- Solidigm 122.88TB SSD supplied the storage for a check involving Nvidia's Nano Tremendous
- The system was used to run DeepSeek and though it labored, it wasn’t quick
- The Gen 4 PCIe SSD’s pace was restricted by the Nano Tremendous’s Gen 3 connection
On the finish of 2024, Solidigm added a 122.88TB QLC SSD to its product line. The D5-P5336 might be out there in U.2 15mm to begin after which in E1.L later in 2025, which means it received’t slot in a typical shopper PC. Its value is anticipated to exceed $10,000 anyway, so that you’d want deep pockets if you wish to purchase one.
In case you’re questioning how such a giant-capacity SSD would possibly carry out, we’ve got the reply – type of – however it doesn’t come within the type of a standard evaluate.
StorageReviewexamined the Jetson Orin Nano Tremendous – Nvidia’s compact AI single-board laptop for edge computing – to see the way it carried out on AI improvement duties, particularly LLM inference. The Nano Tremendous comes with a 6-core Arm CPU, a 1024-core Ampere GPU, and 8GB of LPDDR5 reminiscence. At $249, it’s an inexpensive selection for AI builders, however its restricted VRAM presents a problem for working LLMs.
Not clean crusing
“We acknowledged that onboard reminiscence limitations problem working fashions with billions of parameters, so we carried out an progressive strategy to bypass these constraints,” the location defined. “Sometimes, the Nano Tremendous’s 8GB of graphics reminiscence restricts its functionality to smaller fashions, however we aimed to run a mannequin 45 occasions bigger than what would historically match.”
Doing this concerned upgrading the Nano Tremendous’s storage with Solidigm’s new U.2 drive, which has a Gen 4 PCIe x4 interface and guarantees sequential learn/write speeds of as much as 7.1 GB/s (learn) and three.3 GB/s (write), together with random efficiency of as much as 1,269,000 IOPS.
The Nano Tremendous has two M.2 NVMe bays, each of which provide a PCIe Gen3 connection. The crew linked the SSD to an 80mm slot supporting a full 4 PCIe lanes utilizing a breakout cable to get probably the most bandwidth and used an ATX energy provide to ship 12V and three.3V to the SSD.
Whereas the complete potential of the drive was restricted by the Jetson’s interface, it nonetheless managed as much as 2.5GB/s of learn speeds. Utilizing AirLLM, which hundreds mannequin layers dynamically somewhat than all of sudden, the location managed to run DeepSeek R1 70B Distilled, an AI mannequin 45 occasions bigger than what would historically match on such a tool.
Signal as much as the TechRadar Professional publication to get all the highest information, opinion, options and steering your small business must succeed!
Processing pace turned out to be a significant bottleneck for the experiment. Operating smaller fashions labored nicely, however producing a single token from the 70B mannequin took 4.5 minutes. Whereas not sensible for real-time AI duties, the check demonstrated how huge storage options, just like the D5-P5336, can allow bigger fashions in constrained environments.
You possibly can see how the check was achieved, and the issues that had been encountered and overcome alongside the best way, on this YouTube video.
YouTube 
Watch On
You may also like
- These are the quickest SSDs you should buy proper now
- And these are the biggest HDDs and SSDs available on the market in the mean time
- Solidigm exits shopper SSD market because it prepares to go large on enterprise SSDs