Venice adds cores, memory bandwidth and I/O for the work around inference, while AMD's rack-scale claims still await independent validation.
GPUs are creating more work for CPUs.
That is the premise behind AMD's 6th Gen EPYC 9006 series, codenamed Venice. A production AI system does more than generate tokens. It retrieves context, queries databases, calls tools, moves data and schedules work across services. The accelerator handles the model computation. General-purpose processors handle much of what surrounds it.
AMD has organized Venice around that division of labor. The company describes three CPU roles in an agentic data center: running sandboxed agents, hosting accelerators and supporting the enterprise services those agents use. The flagship SP7 family is positioned for all three roles, with up to 256 cores and 512 threads per socket. It also supports up to 16 memory channels, MRDIMM memory at 12,800 MT/s and PCIe 6.0 connectivity, according to AMD's product specifications.
Those specifications make a credible case for more host-side CPU capacity. AMD's performance claims need a more careful reading. Some are projections built from a modeled 100 kW rack. Others come from vendor-run tests on physical systems. They answer different questions and should not be treated as one body of evidence.
AMD announced EPYC 9006 on July 23, but the launch did not put the flagship SP7 family into customers' hands that day. SP7 availability is planned for the fourth quarter of 2026. The briefing also omitted SKU tables, although AMD's public product pages now list model specifications and some pricing. EPYC 9006X, the high-cache family aimed at supercomputing and technical computing, is due in 2027. The LPDDR5X-based EPYC 9006 LP family, codenamed Verano, is planned for the second half of 2027, according to Phoronix's launch report.
The silicon is further along than the sales calendar might suggest. AMD said in May that Venice had entered production ramp in Taiwan and called it the industry's first high-performance computing product to enter production on TSMC's 2 nanometer technology. The company plans a later production ramp at TSMC's Arizona facility. That’s a manufacturing milestone that puts the process claim on firmer ground than the performance projections surrounding the launch. AMD's production announcement supplies the exact boundary.
The initial SP7 family can already fill AMD's three stated roles. Later products specialize the portfolio rather than complete an otherwise missing architecture. EPYC 9006X adds up to three times the L3 cache per core of comparable SP7 parts for memory-sensitive simulation and modeling. EPYC 9006 LP is designed as a host processor for dense accelerator racks. SP8 covers lower core counts for enterprise and power-constrained deployments. AMD lays out those distinctions in its portfolio description.
An agentic pipeline shifts repeatedly between model inference and conventional systems work. A request may pass through a gateway, a planner, a retrieval system and several enterprise applications before the model produces a useful response. Storage, networking, orchestration and security controls add more CPU work around the accelerator.
This is the infrastructure gap SCN identified in its earlier analysis of what agentic AI requires from the data center. The model is one part of a distributed application. Its end-to-end behavior depends on the services around it, which is also why time to first token cannot describe the whole system.
Microsoft provides evidence from a buyer and cloud operator. Its upcoming Azure HDv2 virtual machines will use nearly 500 physical 6th Gen EPYC cores, 4 TB of memory, 32 TB of local NVMe storage and 400 Gb Azure Boost networking. Microsoft says the systems are intended for data preparation, search, reinforcement learning and agent coordination. It also plans HXv2 virtual machines with 176 EPYC cores and 800 Gb InfiniBand for electronic design automation, scientific simulation and other distributed-memory workloads. These are forthcoming services, but their configurations support AMD's broader point: accelerator demand is pulling substantial CPU, memory and I/O capacity into the same infrastructure. Microsoft described both Azure offerings.
AMD's headline rack comparison puts a two-socket EPYC 9996 system at up to 3.3 times the performance of a two-socket NVIDIA Vera system within a modeled 100 kW rack. The figure is a projected geometric mean across six workloads: estimated SPECrate 2017 integer throughput, Java multi-JVM, NGINX web serving, Redis, Memcached and TPROC-C. It is not a measured SPEC CPU 2017 rack result.
AMD combines its own testing, third-party data and projections, then estimates how many two-socket nodes fit inside the same rack power budget. System configuration, software and workload mix can change the result. Vera was not available to AMD for direct comparison. AMD labels the result as preliminary and directional in its rack-performance methodology.
AMD uses a similar abstraction for "agents per watt," "agents per dollar" and "agents per rack." The company defines estimated agent count through available CPU threads under a consistent theoretical workload. That gives the comparison a numerator, but a hardware thread is not a unit of completed agent work. Real capacity varies with the model, memory footprint, tool latency, orchestration software and quality-of-service target. AMD acknowledges those variables in the EPYC 9006 methodology footnotes.
There are independent efforts to benchmark agentic systems. Artificial Analysis, for example, publishes AA-AgentPerf, which measures inference capacity by replaying trajectories from coding agents. Its authors list tool-execution testing as future work. The framework does not reproduce AMD's CPU thread proxy, so it cannot validate the rack claims. It does show why a useful agent metric must connect infrastructure to workload behavior rather than concurrency alone.
AMD separately published results from physical systems, including a preproduction one-socket EPYC 9996 reference platform, an EPYC 9965 system, an Intel Xeon 6980P server and an AWS Graviton5 bare-metal instance. Its suite measures CPU-centric stages such as gateway response, planning and routing, vector search, enterprise databases and enterprise tool workloads. AMD reports a consolidated geometric mean of 2.74 times the Xeon baseline for EPYC 9996, which it expresses as a 174 percent uplift.
That evidence is stronger than a modeled rack because the listed systems ran workloads. It remains vendor-run testing. AMD chose the components, tuned the platforms and normalized results from systems with different configurations. The suite excludes the GPU reasoning stage from the comparison. It also uses self-contained modules so external network and disk bandwidth do not affect measured throughput or latency. Those choices isolate CPU behavior, but they make the result a test of decomposed pipeline stages rather than a production agentic application.
AMD provides unusually detailed configurations and per-tier results in its agentic pipeline benchmark. The data supports a narrower claim than the marketing language: under AMD's controlled, saturated tests, the preproduction 256-core part processed more of the selected CPU-side work than the comparison systems. Independent reproduction will determine how much of that advantage survives different software, memory allocations, network conditions and end-to-end service constraints.
Venice also arrives during a renewed debate about which mathematical work belongs on a CPU. In recent SCN coverage, Jack Dongarra, Torsten Hoefler and Satoshi Matsuoka argued over whether CPUs can reclaim work from GPUs. China's all-CPU LineShine supercomputer, which TOP500 lists at number one, adds a concrete, if specialized, example. Its architecture also serves domestic supply and sovereignty goals, so it should not be reduced to a general verdict on GPU-free computing.
Venice omits an Intel AMX-style tile matrix engine while extending AMD's AVX-512 vector support, which includes FP64 operations used in classical supercomputing. Phoronix's architecture review also notes the addition of BMM. That abbreviation means Bit Matrix Multiply, according to AMD's instruction documentation. The missing tile engine is a choice about specialized matrix throughput. CPU floating-point execution remains.
Nor does the missing tile engine establish AMD's permanent direction for x86. AMD and Intel created the x86 Ecosystem Advisory Group with other industry participants. AMD says the group has accepted ACE, or Advanced Matrix Extensions for Matrix Multiplication, as a standard intended to provide consistent matrix capabilities across x86 devices. Venice can prioritize cores, memory bandwidth and I/O today while AMD participates in a common route toward CPU matrix acceleration. AMD's account of the ACE work makes a permanent rejection thesis hard to sustain.
The two ideas can coexist. Agentic systems may require more CPU capacity beside accelerators, while some simulation and AI workloads move more matrix work onto future CPUs. One increases the amount of host and orchestration work. The other changes which processor performs part of the mathematics.
Venice gives operators more threads, bandwidth and I/O. Software determines whether those resources become useful agent throughput.
AMD's benchmark keeps cores busy with pinned instances and controlled modules. Production agents are less tidy. Requests stall on remote services, permissions, databases and human approval. They branch unpredictably and contend for memory. A scheduler that cannot place that work efficiently can leave a 256-core socket underused even when the surrounding application feels busy.
That is the software ceiling in practical form. The risk is that hardware capacity grows faster than orchestration layers can consistently use it. The same systems problem appears in d-Matrix's acquisition of GigaIO's data-center business, where inference scale depends on memory movement and fabric design as much as arithmetic throughput.
EPYC 9006 therefore makes a plausible architectural case before it makes an independently verified performance case. The specifications fit the work accumulating around GPU inference, and Microsoft's planned systems reinforce the demand signal. AMD's physical tests show that Venice can move selected CPU-side stages quickly under controlled conditions. The modeled rack and agent-capacity figures remain projections.
The next useful evidence will come from production systems that report completed agent work, latency, accelerator utilization and facility power together. Until then, Venice supports the case for a larger CPU role in the agentic data center. How much of that role AMD wins is still a benchmark question.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
