News
CAPACITY TEST:

Imec Says AI Inference Pushes Optical I/O Closer To Chips

Newsroom brief

Imec researchers told EE Times that AI inference is shifting the interconnect problem from rack-scale optics towards 2.5D and 3D optical I/O near processors. The group discussed a 250 Tb/s bandwidth projection but did not disclose commercial product timing, customer deployments or cooling test results.

Verified against source materialEdited by SendTech Times Chips & Compute Desk
Imec Says AI Inference Pushes Optical I/O Closer To Chips
Image source: EE Times

Imec researchers say AI inference is pushing the optical I/O roadmap closer to processors, because large models now need faster movement of data between accelerators as well as more compute.

The AI inference optical I/O roadmap described by imec starts with co-packaged optics, but the research group says future systems may need 2.5D and 3D integration to control bandwidth, latency and power.

Imec Projects 250 Tb/s Bandwidth Demand For Future AI Processors

Peter Ossieur, portfolio director at imec and professor at Ghent University, told EE Times that inference needs to be reactive because it is the moment when users interact with a model.

EE Times said those workloads can combine extended prompts, retrieval tools, reasoning chains, multimodal inputs and agent-style systems.

EE Times described scale-up networking as the more difficult domain.

Scale-out networks connect racks and systems across a data centre, while scale-up links GPUs or other accelerators tightly enough that applications can treat them as one larger processor.

EE Times said today's scale-up fabric is often copper-based and mostly constrained to the rack.

It said that model becomes harder as AI systems move towards hundreds, thousands or tens of thousands of accelerators working together.

Co-Packaged Optics Is Not The End Of Imec's Roadmap

Imene Jadli, portfolio manager for optical interconnect at imec, told EE Times that co-packaged optics is a logical next step but not the end of the roadmap.

Co-packaged optics moves optical engines into the package and shortens the electrical path between the optical module and the processor or switch.

The power limit is the main constraint in the figures discussed by imec.

EE Times said one future-system projection used by imec puts processor input and output bandwidth at roughly 250 Tb/s.

It said anticipated co-packaged optics approaches could require about 1.25 kW for the optics alone, on top of processors that may already consume several kilowatts.

Ossieur told EE Times that imec needs an approach that is better than co-packaged optics.

The research group's proposed path is 2.5D optical I/O, where optics is integrated at the interposer or substrate level to shorten the electrical path and lower the power needed to move data.

2.5D Optical I/O Could Cut Optical Power Below 200 W

Ossieur said the same imec projection showed 2.5D optical I/O reducing optical power from about 1.25 kW to below 200 W.

EE Times said imec describes the approach as "wide and slow", using many lanes at moderate speeds instead of a smaller number of very high-speed lanes with heavier signal processing.

The materials work remains broad.

EE Times said co-packaged optics still needs compact, efficient and high-bandwidth devices, including electro-absorption modulators and high-speed photodetectors.

It said imec is exploring barium titanate and III-V compounds alongside silicon photonics because silicon remains attractive for CMOS manufacturing but may not provide every property needed for future optical I/O.

Imec's longer-term endpoint is 3D optical I/O, where optics becomes part of the 3D compute stack.

EE Times said that roadmap could move optical links into advanced packages for processor-to-processor and memory-to-processor data movement, including high-bandwidth memory and stacked processors.

Imec did not disclose a commercial product timetable, named AI-system customers, manufacturing-yield data, optical-coupling results or cooling test results for 3D optical I/O.

Share this article
inXf

Related articles

More
Arm and Supermicro Put Agentic AI Servers to a CPU Test
Chips & Semiconductors

Arm and Supermicro Put Agentic AI Servers to a CPU Test

Supermicro has introduced new server platforms built around Arm’s AGI CPU for inference-heavy and agentic AI workloads across cloud, enterprise and edge deployments. Arm says the AGI CPU includes up to 136 Arm Neoverse V3 cores, 12 DDR5 memory channels running at up to 8800 MT/s and PCIe Gen6 connectivity within a 300W power envelope. The key test is whether operators can use these CPU-heavy designs to add inference capacity without creating new pressure on power and cooling.

Samsung Sets HBM And Server SSD Efficiency Targets With Fab Water Gap Still Open
Chips & Semiconductors

Samsung Sets HBM And Server SSD Efficiency Targets With Fab Water Gap Still Open

Samsung’s 2026 Sustainability Report links lower-power HBM and server SSD targets to AI infrastructure, while its semiconductor division still has a 2050 net-zero timetable and limited product-level commercial detail.

FuriosaAI Starts RNGD Accelerator Deployment At Equinix Lisbon Datacenter
Chips & Semiconductors

FuriosaAI Starts RNGD Accelerator Deployment At Equinix Lisbon Datacenter

FuriosaAI has begun deploying its RNGD AI accelerators at Equinix’s LS2 datacenter in Lisbon as the South Korean chip startup looks for European sovereign AI demand. The company describes 48 GB of HBM3, 1.5 TB/s memory bandwidth and 512 teraFLOPS of dense FP8 performance per card, but its Broadcom-linked next-generation accelerator still depends on HBM4 and HBM4e timing.

FuriosaAI and Broadcom Target the Next Layer of AI Inference Infrastructure
Chips & Semiconductors

FuriosaAI and Broadcom Target the Next Layer of AI Inference Infrastructure

FuriosaAI said it will work with Broadcom on a next-generation AI inference platform built around its TCP architecture and Broadcom networking and packaging technologies. The planned third-generation accelerator will use a 2-nanometer compute die, HBM4/HBM4E memory and multi-die packaging, with sampling planned for the first half of 2028. The deal points to AI infrastructure competition shifting from single-chip performance toward memory, networking, power efficiency and rack-level system design.

Micron 9650 SSD Promises 28GB/Second Reads Before PCIe Gen6 Servers Ship
Chips & Semiconductors

Micron 9650 SSD Promises 28GB/Second Reads Before PCIe Gen6 Servers Ship

Micron’s 9650 enterprise SSD is rated for up to 28GB/second reads and 5.5M random-read IOPS, with liquid cooling expected for the E1.S form factor. Public PCIe Gen6 server platforms were not yet available for a Computex demo.

SK hynix Expert Frames AI Chip Race Around Power, Water And Megafabs
Chips & Semiconductors

SK hynix Expert Frames AI Chip Race Around Power, Water And Megafabs

An SK hynix Newsroom expert column argues that AI semiconductor competition is moving beyond faster chips toward megafab capital, grid delivery, industrial water and memory supply stability.

Keep Reading

More Stories

Latest
UAE AI Infrastructure Draws 29 Percent Of Greenfield FDI Capital SpendingCapital & PolicyJul 9, 2026UAE AI Infrastructure Draws 29 Percent Of Greenfield FDI Capital SpendingUAE Investment Minister Mohamed Hassan Alsuwaidi said communications projects including data centres and AI infrastructure accounted for 29 percent of greenfield FDI capital expenditure in 2025. The ministry report also put inward FDI at $48.3 billion, while named cloud operators, project-level financing and delivery dates were not disclosed.Samsung Starts PM1763 PCIe 6.0 SSD Production For AI ServersChips & SemiconductorsJul 9, 2026Samsung Starts PM1763 PCIe 6.0 SSD Production For AI ServersSamsung has started mass production of the PM1763, a PCIe 6.0 enterprise SSD for AI and HPC servers with 28,400 MB/s sequential reads in a 16TB configuration, according to ServeTheHome. The report did not name customers, pricing, shipment volumes or server qualification dates.Kaspersky Flags AI And Password Risks For Malaysian FirmsCybersecurityJul 9, 2026Kaspersky Flags AI And Password Risks For Malaysian FirmsKaspersky cited Malaysian AI use, weak passwords and hybrid work as workplace cybersecurity risks after recording higher 2025 spyware and backdoor detections. It did not name victim companies, incident costs or confirmed breaches tied to AI-themed lures.AscendEX Shuts Down With User Payouts UnclearFintech & Digital PaymentsJul 9, 2026AscendEX Shuts Down With User Payouts UnclearAscendEX said it ceased operations on July 1 after MiCA authorisation, liquidity and operational pressures, with account access limited to offboarding. The exchange said it could not give assurances on withdrawal timing or amounts, while ZachXBT had flagged nearly empty hot wallets.SpaceXAI Prices Grok 4.5 Below Frontier AI RivalsAIJul 9, 2026SpaceXAI Prices Grok 4.5 Below Frontier AI RivalsSpaceXAI released Grok 4.5 with company pricing of $2 per 1 million input tokens and $6 per 1 million output tokens, below several named frontier AI rivals. The company did not publish independent benchmark validation, customer adoption data or enterprise deployment terms.Meta Faces Backlash Over AI Images From Public Instagram PhotosCapital & PolicyJul 9, 2026Meta Faces Backlash Over AI Images From Public Instagram PhotosMeta is facing criticism after BBC reported that Muse Image can generate AI pictures using public Instagram profile photos for US users. Meta said public-account holders can opt out, while campaigners warned that the setting is separate from normal account privacy controls.Telstra Outage Stops 300 Triple Zero Calls In AustraliaTelco & ConnectivityJul 9, 2026Telstra Outage Stops 300 Triple Zero Calls In AustraliaA Telstra outage in Australia disrupted mobile voice and data services and stopped 300 triple zero emergency calls, according to Light Reading. Telstra said just under 90% of calls and data services had been restored after 12 hours, but the report did not include total affected users, the full systems involved or any new redundancy plan.SpaceX And Orbital File 100,000-Satellite Plans At FCCCloud & Data CentersJul 9, 2026SpaceX And Orbital File 100,000-Satellite Plans At FCCSpaceX and Orbital Compute have filed separate FCC plans for 100,000 satellites each, with SpaceX targeting Starlink Gen 3 broadband and Orbital proposing space-based AI data-centre capacity. The filings do not include approvals, named customers, launch cadence or final commercial timetables.Rapidus Targets 2027 2nm Fab Ramp Without Volume CustomerChips & SemiconductorsJul 9, 2026Rapidus Targets 2027 2nm Fab Ramp Without Volume CustomerRapidus has EUV equipment running at its Chitose fab and plans 2nm mass production in 2027, but Tom's Hardware reported that the Japanese chipmaker still lacks a committed high-volume customer.Sony Wins OCC Conditional Trust Charter For Stablecoin UnitFintech & Digital PaymentsJul 9, 2026Sony Wins OCC Conditional Trust Charter For Stablecoin UnitSony disclosed that the OCC has conditionally approved a U.S. national trust bank for Connectia Trust, a Sony Financial Group subsidiary that would issue and manage dollar-denominated stablecoins. Sony said the unit will be capitalised at $40 million, while trade groups and consumer advocates objected to the charter path.Mastercard Says 77% Of Saudi SMEs Accept Online PaymentsFintech & Digital PaymentsJul 8, 2026Mastercard Says 77% Of Saudi SMEs Accept Online PaymentsMastercard research says 75% of Saudi SME leaders remain confident despite a fall from a 92% benchmark, while 77% accept online payments and the Saudi sample size is not disclosed.Naver And KAI Sign Defence AI MOU Without Weapons-Control TimelineAIJul 8, 2026Naver And KAI Sign Defence AI MOU Without Weapons-Control TimelineNaver, Naver Cloud and KAI signed a defence AI MOU for sovereign models and physical AI combat platforms, but the companies did not disclose budget, deployment timing, test results or weapons-control approvals.