Sandisk reveals monstrous 256TB SSD for AI, skipping cache and raising serious questions about speed under pressure
- The SSD of 256 TO of Sandisk completely jumps the cache, which raises concerns concerning the performance of the short -term workload
- The faster speed allegations remain not verified without a public reference or IOP performance numbers
- Direct writing QLC can sacrifice speed in exchange for higher reliability and data integrity
Sandisk announced an SSD of 256 TB, the SN670 Ultraqc, which should be shipped in the first half of 2026.
This model represents the largest SSD ever revealed by the company, marking a daring step towards high density storage solutions adapted to AI and hyperscal infrastructure.
Although the company plans to publish the 128 TB version in the testers in a few weeks, the full commercial availability remains for months.
An architecture built for the scale, not the speed
At the base, the SN670 is built on a NAND BICS 3D 218 layers architecture and has a CBA (CMOS directly linked to the matrix) 2 TB.
It connects via a PCIe Gen5 NVME interface and is part of the new Ultraqlc platform from SanDisk.
Unlike conventional SSDs that stamp data via pseudo-SLC caches, this model uses a “Direct writing QLC” approach.
This simplifies the writing process and makes the reader more sure of the loss of power, but it also introduces compromises, in particular with regard to performance under heavy or short charges.
Without a SLC cache, the SN670 may suffer from slower short -term writings, incoherent performance under load and increased demand from the controller, which makes it less reactive during intensive or unpredictable workloads.
However, Sandisk claims that the SN670 offers more than 68% faster random readings and 55% faster random entries compared to a QLC SSD of 128 TB Gen 5.
Sequential reading speeds are greater than 7%, while sequential writing speeds improve more than 27% in internal comparisons.
Sandisk has emphasized advantages such as scaling of dynamic frequency, which would improve performance up to 10% at the same level of power
He also claims that the data retention profile could reduce the wear of recycling by 33%.
The two features are intended to improve longevity and reduce energy consumption.
However, none of these claims is supported by disclosed performance data such as reading / writing speeds or endurance figures.
Internally, the UltraqLC SN670 is supported by a personalized controller and firmware, which, according to SanDisK, allows better latency and a better bandwidth, but without real reference or IOPS comparisons, these declarations remain projections focused on marketing.
It should be noted the previous iterations of Sandisk business players using QLC NAND showed limitations compared to models based on TLC.
In this case, the native QLC programming latencies could reach 800 to 1200 microseconds, several times slower than SLC -based conceptions.
Sandisk can rely on optimizations such as large Dram pads or advanced parallelism, but these architectural details must still be confirmed.
The final product will arrive in U.2 form initially, with more variants expected later in 2026.
For the moment, the 256 TO of Sandisk’s reader is a symbolic leap to the infrastructure of future data, not a realistic option for users for consumer public.
Via blocks and files
You might also love
