AMD 為智能體 AI 調校「Venice」Epyc 處理器
AMD Tunes Up CPUs For Agentic AI With 'Venice' Epyc Processors
AMD 表示,Venice Epyc 9006 第 6 代處理器適合智能體 AI 工作負載,吞吐量達前代 Turin Epyc 9005 的 1.9X。處理器最高速度由 Turin 的 5.0 GHz、64 核心提升至 5.3 GHz、96 核心,記憶體頻寬則由 614 GB/sec 增至 1.6 TB/sec。
The introduction of the ChatGPT chatbot by OpenAI in late 2022 supercharged the launch of of the GenAI era that shows only a few signs of slowing down, at least in the near term.
The Democratic “Blue Wave” in the recent midterm elections in the United States, which could introduce more regulation for the AI industry and become a drag on the datacenter building boom around the country, could tap the brakes, and the latest scandal involving Nvidia co-founder and chief executive officer Jensen Huang could be a blip on the screen or a longer-term issue. It’s hard to tell how this will shake out.
NEXTPLATFORM AD
That said, the emerging agentic AI market is forecast to expand rapidly, with some analyst firms calling for it to grow almost 43 percent annually between now and 2034. Agentic AI also will drive a shift within datacenter infrastructure. Over the past four years, the various waves of generative AI like reasoning and mixture-of-experts (MoE) models have been a boon for GPU makers – none more than Nvidia, which has seen its valuation soar into the trillions – and pushed the CPU, long the processor of choice for datacenter operators, into a supporting role.
But the computation demands of agentic AI are different from those of chattybot GenAI models froma few years back and require the capabilities that CPUs bring to the table for both AI host processors and AI sandboxes that run code generated by AI models.
In the agentic AI era, “the shape of the workload changes entirely,” Dan McNamara, senior vice president and general manager for compute and enterprise AI at AMD, wrote in May. “Instead of answering one prompt, an agent breaks a goal into steps, decides what to do next, calls multiple models, queries databases, connects with APIs, runs enterprise applications, checks permissions, retrieves memory, validates output and then loops back through the process again. It’s a very different infrastructure profile than prompt-in-answer-out chatbot AI.”
CPUs are better than GPUs at orchestrating the various complex AI agent tasks, as well as policy, security, and agent execution. That promises to change what’s needed in datacenter systems. Right now, the ratio of GPUs to CPUs is somewhere between 1:4 and 1:8. The expectation is that with agentic AI, that ratio will be brought to 1:1, more comfortable numbers for AMD, Intel, and other CPU makers.
NEXTPLATFORM AD
That is a point AMD executives are making this week as the give more details about the 6th Gen Epyc CPUs – formerly known as “Venice” – and the 6th generation of the Zen microarchitecture. As Mike Clark, Corporate Fellow, senior vice president, and lead Zen architect at AMD, told journalists and analysts during recent briefing, “we didn't have a foresight that agentic was going to be what it has become, but we have always believed that the Core Complex Die was going to play a critical role in AI. We did a big investment in Zen 5 to do that, but with Zen 6, we continued to invest down that path and enhance our AI capability and it's paying off as agentic has taken off.”
As we’ve written before, agentic AI doesn’t mean that AMD will be able to chase down Nvidia in the still booming GPU market, but the company will continue to fair well against long-time CPU rival Intel – AMD’s share of the datacenter CPU market is about 46 percent – and most importantly, make a lot of money doing so.
The message from Madhu Rangarajan, corporate vice president for compute and enterprise AI products, during the call with journalists and analysts, was that the developments in the “Venice” sixth generation Epyc X86 CPUs that make them fit so well with the demands of agentic AI were a natural progression in the evolution of the server CPUs over the past eight years, since the introduction of the first generation “Naples” chips in 2018.
With the fourth generation “Genoa” CPUs in 2022, AMD began to expand its roadmap to serve multiple datacenter market segments, Rangarajan said.
NEXTPLATFORM AD
“So with your Genoa and your Turin and your Bergamo and your Siena, you had your optimized design points going after different sub-segments of the market,” he said. “That's where we really started growing up to have a full portfolio of products here, and continued that.”
Rangarajan echoed Clark, saying he couldn’t claim that AMD purpose-built the Venice CPU family for agentic AI, but said the chips fit well with the needs of AI agents because they spend a broad set of workloads, adding that “we really now have a portfolio ready for the agentic AI era. We’ve got 1,000-plus instances, and what that allows customers to do is get a consistent offering, no matter how they want it to fly.”
With the Venice Epyc 9006 series chips, AMD is essentially doubling the throughput from one generation to the next, with the Venice chips delivering 1.9X the throughput of the prior “Turin” Epyc 9005 processors. The CPUs also measure up to throughput and per-core performance when compared to Nvidia’s Vera, Intel’s 6th Gen Xeon 6980P, and Arm’s AGI processors.
Here’s the first comparison to Nvidia Vera:
NEXTPLATFORM AD
And here is the comparison to Intel and Arm CPUs:
AMD is increasing the speed of the Venice chips, from 5.0 GHz with 64 cores with Turin to 5.3 GHz with 96 cores. Memory bandwidth jumps from 12 channels of DDR5 at 614 GB/sec to 16 channels at 1.6 TB/sec while I/O bandwidth doubles from PCI-Express Gen 5 to Gen 6, too.
Concurrent agents is an important measure these days, according to Rangarajan. AMD’s 6th Gen 256-core chip at 400 watts offers 3.3 times the agents-per-watt that Nvidia’s Vera does, 1.8 times the agents-per-dollar compared with Intel’s 6th Gen Xeons, and 1.5 times the agents per Helion AI Rack than fit in a Vera rack.
The Epyc 9006 chips also deliver better density and efficiency than Vera, according to AMD.
As AMD pushes its latest generation of CPUs into a hot agentic AI market, such numbers matter as the types of workloads expand, Rangarajan said.
“What agentic AI is doing is it's driving CPU demand across three networks vs. just general-purpose servers,” he said. “Agents like to do searches, in search clusters. Agents execute enterprise tools – CRM, ERP – we increasingly have agents doing EVM workloads. Your agents are accessing databases. All of this needs a broad CPU portfolio for these workloads.”
Rangarajan pointed to the pipeline of compute stages for agentic AI, noting that while the 192-core fifth generation Epyc 9965 from the Turin family outperformed Intel’s 128-core Xeon 6980P and Amazon’s 192-core Graviton5 chips, the 256-core Epyc 9996 Venice processor out performs them by greater margins.
He zeroed in on the performance with ephemeral tools, which are short-lived, isolated runtimes created during an agent session or task and are destroyed automatically.
“Ephemeral tools are the more interesting part of agentic AI because there is no well-defined benchmark there,” Rangarajan said. “It's a very, very, broad set of workloads. What we did was picked a few workloads – group graph, document parsing, coding agent, compilation – so we picked a big step we thought was representative of some of the agentic workloads we are seeing run inside of AMD. ... Overall, if you just break it down stage by stage in the pipeline, you're doing pretty well in each of the stages and the mix of these workflows could be different or different agentic workflows. But overall, we think we're doing well with Venice for agentic workloads.”
We have an analysis of the Venice SKU stack coming right up and will do an architectural deep dive on the Venice CPUs as well as what to expect from forthcoming "Verano" and "Venice-X" CPUs that are due next year. Stay tuned.
來源:The Next Platform · nextplatform.com