Trade Ideas
Broadcom demonstrated a 6.4T optical engine with 64 lanes of ~100G PAM4 in a Tomahawk 5 51.2T CPO system, using Fan-Out WLP. The paper shows egress transmitter performance meeting specs, reinforcing B
Broadcom demonstrated a 6.4T optical engine with 64 lanes of ~100G PAM4 in a Tomahawk 5 51.2T CPO system, using Fan-Out WLP. The paper shows egress transmitter performance meeting specs, reinforcing Broadcom's leadership in CPO for scale-out networking. While Broadcom will migrate to COUPE later, this generation proves their packaging and optical engine capability.
Risk: Competition from Nvidia's COUPE and OCI MSA could shift standards; Broadcom's CPO margins may compress as volume ramps.
Samsung's HBM4 demonstrated best-in-class pin speed (13 Gb/s) and aggressive adoption of SF4 logic base die, closing the gap with SK Hynix. However, lower 1c yields (~50%) and higher base die cost (SF
Samsung's HBM4 demonstrated best-in-class pin speed (13 Gb/s) and aggressive adoption of SF4 logic base die, closing the gap with SK Hynix. However, lower 1c yields (~50%) and higher base die cost (SF4 vs N12) pressure margins. The LPDDR6 PHY on SF2 also shows strong efficiency features. Overall, technological progress supports Samsung's HBM market share recovery, but reliability and margin headwinds remain.
Risk: Yield improvement is uncertain; SK Hynix retains reliability advantage and could maintain HBM dominance.
AMD's MI355X improvements detailed at ISSCC include doubling matrix throughput per CU on N3P, reducing IOD count from 4 to 2 (saving area/latency), and custom wire optimization cutting interconnect po
AMD's MI355X improvements detailed at ISSCC include doubling matrix throughput per CU on N3P, reducing IOD count from 4 to 2 (saving area/latency), and custom wire optimization cutting interconnect power ~20%. The TSMC Active LSI test vehicle matches AMD's MI450 design (2 base dies, 12 HBM4 stacks), suggesting early adoption of aLSI for next-gen AI GPUs. These architectural gains reinforce AMD's competitive positioning in AI accelerators.
Risk: Software ecosystem and customer adoption remain key hurdles vs NVIDIA's CUDA; monolithic MI355X still trails NVIDIA's multi-die scaling in aggregate flops.
Intel's UCIe-S die-to-die interface on 22nm achieved 48 Gb/s/lane over 30mm organic substrate, outperforming Cadence's N3E implementation in speed and reach. The presentation is likely a prototype for
Intel's UCIe-S die-to-die interface on 22nm achieved 48 Gb/s/lane over 30mm organic substrate, outperforming Cadence's N3E implementation in speed and reach. The presentation is likely a prototype for Diamond Rapids Xeon, enabling multi-die packaging without advanced interposers. If scaled to Intel 3, efficiency could improve significantly, strengthening Intel's server CPU competitive position.
Risk: 22nm test chip efficiency is behind advanced nodes; adoption in Diamond Rapids requires die shrinks and validation at scale.
SK Hynix unveiled first 1c LPDDR6 and GDDR7, with GDDR7 hitting 48 Gb/s and LPDDR6 density estimated at 0.59 Gb/mm², leading the DRAM density race. Their GDDR7 improvement from 0.309 to 0.412 Gb/mm² i
SK Hynix unveiled first 1c LPDDR6 and GDDR7, with GDDR7 hitting 48 Gb/s and LPDDR6 density estimated at 0.59 Gb/mm², leading the DRAM density race. Their GDDR7 improvement from 0.309 to 0.412 Gb/mm² indicates strong node progression. Despite trailing Samsung in low-voltage LPDDR6 efficiency, SK Hynix maintains overall DRAM technology leadership.
Risk: Samsung's aggressive HBM4 performance could erode SK Hynix's HBM premium; GDDR7 demand may shift as NVIDIA de-emphasizes GDDR7 for Rubin CPX.
Marvell presented an 800G coherent-lite transceiver targeting datacenter campus links up to 40km, using O-band for low dispersion and achieving 3.72 pJ/b (half of full coherent). This product fills a
Marvell presented an 800G coherent-lite transceiver targeting datacenter campus links up to 40km, using O-band for low dispersion and achieving 3.72 pJ/b (half of full coherent). This product fills a gap between direct-detect and long-haul coherent, expanding Marvell's optical portfolio for AI cluster interconnects.
Risk: Coherent-lite is a niche application; adoption depends on hyperscaler campus buildouts and could be bypassed by DWDM solutions.
This newsletter, published April 15, 2026,
features Afzal Ahmad
discussing AVGO, 005930.KS, AMD, INTC, 000660.KS, MRVL.
6 trade ideas extracted by AI with direction and confidence scoring.
Speakers:
Afzal Ahmad
· Tickers:
AVGO,
005930.KS,
AMD,
INTC,
000660.KS,
MRVL