AMD’s Venice-X CPU launches in 2027 with 1152 MB of 3D V-Cache, 96 cores, and 5.15 GHz boost clock
Finally, we have "agentic AI workload performance," which uses actual benchmarks for comparison, despite what the names in the chart above suggest. From left to right, AMD used NGINX, TPCx-AI kit, FAISS, TPC-H and TPC-C, and a replay of a multi-persona agent. For TPC-H and TPC-C, AMD says it derived workloads from those benchmarks, so the results here aren't comparable to published results.
Although looking at benchmark results is always interesting, it doesn't say much in the context of a server deployment, at least at the scale that AMD is targeting. Peak performance is only one of the major factors that go into server deployments, after all, and even then, performance can vary wildly depending on what software you're running and how it's built.
Still, Venice looks impressive, perhaps more so in the gen-on-gen comparison than any competitive comparison. Hopefully that bodes well for AMD's future Zen 6 rollout on consumer desktops, but we'll have to wait until Team Red has more to share before drawing any conclusions on that front.
Follow Tom's Hardware on Google News , or add us as a preferred source , to get our latest news, analysis, & reviews in your feeds.
Jake Roach Social Links Navigation Senior Analyst, CPUs Jake Roach is the Senior CPU Analyst at Tom’s Hardware, writing reviews, news, and features about the latest consumer and workstation processors.
Stomx This fat toy is for the future world "elite" and their personal AI datacenters. As for consumers — I honestly don't know who this processor is even for. Maybe it will end up in servers. Maybe it will run local AI on the CPU: with 2.5x the RDRAM bandwidth of AMD Turin, it could reach ~20 tokens per second on decent 4-bit quantized trillion-parameter models — though even 20 cores would be enough for that… But there is one piece of bad news for consumers: the massive socket size and the larger RAM footprint will most probably kill dual-socket designs within the old extended-ATX form factor. A dual-socket motherboard, if one ever gets built, will look monstrous (just take a look at the first single-socket example, Gigabyte's MD25-CG0) and won't fit into any existing cases. That's fine by me — the design could go case-free — but I'm afraid motherboard manufacturers will be scared off by the smaller demand and simply stop even considering making them. As a single-socket replacement for the previous dual-socket Turin, Venice makes little to no sense (and it's still a C-core CPU, not a P-core one). To keep going single-socket, CPUs will have to push core counts (to 500 and beyond), the same way GPUs are doing. Unfortunately, while GPU designers are bulking up at 2–2.5 kW per GPU, the CPU crowd is still in kindergarten, scared to touch 600 W. With NVIDIA, you buy an 8-GPU motherboard and then populate it with 1, 2, 4, or all 8 GPUs (16 were possible before, too) — the 900 GB/s NVLink is already there. With CPUs all we ever had was a dual-socket design with a barely-there ~40 GB/s, and even that will most probably go extinct. Reply
usertests Stomx said: But exist also one bad news for consumers: its massive socket size and larger RAM area most probably will kill dual socket designs within the previous extended ATX motherboard size. Dual socket motherboard if it will be build will look monstrous (take a look at the first single socket example like the MD25-CG0 from Gigabyte) and do not fit into any existing boxes. It is OK with me, the design could be box-free, but I afraid the manufacturers of motherboards will be scared of smaller demand and will stop making them. As a single socket arrangement instead previous dual socket Turin, Venice has little to no sense (and it is is still C-core not a P-core CPU). To continue with single socket, CPUs have to rise core counts to 500 and beyond same as GPUs doing. To the extent that "consumers" are using Epyc… single socket is cheaper and more accessible. AMD's "Siena" Epyc is single socket only, for example. Venice SP8 doesn't support the full 256 cores, but it also cuts the memory channels down to 8 (16 DIMMs possible). A board could offer only 8 DIMMs for 8 channels, making it much smaller. So I guess you got it right in your first sentence, the 256-core is not for plebeians at all. Not a really big surprise. We're long past the days of the $200 Threadripper special too. 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-shares-first-official-benchmarks-for-epyc-venice-cpus-targets-nvidia-company-claims-256-core-chip-is-more-than-twice-as-fast-as-nvidia-vera-96-core-model-20-percent-faster-per-core#main
- https://www.tomshardware.com/membership
- The exceptional Bambu Lab P1S 3D printer could be yours for just $349 — Fall sale drops the popular beginner-friendly printer to a new low price
- d-Matrix Adopts NVIDIA NVLink Fusion for Rack-Scale XPU Deployment
- Microsoft director called AI scraping ‘the largest theft of labor in human history,’ while OpenAI head brands ChatGPT an ‘existential threat’ to publishers — re
- Chinese state media counters Anthropic's call to put brakes on AI development — paper says move is ‘a response to Chinese competition’
- Perplexity Portable Computer Is Now Available on Windows, Powered by NVIDIA RTX
Informational only. No financial advice. Do your own research.