“Did we want to discuss the storage and memory needs for AI PCs?” asked Kingston. Yes, B&F did, and so we interviewed Elliott Jones, who looks after B2B Strategic Marketing at Kingston.
Kingston Technology, headquartered in Fountain Valley, CA, is a global manufacturer of memory devices, SSDs and USB-connected storage devices. It’s the largest independent DRAM module producer, employs more than 3,000 people and has factories and logistics bases in manufacturing and logistics facilities in the United States, United Kingdom, Ireland, Taiwan, and China.
The AI PC is reckoned to need more memory and storage capacity than an average PC because AI workloads are more intensive and demanding than running a spreadsheet or word processor but, apart from a 40+ TOPs (Tera Operations per second) requirement there doesn’t appear to be a basic specification for the device.
Blocks & Files: What is the background to the AI PC development and what components will an AI PC have, in general, that are different from an average PC?
Elliott Jones: Window AI PCs have been around in various forms since late last year, but they are still in a transitional phase. That said, AI PCs offer way more autonomy than their cloud counterparts. For one, they have NPUs connecting with memory vs. the HBM/DRAM memory that their server counterparts offer.
Blocks & Files: Is there a real need for an AI PC or is it just marketing hogwash aiming to boost a PC refresh surge for suppliers?
Elliott Jones: Both — it’s a case of hype and purpose. It’s no secret that OEMs have been trying to push new PC/laptop purchases after a significant drought.
However, that doesn’t mean AI PCs are devoid of value. By design, they offer autonomy and privacy over data and the promise of employee efficiency on a local basis. But beyond current creative, video conferencing, and early Co-Pilot applications, there is much to be found out. This will come, and we will see the benefits — but there are hurdles ahead.
So today, it’s somewhat limited — the hardware appears to be ahead of the software, so there is much to do. That’s why we are advocating short-term upgrades — to buy organizations time to figure out how the tech and market moves.
Blocks & Files: Is there a reference architecture for an AI PC?
Elliott Jones: Not really — we see AI PCs progressing without reference architecture. This is evident in the market today. AI PC laptops offer just 256GB storage with 8GB memory, which is woefully under specced. While we have seen Co-pilot+ come out with minimum specifications — even then they admit these specs will change with the inevitable demands of software applications. In short, if you are buying today, make sure it is over specced and you can upgrade; otherwise, your devices will get found out in a short amount of time.
Blocks & Files: Who should develop it? Will this define a standard AI application set it
will have to support and run?
Elliott Jones: The problem is that ARM CPUs are now competing with x86 CPUs — it’s very much an arms race on TOPs (pun intended). While ARM CPUs appear faster and have more TOPs, this offers no guarantee for the future. AMD/Intel will likely be trading blows between themselves and ARM.
The other challenge for ARM is that it is trying to displace x86 chipsets. For organizations with internal applications, there is no guarantee they will work with ARM CPUs — even with the Prism emulator.
Taking a step back — things will likely converge — certainly with the x86 chipsets.
Blocks & Files: In the absence of an AI PC reference architecture, what can you say about likely AI PC memory
requirements?
Elliott Jones: The biggest shift is whether users or organizations will swallow the inability of not being able to upgrade. With “right to repair” growing in popularity, in turn, it suggests “right to upgrade” will be next. At Kingston, we anticipate CAMM2 (Compression-Attached Memory Module) becoming a potential player in this discussion. Some OEMs will elect to have soldered memory. But they are exchanging short term wins –the benefits of selling to those who are unsure – with an upgradable device.
Blocks & Files:In the absence of an AI PC RA, what can you say about likely AI PC storage requirements?
Elliott Jones: In the past few days, I have heard reports that OEMs are using cheaper DRAM-less storage, and the existing native AI PC applications are suffering as a consequence. This indicates a misunderstanding or confusion in how AI PCs could or should work, not only today – but also in the future. If major challenges are being realized today – it offers a scathing view of what will happen in the medium and long term – as organizations battle with their refresh objectives.
Blocks & Files: Do you think an AI PC reference architecture will emerge?
Elliott Jones: A common ISV portfolio would have been the answer – but that view changed since the inclusion of the ARM chipset with Windows 11. We have now a potential Android vs iOS moment – bigger ISVs can cope with developing for both platforms, while the smaller ISVs will have to pick a route. Factor in that Windows 10 will become EOS in October next year, and organizations will find themselves in a pinch. For AI PC RA to be realized – it requires common ground – but this might not be likely in the short term due to x86 and ARM platforms competing for market share.