- Sligigm 122.88 to SSD provided storage for a test involving NVIDIA Nano Super
- The system was used to execute Deepseek and although it worked, it was not fast
- The speed of the Gen 4 PCIe SSD was limited by the Gen 3 connection of the Nano Super
At the end of 2024, Soligm added an QLC SSD of 122.88 TB to its product range. The D5-P5336 will be available in U.2 15 mm to start, then in E1.L later in 2025, which means that it will not fit into a typical general public PC. Its price should exceed $ 10,000 anyway, so you would need deep pockets if you want to buy one.
If you are wondering how a giant capacity SSD could work, we have the answer – in a way – but that does not take the form of a traditional review.
Storagereview Tested the computer with compact monomarous panels from Jetson Orin Nano Super – NVIDIA for IT EDGE – To see how IA development tasks, in particular LLM inference. The Nano Super is delivered with a 6-core CPU, an Ampère 1024-Core GPU and 8 GB of LPDDR5 memory. At $ 249, it is an affordable choice for AI developers, but its limited VRAM presents a challenge for LLM management.
No smooth navigation
“We have recognized that the integrated memory limitations question the execution models with billions of parameters, so we have implemented an innovative approach to bypass these constraints,” said the site. “As a general rule, the 8 GB of graphic memory of the Nano Super restricted its capacity to smaller models, but we have sought to execute a model 45 times larger than what would traditionally suit.”
This consisted in upgrading the storage of the Nano Super with the new U.2 reader from Solidigm, which has a Gen 4 Pcie X4 interface and promises sequential reading / writing speeds up to 7.1 GB / S (reading) and 3.3 GB / S (writing), as well as random performance from up to 1,269,000 IOP.
The Nano Super has two bays NVME M.2, both of which offer a PCIe Gen3 connection. The team connected the SSD to an 80 mm location supporting a full track of four PCIE using a break -up cable to get the most bandwidth and used an ATX power to supply 12 V and 3.3 V to the SSD.
Although the reader’s full potential is limited by the interface of Jetson, it still managed up to 2.5 GB / s of reading speeds. Using Airllm, which loads the models of dynamically rather than all in one time, the site has managed to execute DEEPSEEK R1 70B distilled, an AI 45 times larger model than what would traditionally adapt on such a device.
The treatment speed has proven to be a major bottleneck for experience. The operation of smaller models worked well, but the generation of a unique token from the 70B model took 4.5 minutes. Although it is not practical for real-time AI tasks, the test has demonstrated how massive storage solutions, such as D5-P5336, can allow larger models in forced environments.
You can see how the test was done and the problems encountered and overcome along the way, in this YouTube video.
