珠光宝气

凤凰网科技讯 8月24日，摩尔线程正式发布《MTT S5000 Prefill-as-a-Service技术白皮书》，基于旗舰级AI训推一体智算卡MTT S5000，提出Prefill-Decode异构解耦方案，面向AI Agent、代码生成、超长文档分析等长上下文推理场景，旨在破解推理成本瓶颈。白皮书指出，大模型输入上下文规模进入百K至1M级别后，Prefill（预填充，计算密集型）与Decode（解码，访存密集型）在同一资源池混跑造成结构性算力错配。    Commerce Secretary Gina Raimondo said this weekend that U.S. business leaders are "very worried" about a partial federal government shutdown if Congress cannot pass a short-term funding measure ahead of an Oct. 1 deadline."They are, I think, in some cases, frustrated that this is how government operates," Raimondo said in an interview that aired Sunday on "Face the Nation."  The White House has asked Congress to pass a short-term spending bill to avoid a partial shutdown on Oct. 1. Raimondo said she hopes there will be a "speedy resolution" to the situation.       "It's a challenge for the economy and (has) the potential to set us back," she continued. 

• Transcript: Commerce Secretary Gina Raimondo on "Face the Nation," September 3, 2023

Earlier this summer, Republican Rep. Garret Graves of Louisiana told "Face the Nation" that he couldn't rule out a potential government shutdown after struggles to make a deal to raise the nation's debt ceiling stretched on, bringing the country dangerously close to a default.                 Fitch Ratings cited the standoff as one of the reasons why the ratings agency recently downgraded U.S. credit from its highest rating. The agency said that the government's "repeated debt-limit political standoffs and last-minute resolutions have eroded confidence in fiscal management."The August jobs report, released Friday morning, showed 187,000 jobs added last month to the workforce. However, it also showed that underemployment – the number of people working more than one job, or jobs they are overqualified for – has gone up."The reality is that inflation still exists," Raimondo said. "I don't want to minimize what Americans are feeling, and that's why we get up and go to work every day, but this economy by any measure is doing incredibly well, and much better than anyone could have predicted, I think, three years ago when we started."

On China

Raimondo also spoke on her recent visit to China, saying that "a great deal of progress" was made. Raimondo is the first U.S. commerce secretary in five years to be on the ground in China.                       "In terms of concrete deliverables, I was able to open three specific lines of communication with the Chinese," Raimondo said. These included sharing information about export controls, opening a dialogue on commercial issues, and discussing how trade secrets are handled.American business leaders have expressed concern about the handling of trade secrets, Raimondo said. Another is the lack of transparency in Chinese business regulations. American companies in China have had their offices raided and employees detained, without Chinese authorities presenting hard evidence of wrongdoing. "When I expressed these concerns, I was heard," Raimondo said, before emphasizing that she  "wasn't given any promises" by Chinese leaders. When asked if she trusted the Chinese government after her visit, Raimondo said that "trust is probably not the word" she would use. "Actions speak louder than words," Raimondo said. "So time will tell whether they are going to really be as they say they want to be, you know, more market oriented and open and reform minded when it comes to their economy, their regulatory regime."Raimondo was the fourth high-level official from the Biden administration to visit Beijing this summer, following Secretary of State Antony Blinken, Treasury Secretary Janet Yellen, and the Presidential Envoy for Climate, John Kerry. "But one thing that I think we can take a bit of heart in is it's in China's interest," Raimondo continued. "It's in their economic interest to have foreign direct investment and to make sure, you know, they had a very good economy when they were market oriented and reform minded. And so we have to hope that self interest will affect their future decisions."                      — Richard Escobedo contributed to this report.                                  。解决方案是将二者彻底解耦，分别运行在各自最契合的硬件资源池上，实现单位Token基础设施成本大幅下降。

二 | MTT S5000硬件特性与Prefill任务高度契合：提供高稠密算力压缩首字时延，原生支持FP8全精度计算，并全面兼容CUDA及PyTorch、vLLM等主流推理框架。

三 | 实测数据显示，在64K上下文场景下，单机Prefill吞吐达95,920 tok/s，按智谱API定价测算，满负载下单机月收入可达64.6万元；400K极限长序列场景TPM达2.6MTokens，吞吐平稳可控。

四 | 摩尔线程表示，大模型产业竞争下半场本质上是推理效率与商业回报的较量，该方案旨在帮助企业在万亿参数与Agent时代降低总拥有成本。完整白皮书已在官网发布。