Two days ago we told you something very interesting: the Chinese state medium Securities Times had revealed that Huawei was about to present a technological advance that pursued Reduce China dependence of HBM memory chips (High Bandwidth Memory) from abroad. According to this source Huawei was going to officially announce his technological milestone a few hours later, during the celebration in Shanghai (China) of the Applications Forum and Development of Financial Reasoning 2025.
Huawei has fulfilled what was promised, although not as we had planned. In any case, before getting into flour it is important that we remember that Chinese memory chips manufacturers They are not producing solutions capable of competing with the most advanced memories manufactured by South Korean companies Samsung and SK Hynix, or the American Micron Technology. GPUs for Ia work side by side with HBM memory chips. In fact, its performance is largely conditioned by these memories.
As the editors of SEMIANALYSISthe total bandwidth of the HBM3 memory chips that live with some of THE GPU FOR THE MOST ADVANCED Nvidia or AMD exceeds 819 GB/s, while DDR5 and GDDR6X memories reach much more modest 70.4 GB/Sy 96 GB/s. HBM3E memories and future HBM4 are even better. Chinese manufacturers of this type of chips do not yet produce this kind of memories, but it seems that Huawei will deeply alter this scenario.
An algorithm expressly designed to accelerate inference in AI
The filtration that occurred scarcely 48 hours suggested that probably what Huawei was going to present was a avant -garde packaging technology that, perhaps, would rival those used by SK Hynix, Samsung or Micron to produce their HBM3 and 3E memories. And it is that manufacturing these integrated circuits is complex because they require stacking several DRAM chips and implementing an interface between the XPU (Extended Processing Unit) or extended processing unit and extraordinarily dense HBM chips. As a button shows: in a HBM3E stack the XPU and the HBM memory are linked through more than 1,000 drivers.
According to Huawei, the UCM algorithm is capable of drastically accelerating inference in the great AI models
However, finally Huawei has presented a different technology: an advanced algorithm called UCM (Unified Cache Manager) that, according to this company, it is capable of drastically accelerate inference In the great models of artificial intelligence (AI). A relevant note: inference is broadly the computational process carried out by language models with the purpose of generating the responses that correspond to the requests they receive.
To achieve its purpose, the UCM algorithm displays a very ingenious strategy: decide in what type of memory it is necessary to store each data taking as a fundamental indicator the latency requirements. In practice, this algorithm behaves as a gigantic cache that guarantees that each data will go to the right memory, including HBM3, with the purpose of minimizing latency during inference. If it is a very often used data, it will be stored in a very fast memory, such as HBM3. According to Huaweithis technology is able to reduce the latency of inference by 90%. Interestingly, this company plans to do the UCM Source Open Algorithm in September.
More information | SCMP
In Xataka | Nvidia has to deal with the absolute distrust of several US legislators. His plan in China is in danger
In Xataka | The US wants to end the chips for the Chinese that are sold abroad. And China knows how to defend oneself
GIPHY App Key not set. Please check settings