AMD Unveils 2nm 256-Core Zen 6 EPYC Monster to Crush Intel
AMD has just pushed server silicon into territory that seemed like sci-fi standard a few years ago. As artificial intelligence workloads devour power grids worldwide, the semiconductor race has officially hit overdrive.
Production for the server processors, codenamed Venice, is rolling out at TSMC facilities using a 2-nanometer manufacturing process. Moving away from FinFET to GAAFET nanosheets boosts transistor density by 15% while cutting power draw by up to 30%.
Under the hood, AMD split the Zen 6 architecture into standard high-performance cores and dense compute units. The top-tier EPYC 9996 hits 256 cores and 512 threads running at boost speeds up to 4.1 GHz, demanding a beefy 600-watt power supply and a wallet-draining price tag of $14,904.
To feed these processing beasts, AMD updated memory controllers to support 16 channels alongside full support for 512-bit AVX-512 instructions. Dedicated hardware extensions like AVX512_FP16 and AVX_VNNI_INT8 were baked directly into the die to accelerate AI matrix calculations without melting server racks.
On the connectivity front, the platform brings 128 lanes of PCIe Gen 6.0, pushing memory bandwidth to 1.6 TB/s. By comparison, competitor Intel counters with its Xeon 6+ chips featuring 288 cores on the Intel 18A node, though capped at 96 PCIe lanes and 1.3 TB/s bandwidth.
Server rooms are rapidly transforming from computer racks into mini power plants that demand their own electrical substations. When spending fifteen grand on a single chip becomes a bargain for AI training, traditional cloud economics have officially left reality behind.
Source: AMD Newsroom
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.