[Talk Proposal] Inside AMD's NTB Stack — Engineering Challenges from Silicon to ntb_perf

[S, Sanath]

Hi All,

I'd like to submit a regular talk proposal (25+5 minutes) for the upcoming BLR Kernel Meetup.

Title: Inside AMD's NTB Stack — Engineering Challenges from Silicon to ntb_perf

Abstract:

PCIe Non-Transparent Bridges let two independent host systems communicate directly over PCIe at wire speed — no network stack, just memory windows, doorbells, and scratchpad registers. Sounds simple, but shipping a production NTB stack on AMD EPYC meant solving hard problems at every layer.

AMD's NTB journey spans a decade — from the original ntb_hw_amd driver in kernel 4.5 through seven EPYC generations to the latest device IDs in kernel 6.18. Along the way, we ran into the kind of problems that only show up when two machines coordinate without a shared OS or a guarantee that both sides are even running at the same time — asymmetric boot races, surprise reboots, a scratchpad negotiation protocol with no hardware atomicity, and an IOMMU bug that silently broke DMA on every IOMMU-enabled system.

This talk walks through those challenges, explains why AMD's NTB is strictly point-to-point — a deliberate design decision, not a PCIe limitation — covers performance tuning with ntb_perf and AE4DMA offload, and shares what we've learned from a decade of evolving this stack across seven EPYC generations.

Outline:

• NTB fundamentals
• The asymmetric lifecycle problem
• Scratchpad negotiation
• Bugs that bit us
• Why AMD's NTB is strictly point-to-point
• Performance tuning

 

Speaker Bio:
Sanath S

I am a Senior Linux kernel developer at AMD, spearheading upstream development across USB4/Thunderbolt, NTB, xHCI, and DMA subsystems. I am the go-to engineer within AMD for all NTB and USB-related upstream activities, and with deep expertise across AMD's platform driver ecosystem — including PMF, I2C, I3C, SFH, UART, and SPI — I own first-level triage of all AMD platform driver issues surfaced by the community. I own feature enablement across all of AMD's client and EPYC platforms, ensuring that every platform driver is enabled and functioning seamlessly from post-silicon through upstream acceptance. Outside of my core subsystem work, I am actively exploring how AI can be leveraged to transform Linux kernel development workflows.

Co-author: Basavaraj Natikar

Looking forward to the meetup!

Thanks,
Sanath S