At the intersection of silicon and intelligence, SK hynix and SanDisk have jointly proposed a new grammar for how machines remember and retrieve — the High Bandwidth Flash standard, unveiled at FMS 2026 under the Open Compute Project's stewardship. Targeting data transfer speeds of up to 3 terabytes per second, the specification confronts one of AI's most quietly consequential constraints: the gap between what processors can think and how fast memory can feed them. In standardizing this bridge, two of the industry's largest manufacturers are not merely announcing a product — they are sketching
SK hynix and SanDisk Unveil High Bandwidth Flash Standard to Accelerate AI Inference
The bottleneck that has constrained AI inference finally has a solution
Mark
Why does the speed of memory matter so much for AI? Can't the GPU just wait a moment for data?
Mimi
It can wait, but every moment it waits is a moment it's not doing useful work. In AI inference, you're running billions of calculations per second. If the GPU has to sit idle because data isn't arriving fast enough, you're burning power and time on nothing. Scale that across a data center running thousands of requests, and the inefficiency becomes enormous.
Mark
So this new standard is basically making the pipe between memory and processor much wider?
Mimi
Exactly. Wider and faster. Right now that pipe is relatively narrow—designed for storage tasks, not for feeding a ravenous processor. The HBF standard redesigns the connection from the ground up to handle the kind of throughput AI systems actually need.
Mark
Why did it take SK hynix and SanDisk working together to solve this? Couldn't one company have done it alone?
Mimi
They could have, but then you'd have a proprietary solution that only works with their hardware. By establishing an open standard through the Open Compute Project, they're creating an ecosystem. Other manufacturers can build compatible products. Data centers can mix and match. That's what makes it actually transformative rather than just a clever engineering trick.
Mark
What's the risk here? Is there a downside to this much bandwidth?
Mimi
The engineering challenges are real—you're talking about moving data at speeds that create heat, power consumption issues, and signal integrity problems. But those are solvable. The bigger question is whether the industry will actually adopt it, and whether the performance gains justify the cost of new infrastructure.
Mark
When will we actually see this in data centers?
Mimi
That's unclear. The specification is released now, but manufacturing and integration take time. You're probably looking at a year or more before products start shipping, and longer before they're widely deployed. But the fact that it's been standardized means it's coming.
Il Polso
- AI inference workloads are being quietly strangled by a memory bottleneck — GPUs sit idle, waiting for data that current flash architectures were never designed to deliver at this speed.
- The gap is stark: today's best NVMe drives peak near 14 gigabytes per second, while the HBF standard targets 3 terabytes per second — roughly 200 times faster, a difference that redraws the map of what's computationally possible.
- SK hynix and SanDisk are not just building faster chips — they are establishing a common technical language through OCP specifications, betting that standardization will pull the rest of the industry into alignment.
- The ambition extends beyond speed: HBF could allow GPU memory pools to scale into multiple terabytes, fundamentally expanding how much data AI systems can access without latency penalties.
- Adoption remains the open question — GPU designers, competing memory manufacturers, and data center operators must each choose to follow the blueprint before the standard becomes reality.
At the intersection of silicon and intelligence, SK hynix and SanDisk have jointly proposed a new grammar for how machines remember and retrieve — the High Bandwidth Flash standard, unveiled at FMS 2026 under the Open Compute Project's stewardship. Targeting data transfer speeds of up to 3 terabytes per second, the specification confronts one of AI's most quietly consequential constraints: the gap between what processors can think and how fast memory can feed them. In standardizing this bridge, two of the industry's largest manufacturers are not merely announcing a product — they are sketching the architecture of the next era of artificial intelligence.
Two of the world's largest memory manufacturers have jointly released a technical specification for a new kind of flash memory designed to address one of artificial intelligence's most pressing constraints. SK hynix and SanDisk announced the High Bandwidth Flash standard at FMS 2026, backed by the Open Compute Project, targeting data transfer speeds of up to 3 terabytes per second.
The problem is both real and urgent. As AI models grow larger, the systems running them face a critical chokepoint: getting data from memory into the GPU fast enough to keep the processor working. Current flash architectures were built for traditional storage — files, databases, archives — not for the relentless throughput demands of AI inference. The result is idle GPU time, wasted computational power, and slower responses.
The HBF standard represents a fundamental rethinking of how flash memory connects to processors. By releasing a common specification through the Open Compute Project, SK hynix and SanDisk are creating a path for the broader industry to build compatible hardware — enabling data center operators to mix components freely and manufacturers to invest in production with confidence. At 3TB/s, the standard would be roughly 200 times faster than today's best NVMe drives, potentially allowing GPU memory pools to expand to multiple terabytes.
The collaboration carries weight precisely because of who is involved. SK hynix is among the world's largest DRAM and NAND producers; SanDisk commands significant enterprise storage market share. When manufacturers of this scale align on a standard, the industry tends to follow. The specification is now public, the technical roadmap is clear, and the bottleneck constraining AI inference performance may finally have a credible answer — if adoption follows ambition.
Two of the world's largest memory manufacturers have jointly released a technical specification for a new kind of flash memory designed to solve a growing problem in artificial intelligence: the bottleneck between processors and storage. SK hynix and SanDisk announced the High Bandwidth Flash standard at FMS 2026, a major industry conference, with the backing of the Open Compute Project—a consortium that sets standards for data center hardware. The new standard targets data transfer speeds of up to 3 terabytes per second, a dramatic leap from current capabilities.
The problem they're addressing is real and increasingly urgent. As AI models grow larger and more complex, the systems that run them face a critical constraint: getting data from memory into the GPU fast enough to keep the processor fed. Current flash memory architectures weren't designed with this kind of throughput in mind. They were built for traditional storage tasks—holding files, databases, archives—where speed matters less than capacity and cost. But AI inference, the process of running a trained model to generate predictions or responses, demands something different. The GPU sits idle waiting for data, and that idle time translates directly into wasted computational power and slower responses.
The High Bandwidth Flash standard represents a fundamental rethinking of how flash memory connects to processors. By establishing a common technical specification, SK hynix and SanDisk are creating a path for the entire industry to build compatible hardware. This matters because standardization drives adoption. Manufacturers can invest in production knowing their products will work across different systems. Data center operators can mix and match components without worrying about compatibility. The specification released by the Open Compute Project provides the technical blueprint that others can follow.
What makes this announcement significant is the scale of the ambition. Three terabytes per second is not a modest improvement—it's a fundamental expansion of what's possible. To put that in perspective, current high-end NVMe drives max out around 14 gigabytes per second. The new standard would be roughly 200 times faster. That kind of bandwidth could allow GPU memory pools to expand to multiple terabytes, effectively giving AI systems access to vastly more data without the latency penalties that come with traditional storage hierarchies.
The collaboration between SK hynix and SanDisk signals that this isn't a niche innovation. These are companies with the manufacturing scale and market position to make something like this real. SK hynix is one of the world's largest producers of DRAM and NAND flash. SanDisk, owned by Western Digital, is a major player in consumer and enterprise storage. When companies of this caliber align on a standard, it typically means the industry is moving in that direction.
The timing is also telling. AI infrastructure is becoming a central concern for every major technology company and data center operator. The race to build faster, more efficient AI systems is intensifying. Any technology that can reduce the bottlenecks between memory and processor becomes strategically important. By releasing the specification now, SK hynix and SanDisk are positioning themselves at the center of the next generation of AI hardware architecture.
What happens next will depend on adoption. Other memory manufacturers will need to decide whether to build products that conform to the HBF standard. GPU makers will need to design systems that can take advantage of the new bandwidth. Data center operators will need to see real performance gains in their workloads before committing to new infrastructure. But the specification is now public, the technical roadmap is clear, and two major manufacturers have committed to the vision. The bottleneck that has constrained AI inference performance may finally have a solution.
Citazioni salienti
The specification is now public, the technical roadmap is clear, and two major manufacturers have committed to the vision— Industry analysis
Vuoi la storia completa?
Leggi l'originale su Google News ↗
Technology