Better inference starts with the right architecture. @AndrewDFeldman, Founder and CEO at @Cerebras, highlights how we are combining pre-fill and decode to deliver high throughput, fast tokens and greater efficiency.

Dear AMD and @LisaSu - Please have a product line on GPUs the same as your gaming X3Ds. Nvidia went all in on AI and their cards are absurdly overpriced. Ps5/Xbox run on AMD. Steam OS is optimized for AMD. X3Ds out perform intel in gaming. You sound develop a gaming specific card that performs extremely well and is not expensive. If you can manage a 1k card that performs on par with the last generation of Nvidias flagship while having a better software tech alternative to DLSS with extremely stable drivers you will win the hearts and loyalty of gamers. You can still make high tech cards for AI, but let those be different than your GPU flagship for gamers. Tag it with gamer specific branding or a new line specifically for gaming.

@andrewdfeldman @cerebras Liked the pre-fill vs decode framing in the video, it makes the wafer-scale efficiency story land quickly. AMD pairing its silicon with Cerebras on inference throughput is a smart signal. Where do you see the biggest bottleneck left in the inference stack?

@andrewdfeldman @cerebras Inference is where architecture stops being a diagram and starts becoming a budget. Fast tokens are great; predictable cost is the adoption unlock.

@andrewdfeldman @cerebras combining prefill and decode isn't news, it's called continuous batching

@andrewdfeldman @cerebras the pre-fill + decode part is what caught my attention.. i use ai tools a lot, and if this means faster responses without the system feeling overloaded! that’s a win i’d actually notice😎