companies aiming for a 10- to 80-fold efficiency improvement over GPUs.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://spectrum.ieee.org/sram-memory-z2ram>

## Questions this post answers

### How much faster does Neurophos claim its optical chip is compared to Nvidia's Blackwell B200 GPU?

Neurophos is targeting 50 times the speed of Nvidia's Blackwell B200 GPU and 35 times its energy efficiency, according to CEO Patrick Bowen. The company has a roadmap to reach 250 times the speed and roughly 90 times the energy efficiency in subsequent generations. Its modulators form a 1,000 by 1,000 array on a 5mm chip and switch over 10,000 times faster than standard liquid-crystal spatial light modulators.

_Anyone weighing optical accelerators against GPUs for AI inference can track these claims on daily.dev._

### Why hasn't optical computing replaced GPUs for AI inference yet despite its theoretical energy advantages?

Optical multiplication is passive and nearly energy-free once a matrix is set, scaling energy use with vector length N rather than N squared like electronics. But encoding data into light and reading it out costs energy too, and Peter McMahon notes electronics still wins below roughly 1,000 to 10,000 vector elements. Amdahl's law also limits gains since inference includes nonlinear operations optics can't accelerate, capping realistic improvements at 10- to 80-fold over GPUs.

_Developers evaluating emerging AI hardware can follow these efficiency tradeoffs on daily.dev._

### What is the difference between the prefill and decode phases of AI inference and why does it matter for hardware design?

Prefill processes the input query and propagates it through the model using repeated vector-matrix multiplications against the same matrix, making it compute-intensive. Decode generates output one token at a time and is memory-bound, with the main cost being reading model data in and out of memory rather than the multiplication itself. Companies like Lumai target prefill, while Optalysys targets decode by adding processing into the memory communication path.

_Engineers optimizing inference pipelines can keep up with hardware approaches to each phase via daily.dev._

## Similar posts on daily.dev

- [Robot Videos: Morphing Robots, Dancing Humanoids, More](https://daily.dev/posts/robot-videos-morphing-robots-dancing-humanoids-more-dl8vcudbk) · IEEE Spectrum · 0 upvotes · 0 comments
- [From invisibility cloaks to AI chips: Neurophos raises $110M to build tiny optical processors for inferencing](https://daily.dev/posts/from-invisibility-cloaks-to-ai-chips-neurophos-raises-110m-to-build-tiny-optical-processors-for-in-7oo0cb60y) · TechCrunch · 0 upvotes · 0 comments
- [UK’s ARIA will test Lumai’s light-based AI servers in a data centre](https://daily.dev/posts/uk-s-aria-will-test-lumai-s-light-based-ai-servers-in-a-data-centre-evaz2ameo) · The Next Web · 0 upvotes · 0 comments

---

Tags: [#ai-inference](https://daily.dev/tags/ai-inference)

[View this post on daily.dev](https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Z2RAM Layers Promise Denser SRAM Memory for AI","url":"https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb"},"datePublished":"2026-11-21T14:02:33.312Z","dateModified":"2026-11-21T14:03:05.387Z","description":"AI inference is increasingly limited by power, driving a wave of optical-computing startups aiming to speed up or shrink the energy cost of matrix-vector...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/af3d3b7637e288d4c3d27ff2f7e0c93e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/af3d3b7637e288d4c3d27ff2f7e0c93e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"IEEE Spectrum","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"IEEE Spectrum","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/b123f5e8420e492ab36f6adb30c1793a","url":"https://daily.dev/sources/ieeespectrum"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-inference","timeRequired":"PT18M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"IEEE Spectrum","item":"https://daily.dev/sources/ieeespectrum"},{"@type":"ListItem","position":3,"name":"Z2RAM Layers Promise Denser SRAM Memory for AI"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/z2ram-layers-promise-denser-sram-memory-for-ai-kg2zxmheb#faq","mainEntity":[{"@type":"Question","name":"How much faster does Neurophos claim its optical chip is compared to Nvidia's Blackwell B200 GPU?","acceptedAnswer":{"@type":"Answer","text":"Neurophos is targeting 50 times the speed of Nvidia's Blackwell B200 GPU and 35 times its energy efficiency, according to CEO Patrick Bowen. The company has a roadmap to reach 250 times the speed and roughly 90 times the energy efficiency in subsequent generations. Its modulators form a 1,000 by 1,000 array on a 5mm chip and switch over 10,000 times faster than standard liquid-crystal spatial light modulators. Anyone weighing optical accelerators against GPUs for AI inference can track these claims on daily.dev."}},{"@type":"Question","name":"Why hasn't optical computing replaced GPUs for AI inference yet despite its theoretical energy advantages?","acceptedAnswer":{"@type":"Answer","text":"Optical multiplication is passive and nearly energy-free once a matrix is set, scaling energy use with vector length N rather than N squared like electronics. But encoding data into light and reading it out costs energy too, and Peter McMahon notes electronics still wins below roughly 1,000 to 10,000 vector elements. Amdahl's law also limits gains since inference includes nonlinear operations optics can't accelerate, capping realistic improvements at 10- to 80-fold over GPUs. Developers evaluating emerging AI hardware can follow these efficiency tradeoffs on daily.dev."}},{"@type":"Question","name":"What is the difference between the prefill and decode phases of AI inference and why does it matter for hardware design?","acceptedAnswer":{"@type":"Answer","text":"Prefill processes the input query and propagates it through the model using repeated vector-matrix multiplications against the same matrix, making it compute-intensive. Decode generates output one token at a time and is memory-bound, with the main cost being reading model data in and out of memory rather than the multiplication itself. Companies like Lumai target prefill, while Optalysys targets decode by adding processing into the memory communication path. Engineers optimizing inference pipelines can keep up with hardware approaches to each phase via daily.dev."}}]}
```

