The definition of an AI PC is now even muddier, helping no-one – not even AIs The dominant theme at this year's Computex conf in Taiwan is that tens of millions of "AI PCs" will sell this year, and more the year after. But despite all the enthusiasm, the qualities of an AI PC have become even more uncertain – and clarity is many months away. The Register first became confused by the term "AI PC" earlier …
TOPS is short for a trillion operations per second. In the context of AI, it'll be INT8 or signed byte precision. So 40 TOPS is 40 trillion operations a second using signed byte values. It's a measure of AI performance by systems.
You'll typically see manufacturers say their hardware can hit X TOPS of AI performance, and software and LLM makers say you'll need hardware capable of Y TOPS for this AI application to be useful. You as a user will want X to be greater than Y.
Eg, Microsoft claims a PC needs a minimum of 40 TOPS to run its Copilot+ suite.
I'll define a page for it and link to it.
C.
It’s a pretty arbitrary and irrelevant measure though.
For inference, being able to rapidly traverse a neural network is what is important. So a special purpose LLM oriented ASIC with let’s say 32 tiers of register oriented memory (think d-latch 4-bit registers) but extremely wide with a massively wide multiplex between each tier might have a pathetic theoretical TOPS, but would destroy every alternative general purpose inference engine as it would allow FPGA style representation of a neural network. This would make it probably the most powerful inference engine on earth, but would not meet minimum requirements for any “AI PC”.
We need a measure of “inference operations per second” with a depth and width variable.
Thanks. I had guessed the T might stand for Trillion but then thought "no it couldn't be that many". Inasmuch as I'm familiar with any hardware beyond the user level, I'm totally unfamiliar with AI-specific hardware. Anyway it should be spelled out the first time it's used etc. etc.
The whole point of not waiting is that THEY know this stuff is utter Sh**e, they just want you buy it before YOU realize that.
The fact that they are even talking to Gelsinger about this stuff at all is bad enough, let alone without calling BS on the whole AI PC market.
Much like the clowns that mastered how to red-line the CPU on a high end gaming rig to run a single tab in Google Chrome. The more of their computing they can do on your hardware the less they have to pay for hosting fees. Pretty soon someone will catch the AI kids stealing clock cycles for training other people's workloads, and Simon will get a free article to post.
Here is the quiet part out loud: They expect you to by an expensive and powerful new PC to run a service you subscribe to monthly, but run it on hardware, cooling and electricity you pay for instead of theirs.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
