Sell-side most interesting: Cheers for Robots and Intel, telecom equipment sucks Evercore says we need robots more than ever as labor shrinks, Intel has rising prospects, according to Raymond James, Applied Materials can ride the resurgence of DRAM, says Cowen, and Merrill Lynch
AI HARDWARE
-
Q5 Model Performance with GPU Acceleration Optimization
By
–
I am using Q5, and that's what I would advise anyway, even with a more powerful machine. You get the best compromise. And make sure you turn on GPU acceleration.
-
Google DaTaSeg: Universal Multi-Task Segmentation Model
By
–
Attending #NeurIPS2023? At 12:45pm, visit the Google booth where Xiuye Gu will describe DaTaSeg, a universal multi-dataset, multi-task segmentation model well-suited for memory-limited applications, e.g., for multiple segmentation tasks on mobile devices.
-
AMD MI300x cheaper than Nvidia H100 despite higher manufacturing costs
By
–
Quite insane to think the MI300x (AMD’s new ai chip) costs them 2x to manufacture vs H100 but will be cheaper for us to buy. Nvidia is creaming it!
-
Apple Air lacks dedicated GPU for AI processing
By
–
Air doesn't have a dedicated GPU so that option won't work I think
-
Apple M3 Chip Architecture Designed for LLM Integration
By
–
100%! I purposely got an M3 because I remember when they first presented it, it was clear to me this new architecture was winking at LLM usage. Apple is a master at setting up their systems for changes once they are announced.
-
GPT-5 Hardware Made from Stones: Chip Manufacturing
By
–
Same as gpt-5 hardware actually. Chips literally made out of stones.
-
FPGA ConvNet Implementation for Real-Time Image Recognition
By
–
.
@clmt demonstrating his FPGA implementation of convolutional networks doing real-time image recognition at #NeurIPS 2010.
Yes, before ConvNets were cool, we had them running on FPGAs. -
Batch Size Constraints: GPU SM Saturation and Memory Limits
By
–
For a large enough batch size on a given expert, you'll either saturate the SMs or run out of memory