Alibaba delivers RISC-V server chip optimized to run China’s top AI models

Alibaba has revealed a new server chip that it says is the most powerful processor ever to use the RISC-V instruction set. According to a social media post by Alibaba’s DAMO Academy, which develops some of its chips, the new XuanTie C950 is ready to power cloudy servers, generative AI workloads, high-end robotics, and edge …

[Groo The Wanderer - A Canuck]

But do you really need to match "modern" CPU performance to be useful in serving workloads? From what I understand, most web environments run dozens or hundreds of CPU threads at relatively low performance, but with massive parallelism. So you slap enough of these new RISC-V processors in a cabinet, and you're going to serve some serious web capacity, even if it isn't as "efficient" as the latest and greatest.

How many enterprises have you worked at that run the latest and greatest? How many of you can afford the virtual systems on offer that promise deployment on the latest and greatest hardware instead of a 4-6 year old blade?

Parallelism definitely changes the metrics of "what can I do with it?"

Considering where RISC-V performance was even 2-3 years ago, this is a landmark processor.

[Anonymous Coward]

Yes, but can it run GhostWrite (eg. at 1/9ᵗʰ the speed of a 2019 CPU, or 1/3ʳᵈ for the C950)?!

[Groo The Wanderer - A Canuck]

If you care to actually read the article:

"Per analysis by Google researcher Laurie Kirk puts it nearly on par with Apple’s M1 chip – which the iGiant launched in the year 2020."

That's performance only 6 years behind what was considered one of the most efficient and high performance processors of the year.

[Anonymous Coward]

Yes, but Laurie's looking only at the SPECInt score ... the 'speed of a 2019 CPU' link above (and also this one) do more thorough benchmarking of the XuanTie C920 (of the 64-core Sophon SG2042), and T-Head stated this new C950 is 3x faster than that 'old' C920, hence the values I posted above.

There's also a need for RISC-V to break out of SPECInt-only benchmarking, and show perf on SPECjbb and other more realistic workloads imho.

Plus, GhostWrite was a big issue on C920, C910, ... and it would be important to know if T-Head removed the offending MMU-bypassing memory-access instructions, or provided some way to disable them.

[CheesyTheClown]

Very backwards analysis

1) alibaba has agreements with whoever is fabbing the chip. They didn’t design a 5nm chip unless their supplier could commit to delivering it at a specific price and timeframe. Why in the world would someone doubt that? Are they under the impression that Alibaba doesn’t know how to run a business?

2) the chip as a general purpose processor is comparable to an M1. Does the writer have any idea how impressive this is? Apple is a pretty impressive company with insane amounts of resources and engineering capabilities. It was genius that they could make the M1. Alibaba has made massive in-roads catching up.

3) the chip is an AI chip which means that while Alibaba could have used a lot more real estate to make improve general compute, they were focused instead on an engine to feed NPUs.

4) Qwen 3.5 might be the most impressive open model n the market right now. I’ve been using it for most of my work for weeks and it’s killing Kimi K2.5 and ChatGPT 5.4 has been just a disaster. Alibaba will optimizing their processors and models to perform well with each other. They have already proved they can.

I think this article should be written much more in favor of the massive achievement of Alibaba. America’s top AI companies are so fragmented. Microsoft should be 100% in Mai tech not OpenAI. Google is on the right track kinda but they’re ignoring local AI. OpenAI and Anthropic aren’t even players anymore and they haven’t even noticed. Al is not a product, it’s a feature or a tool. OpenAI and Anthropic lack products because everything they make can be easily replaced free tech.

Look to Deepseek. They figured it out already.