hotaru251 i mean he aint wrong. These take yrs of development and by time these were designed the idea of agentic likely wasnt even a known thing. In 2 to 3 yrs we'd likely see the first purpose built one. Reply
usertests hotaru251 said: In 2 to 3 yrs we'd likely see the first purpose built one. I wonder what that even looks like. What do AI agents need? Imagine if they did chiplets with 1-2 big/classic cores and 64-128 tiny (LP) cores. Reply
sseemaku 'superchip', a minipc called 'super computer', AI PCs, now agentic cpus. Its not over yet, I bet we will see more jargon Reply
hotaru251 usertests said: Imagine if they did chiplets with 1-2 big/classic cores and 64-128 tiny (LP) cores. i fully expect integrated graphics being removed entirely to make room for more "ai" stuff. (as can only fit so much on a chip and server side its un-needed and gaming have dedicated gpu) Reply
bit_user hotaru251 said: In 2 to 3 yrs we'd likely see the first purpose built one. That's mostly just a matter of tuning. The current generation of server CPUs are tuned to run server workloads like database engines, web servers, file servers. Also, bandwidth needed for extremely large amounts of data movement (1.6TB/s, wow!). usertests said: I wonder what that even looks like. What do AI agents need? Probably, the main difference between the server workloads they targeted and what agents do is that you'd have a larger number of small virtual machines for the agents. That works well on Zen, because each of those VMs should comfortably fit within a CCD and thus you have good L3 latency between all of their threads. usertests said: Imagine if they did chiplets with 1-2 big/classic cores and 64-128 tiny (LP) cores. I'll admit I don't have a good intuition of what architectural approach would best support agentic workloads. I think the way they probably approach that is to amass a large sample of agentic workloads and look at where their bottlenecks are, as well as where the current design is underutilized by them. The fix could be something as simple as going from SMT-2 to SMT-4, or it could be changing up the CCD sizes and partitioning. Probably a large number of smaller tweaks. Whatever they do, I'd bet it won't be too drastic, as the same problem exists as before – development takes so long that you can't accurately predict what will be the dominant workloads and if you over-optimize for what people are doing today, it could come at the expense of what people are doing when the thing launches. Reply
bit_user hotaru251 said: i fully expect integrated graphics being removed entirely to make room for more "ai" stuff. (as can only fit so much on a chip and server side its un-needed and gaming have dedicated gpu) This article is about a server CPU, which have no integrated graphics. AMD does make server processors with a mix of AI/HPC compute and CPU cores. Years ago, they launched the Instinct APU line: https://www.tomshardware.com/pc-components/cpus/amd-revolutionary-exascale-apu-under-the-microscope The "GPU" cores in that aren't actually graphics-oriented. They only do AI and HPC compute. Reply
alan.campbell99 At least these chips will still have uses once 'agentic' stops being a thing, or least as much as it's hyped to be by certain hypergrowth-starved tech companies. Reply
alan.campbell99 bit_user said: This article is about a server CPU, which have no integrated graphics. AMD does make server processors with a mix of AI compute and CPU cores. Years ago, they launched the Instinct APU line: https://www.tomshardware.com/pc-components/cpus/amd-revolutionary-exascale-apu-under-the-microscope The "GPU" cores are not actually graphics-oriented. They only do AI and HPC compute. I'm of the opinion these chips should be called something other than GPUs when they seemingly have nothing to do with graphics in the same way a consumer GPU does. Reply
usertests hotaru251 said: i fully expect integrated graphics being removed entirely to make room for more "ai" stuff. (as can only fit so much on a chip and server side its un-needed and gaming have dedicated gpu) Epyc CPUs already use a different I/O die than consumer chips. There's no integrated graphics on Epyc. I think there might have some talk or a patent about optionally including a CDNA/AI chiplet on an Epyc CPU package, but I haven't seen it materialize. There is a rumor that Zen 6 consumer desktop CPUs will ditch the integrated graphics and possibly add in an NPU. But there are multiple competing rumors about what Zen 6 Olympic Ridge I/O will look like so we need to wait on it. Reply
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/pc-components/cpus/SPONSORED_LINK_URL
- https://www.tomshardware.com/pc-components/cpus/amd-vp-and-chief-zen-architect-rebuke-claims-that-nvidia-and-arm-made-cpus-for-agentic-ai-we-didnt-magically-know-four-years-ago-that-agentic-was-going-to-be-a-thing#main
- https://futureplc.com/privacy-policy/
- $10 programmer saves BIOS-locked Lenovo ThinkPads from the dump
- AMD VP and chief Zen architect rebuke claims that Nvidia and Arm made CPUs for agentic AI — ‘We didn’t magically know four years ago that agentic was going to b
- Grab 32GB DDR5 for $371 in this Newegg RAM bundle with the 9800X3D, $70 less than buying it separately
- Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Spark
- Korea Advances Physical AI Across Industries With NVIDIA
Informational only. No financial advice. Do your own research.