Pascari Branding and Accelerating AI

At FMS 2024, Phison devoted important sales space house to their enterprise / datacenter SSD and PCIe retimer options, along with their consumer products. As a controller / silicon vendor, Phison had traditionally been working with drive companions to convey their options to the market. On the enterprise facet, their tie-up with Seagate for the X1 series (and the next Nytro-branded enterprise SSDs) is kind of well-known. Seagate provided the necessities checklist and had a say within the closing firmware earlier than qualifying the drives themselves for his or her datacenter prospects. Such qualification includes a big useful resource funding that’s attainable solely by giant firms (ruling out a lot of the tier-two client SSD distributors).

Phison had demonstrated the Gen 5 X2 platform ultimately 12 months’s FMS as a continuation of the X1. Nevertheless, with Seagate specializing in its HAMR ramp, and likewise preventing other battles, Phison determined to go forward with the qualification course of for the X2 course of themselves. Within the larger scheme of issues, Phison additionally realized that the white-labeling method to enterprise SSDs was not going to work out in the long term. Consequently, the Pascari model was born (ostensibly to make Phison’s enterprise SSDs extra accessible to finish shoppers).

Beneath the Pascari model, Phison has totally different lineups focusing on totally different use-cases: from high-performance enterprise drives within the X collection besides drives within the B collection. The AI collection is available in variants supporting as much as 100 DWPD (extra on that within the aiDAPTIVE+ subsection beneath).

The D200V Gen 5 took pole place within the displayed drives, due to its main 61.44 TB capability level (a 122.88 TB drive can be being deliberate beneath the identical line). The usage of QLC on this capacity-focused line brings down the sustained sequential write speeds to 2.1 GBps, however these are meant for read-heavy workloads.

The X200, then again, is a Gen 5 eTLC drive boasting as much as 8.7 GBps sequential writes. It is available in read-centric (1 DWPD) and blended workload variants (3 DWPD) in capacities as much as 30.72 TB. The X100 eTLC drive is an evolution of the X1 / Seagate Nytro 5050 platform, albeit with newer NAND and bigger capacities.

These drives include all the same old enterprise options together with power-loss safety, and FIPS certifiability. Although Phison did not promote this particularly, newer NVMe options like versatile knowledge placement ought to turn into a part of the firmware options sooner or later.

100 GBps with Twin HighPoint Rocket 1608 Playing cards and Phison E26 SSDs

Although not strictly an enterprise demo, Phison did have a station exhibiting 100 GBps+ sequential reads and writes utilizing a traditional desktop workstation. The trick was putting in two HighPoint Rocket 1608A add-in cards (every with eight M.2 slots) and inserting the 16 M.2 drives in a RAID 0 configuration.

HighPoint Expertise and Phison have been working collectively to qualify E26-based drives for this use-case, and we will probably be seeing extra on this in a later evaluate.

aiDAPTIV+ Professional Suite for AI Coaching

One of many extra attention-grabbing demonstrations in Phison’s sales space was the aiDAPTIV+ Professional suite. Finally 12 months’s FMS, Phison had demonstrated a 40 DWPD SSD to be used with Chia (fortunately, that fad has light). The corporate has been engaged on the acute endurance side and moved it as much as 60 DWPD (which is normal for the SLC-based cache drives from Micron and Solidigm).

At FMS 2024, the corporate took this SSD and added a middleware layer on prime to make sure that workloads stay extra sequential in nature. This drives up the endurance ranking to 100 DWPD. Now, this middleware layer is definitely a part of their AI coaching suite focusing on small enterprise and medium enterprises who wouldn’t have the funds for a full-fledged DGX workstation, or for on-premises fine-tuning.

Re-training fashions by utilizing these AI SSDs as an extension of the GPU VRAM can ship important TCO advantages for these firms, because the expensive AI training-specific GPUs could be changed with a set of comparatively low-cost off-the-shelf RTX GPUs. This middleware comes with licensing elements which can be primarily tied to the acquisition of the AI-series SSDs (that include Gen 4 x4 interfaces at the moment in both U.2 or M.2 form-factors). The usage of SSDs as a caching layer can allow fine-tuning of fashions with a really giant variety of parameters utilizing a minimal variety of GPUs (not having to make use of them primarily for his or her HBM capability).