Artificial intelligence (AI) is all over the news these days with OEMs slapping the AI tag on just about anything and everything — almost even on the humble thermal paste.
While most AI models such as ChatGPT, DALL-E, Gemini, Llama, Copilot, etc. rely heavily on the cloud given their stupendously heavy computing requirements, it is possible to run AI inferencing at the edge or client systems as well. This becomes particularly important if you want to train AI models on sensitive data and IP while keeping them on-prem.
Training AI models and inferencing them on the edge is possible but consumer hardware poses a major limitation in running large models and efficiently processing more tokens per Watt. Gigabyte wants to change that and help you "train your own AI on your desk" with the new AI TOP lineup.
AI TOP Utility
Gigabyte says its AI TOP PCs are built on three structures: AI TOP Utility, AI TOP Hardware, and AI TOP Tutor. The goal is to allow scaling from 8/30 billion parameters to 70/236 billion while being able to run the system off a common wall socket.
The AI TOP Utility allows monitoring various system parameters such as CPU, GPU, RAM, and SSD usage while offering a no-code GUI front-end for AI training tasks.
AI TOP Hardware
On the hardware front, AI TOP PCs come with standardized and fully tested hardware right from the motherboard to the PSU. The components are designed to be upgradable as needed and used in a typical home or office setting.
Apropos to this, the company is offering two motherboards tailored for AI use:
- TRX50 Aero D (US$559 on Amazon) AI TOP with dual 10GbE LAN, up to 2 TB DDR5, 4x PCIe Gen 5 x16 slots, and 4x NVMe SSD for use with AMD Threadripper 7000 Series Storm Peak HEDT CPUs
- W790 AI TOP with 2x Thunderbolt 4, up to 2 TB DDR5, 5x PCIe Gen 5 x16 slots, and 6x NVMe SSDs for use with Intel Xeon-W2400 and W3400 Sapphire Rapids HEDT CPUs
For GPUs, Gigabyte AI TOP PCs can make use of the Nvidia GeForce RTX 4070 Ti Super AI TOP 16G, four of which can be combined for a total of 64 GB of VRAM for AI training. You can also opt for the new AMD Radeon Pro W7900 DS AI TOP 48G or the W700 AI TOP 32G and scale them up to four cards for a whopping 192 GB VRAM.
Gigabyte is also offering an AI TOP 100E NVMe SSD with Phison's aiDAPTIVCache that is tailored to run Phison's aiDAPTIV+ model. This SSD is rated to offer up to 7,200/6,500 MB/s sequential read/write speeds and 150x TBW.
AI TOP PCs come fitted with the UD1600PM PG5 80+ Platinum PSU that provides enough power to run four GPUs and is rated to last twice as long as conventional PSUs.
These AI TOP pre-built PCs start from US$6,499 for 8 billion parameters, US$11,999 for 13 billion, US$16,499 for 30 billion, and go all the way up to US$35,999 for 70 billion parameters with a max model size of up to 236 billion.
The first two options offer AMD Ryzen Threadripper 7960X CPUs (US$1,399 on Amazon) and 2x RTX 4070 Ti Super GPUs. The 30B config takes it up to Threadripper 7975X (US$3,899 on Amazon) and 4x RTX 4070 Ti Supers. Finally, the 70B behemoth utilizes the Threadripper 7975X and 4x Radeon Pro W7900 GPUs.
AI TOP Tutor
Source(s)
Gigabyte Press Release