Samsung and Google Create Hardware-Virtualized SSD for AI

What happened?

Samsung Semiconductor, in collaboration with Google and other key infrastructure players within the NVM Express organization, has ratified the technical standard TP4193, called "PCIe Exported NVM Subsystem Migration." This standard introduces SSD virtualization at the hardware level, moving functions that previously resided in hypervisor software to the SSD controller itself.

Traditionally, storage virtualization was managed through software on the host server, using a method known as "trap-and-emulate." This approach intercepted each command from a virtual machine, hid the real disk identity, and modified instructions, consuming processing cycles and adding latency. With TP4193, the SSD can natively present virtualized storage objects, allowing the host server to act as an orchestrator rather than an executor.

The development of this standard did not happen overnight. NVM Express, the consortium that defines specifications for PCIe-based SSDs, has been working for years to improve virtualization. In 2021, the concept of "Subsystem Local Memory" (SLM) was introduced to allow SSDs to share memory directly with the CPU, but TP4193 goes a step further by allowing the SSD to expose its own virtualized NVM subsystems. The collaboration between Samsung and Google is key: Google, as a hyperscaler, has been one of the main drivers of this technology to optimize its AI data centers, where storage efficiency is critical.

Why is it important?

This change is crucial for AI data centers, where training and inference workloads are highly dynamic and run on GPU clusters. The ability to migrate virtual machines between physical SSDs without the guest operating system noticing changes in the underlying hardware enables more efficient resource management. Additionally, reduced latency and simplified hypervisor improve overall AI application performance.

According to Samsung, the new standard allows virtual machines to directly access SSD administrative queues, eliminating hypervisor overhead. In internal tests, this has reduced I/O latency by up to 30% in AI workloads with high density of random operations, such as those found in recommendation systems or natural language processing. Moreover, secure isolation between tenants becomes native: the SSD ensures that data from one VM is not accessible by another, even if the hypervisor is compromised. This is especially relevant for multi-tenant cloud environments, where security is a growing concern.

Historically, storage virtualization has been a bottleneck. Before TP4193, hypervisors like VMware ESXi or KVM used "trap-and-emulate" techniques that, while functional, added 5-10% overhead in I/O performance. With TP4193, that overhead is reduced to nearly zero, allowing data centers to consolidate more VMs per server without sacrificing performance. This could translate into reduced operational costs for hyperscalers, though it will initially require investment in new hardware.

Market consequences

Since TP4193 requires new capabilities integrated into the SSD controller, existing units cannot be updated via software. This means hyperscalers like Google, who already participated in the standard's development, will have to refresh their storage fleets to reap the efficiency and migration benefits. This refresh, combined with current NAND supply constraints and growing demand for generative AI infrastructure, will put additional upward pressure on enterprise SSD prices.

According to TrendForce, enterprise SSD prices have increased 15-20% year-over-year due to NAND shortages and strong AI demand. With the introduction of TP4193, manufacturers like Samsung, Micron, and Western Digital are likely to launch premium products that justify even higher prices. However, in the long term, competition could moderate prices. The first TP4193-compatible SSDs are expected to hit the market in the second half of 2025, according to Samsung statements. Hyperscalers like Google, AWS, and Azure will be the first to adopt them, followed by large-scale enterprise data centers.

Compared to previous events, such as the transition from HDD to SSD in data centers, adoption of TP4193 will be faster due to the urgency of AI workloads. However, it won't be immediate for everyone: SMEs that don't operate AI data centers will likely wait for prices to drop. Additionally, hypervisor vendors like VMware will need to update their drivers to support the new standard, which could delay adoption in existing environments.

What readers should know

For companies operating AI data centers, TP4193 represents an opportunity to improve efficiency and reduce latency, but it will also involve investments in new hardware. SSD manufacturers and hypervisor providers will need to adapt their products to support the standard. For end users, the impact will be indirect: better cloud services and possibly higher prices in the short term due to infrastructure renewal.

Speculatively, if adoption accelerates, we could see a reduction in cloud costs in the long term, as energy and hardware efficiency translates into lower prices for customers. However, there is no confirmation that this will happen, and it will depend on competition among cloud providers. Additionally, new SSDs with TP4193 could consume less power by reducing hypervisor processing load, which is positive for data center sustainability goals.

"TP4193 marks a milestone in storage virtualization, moving from software solutions to native hardware capabilities. This is especially relevant for AI workloads that require low latency and high availability." — Analyst at TheVortiq