HPC News Bytes – 20260928

SPEAKER_1 (0:05)

Welcome to HPC News Bytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies.

Doug Black (0:14)

Hi, everyone. Welcome to HPC News Bytes. I'm Doug Black, and with me is Shahin Khan. Alibaba used its AppSara conference last week in Hangzhou, China, to unveil the Zhenwu V900 accelerator. The company said its next flagship model will have 5 to 10 trillion parameters and set a target of more than 20 gigawatts of cloud capacity by 2032

All that came during President Xi Jinping's September 23 to 25th state visit to the US last week held during the UN General Assembly week. AI was part of the U.S.-China discussions, and the two sides subsequently agreed to establish an AI dialogue. The UN Security Council separately discussed AI and international security with Chinese and US. AI companies represented. Alibaba is one part of a much broader Chinese advanced computing effort. Huawei, Baidu, and its subsidiary, Hunlun Jin, Chinese Academy of Sciences, referred to as CAS, and the Sougan-Higan ecosystem. And of course, Lineshine, now number one on the top 500 list of the world's most powerful supercomputers.

At the chip manufacturing level, there is SMIC and increasingly Shanghai Wall-E, China's answers to CSMC, alongside efforts in domestic lithography, ranging from immersion-DUV, meaning deep ultraviolet technology, to a reported extreme ultraviolet, or EUV prototype, which to date only Olin-based ASML has and is under an export ban to China.

Shahin Khan (2:04)

Yeah, getting access to the leading edge chip fabrication remains the big structural weakness in technology for China. So its strategy naturally seems to be to focus on areas where it does have access to similar capabilities or even parity, and then steadily build domestic know-how. If your processors are behind, throw more processors at the workload. Improve the interconnect and packaging. Squeeze more utilization out of them through software. Use CPUs wherever you can so you don't need GPUs. And co-design the model and machine together to get more efficiency. We covered Lineshine, the CPU-only ARM-based number one system on the top 500

This happened a few months ago, which demonstrated that China can scale a domestic architecture to the very top, although potentially at a considerable cost in power and efficiency.

In parallel, they have added more players and more redundancy across the stack. Alibaba, Huawei, Baidu and Haigon provide different compute paths. SMIC and Wall-E provide more than one advanced manufacturing path. And their efforts in advanced datography target the hardest remaining choke point.

Of course, a newcomer can leverage the accumulated industry knowledge about how ASML did it. And there was a report in Reuters last December that China appears to have built an operational EUV prototype that can generate EUV light. That's one of the many hard steps.

And they did that with help from recruited former ASML engineers.

That's impressive. And even if they managed to do that scale and with precision, it will be one step in many. The supply chain for ASML includes hundreds of specialist vendors. You might say, what if they gain access to all the technical detail available one way or another? Even that is very different from building, tuning, and operating a production worthy EUV system. It took ASML over two decades to get its first system going. Replicating that could be faster, but not by a lot.

Doug Black (4:12)

Hi, Pearl. The European Union chip development company has delivered the first RAYA-1 CPU to Bull for integration into the Jupiter supercomputer's CPU partition. RAYA-1 is built on TSMC 6 nanometer with 80 ARM Neoverse V1 cores, HBM2e and DDR5 memory, and PCIe Gen5 with C6, that is CCIX connectivity. The ecosystem is also beginning to broaden. Cyperl CPUs are planned for the Italia inference system, and the European domestic cloud project OpenCube, working with HPE's European lab, has delivered a RAYA server based on Cyperl's reference design called the CEN server, as in the River CEN.

The next-generation RAYA 2 is the more important next step. It's described as a dual-chiplet processor with more than 96 Neoverse V3 cores connected to UCIE with tape-out expected in 2027 and deployment beginning around 2028

It is already planned for the CPU partition of Europe's second exascale system after Jupiter called Alice Rococke. Europe also has Intel's advanced manufacturing operation in Ireland, all of which means that the EU is moving ahead even if at a relatively slow pace with its exascale and technology sovereignty strategies.

Shahin Khan (5:47)

RAYA 1 is Europe's first modern foray into the high-end CPU space. So Cyperl has had to do a lot of foundational work before it could build a processor. It had to build the company, engineers, tools, and ecosystems at the same time. Those are already hard areas that you have to push, so it makes sense not to push too hard on technology. C6 is one example. It was baked into RAYA 1 before several industry projects, C6, Gen Z, OpenCAPI, CXL itself, effectively converged around CXL with the idea that future CXL specification would blend them all into a single proposed standard.