AMD Cuts Latency by 40% and Doubles Performance with New Solarflare X4 Ethernet Adapters

In networks there are metrics that are read as a technicality and yet they move real money. In electronic markets, every nanosecond can separate an order executed on time from one that arrives late. In this context it appears AMD Solarflare X4a new generation of Ethernet adapters that does not come to change headlines, but to reduce latencies and raise the bar of predictability in environments where consistency is the only currency that matters.

AMD talks about up to 40% less latency compared to previous generations and a 200% jump in system performance over the X2 series. These are figures that, when put into production, are not a whim: they are basis points.

What the X4 really brings (and why it matters)

Beyond the slogan, the card arrives with two clear ideas: less internal path for the packet and better coupling with the CPU. The official product brief explains that the X4 has a Low latency ASIC with PCIe Gen5 x8, PTP synchronization, and two data paths: a Enterprise Data Path (with offloads, multicast, jumbo) and one Express Data Path designed for the least possible delay, where it appears CTPIO (cut-through packet I/O) to push the frame as soon as it is ready, before completing DMA.

There are also scaling details (8K queues, 32K RX descriptors per queue) and versions X4522 (SFP56) and X4542 (QSFP56). Everything points to less “friction” in the transit of the packet through the NIC.

The other leg is software: Onload (POSIX kernel-bypass) and the TCPDirect layer to cut even more latency at the cost of a more specific API. They are veteran players in trading and have been polishing microseconds for years by avoiding the kernel stack. In X4 they continue to be the glue that turns the silicon improvement into a real advantage for the application.

Where it fits: capital markets and any jitter-sensitive flows

AMD boasts (and the sector recognizes it) its presence in 9 of the top 10 bags of the world with Solarflare. It is not just a marketing fact: it means that the ecosystem (drivers, telemetry, support) is used to operating with sub-microseconds and surviving violent spikes in market messages without queues or surprises. For those who execute market data ingestion, order gateways or matching enginesthe promise of 40% lower latency and higher ingestion is not a “nice to have”, it is operating margin and less jitter in the event queue.

Migrate without re-wiring: backward compatibility and ways

A no small advantage: backwards compatibility. The X4 series maintains HHHL formats, supports 1/10/25/40/50/100 GbE and fits into already deployed infrastructure with Onload. The reading is clear: you can plan the refresh in phases (first critical gateways, then feed handlers) without raising the CPD or touching cross-connects. In networks where stopping minutes is unfeasible, this detail is worth more than any laboratory figure.

X4 with EPYC 4005: addition that subtracts latency

The announcement comes from the hand of EPYC 4005 and AMD highlights -12% latency compared to the competition when combining in-house CPUs and NICs. It is no coincidence: extreme network performance relies on CPU-NIC affinity (NUMA, per-core queues, interrupts) and cutting out superfluous hops. Those who already fine-tune with thread pinning, busy-polling and BIOS “low-latency” will find in this couple a sensible starting point.

What changes compared to X2 (and intermediate generations)

The comparison with X2 serves to ground expectations. AMD talks about up to double system performance and more market data ingestion capacity, in addition to the latency reduction already mentioned. In practice, the It’s a more elastic approach than previous generations.

Since Xilinx bought Solarflare (2019) and AMD bought Xilinx (2020), the company has woven a clear narrative: control computing and network in the same solution package. With EPYC + Solarflare + Onload/TCPDirect, AMD sells a coherent stack for those who need deterministic performance, from matching to analytics in near real time. In capital markets they were already in; With X4 they reinforce that site and prepare the jump to other jitter-sensitive niches (telco, tick-to-trade in crypto, in-memory analytics).

Solarflare If your business lives by reacting sooner (and always reacting the same), that is where the improvement is felt. The rest is method: measure, tune and deploy in sections until the p99 says you are on the right track.