- Requestigm 122.88TB SSD provided storage for a test that involves Nano Super Nano
- The system was used to execute Depseek and, although it worked, it was not fast
- The speed of gene 4 pcie ssd was restricted by the gen 3 connection of Nano Super
At the end of 2024, Soldigm added an QLC SSD of 122.88TB to its product line. The D5-P5336 will be available in U.2 15 mm to begin and then in E1.L later in 2025, which means that it does not fit on a typical consumption PC. Its price is expected to exceed $ 10,000 anyway, so it would need deep pockets if you want to buy one.
If you ask how a giant capacity SSD could work, we have the answer, more or less, but it does not come in the form of a traditional review.
Storagereview I tried the single AI Board computer from Jetson Orin Nano Super – Nvidia for edge computing, to see how it performed in the development tasks of AI, specifically the inference of LLM. The super nano comes with a 6 -core CPU ARM, a 1024 -core amps and 8 GB of LPDDR5 memory. At $ 249, it is an affordable option for AI developers, but its limited vram presents a challenge to execute LLMS.
No soft key
“We recognized that memory limitations on board challenge execution models with billions of parameters, so we implement an innovative approach to avoid these restrictions,” the site explained. “In general, the 8GB of graphic memory of Nano Super restrict its capacity to smaller models, but our goal was to execute a 45 times larger model than would traditionally fit.”
Doing this involved updating the storage of Nano Super with the new U.2 unit of Solidigm, which has a PCIE X4 Gen 4 interface and promises sequential reading/writing speeds of up to 7.1 GB/s (READ) and 3.3 GB/s (writing), together with random performance of up to 1,269,000 IOPS.
The super nano has two m.2 NVME bays, which offer a PCIE GEN3 connection. The equipment connected the SSD to an 80 mm slot that supports four full PCIE lanes using a break cable to obtain the largest amount of bandwidth and used an ATX power supply to deliver 12V and 3.3V to the SSD.
While the complete potential of the unit was limited by the Jetson interface, it still achieved up to 2.5 GB/s of reading speeds. Using Airllm, which loads the layers of the model dynamically instead of all at once, the site managed to run Depseek R1 70b distilled, a 45 -time AI model than what would traditionally fit on that device.
The processing rate turned out to be an important bottleneck for the experiment. Running smaller models worked well, but generating a single token of the 70b model took 4.5 minutes. While it is not practical for real-time AI tasks, the test demonstrated how mass storage solutions, such as D5-P5336, can allow larger models in restricted environments.
You can see how the test was achieved and the problems that were found and overcome along the way, in this YouTube video.
