Rohan Paul@rohanpaul_ai
38AI 编辑部评分,满分 100
2026-08-05 19:34· 4分钟前
AI 导读

NVIDIA 宣称其自研 Arm CPU Vera 在存储数据路径上比传统 x86 芯片快 3.67 倍,压缩吞吐量高 3.29 倍。该芯片内置于 BlueField-4 STX 存储处理器,在 GPU 接触数据前完成加密、校验、压缩等操作。NVIDIA 称此举意在挑战 Intel 和 AMD 长期主导的存储处理器市场,因 Agentic AI 使存储操作量激增。

NVIDIA is going after a job Intel and AMD have owned for decades, and claims a 3.67x speed lead.

NVIDIA is claiming its own Arm CPU does the routine work storage systems perform, encrypting, checksumming, compressing, rebuilding, up to 3.67x faster than the x86 chip that has always done that job.

Data has to be locked, squeezed, and checked before it lands on a drive, and NVIDIA now does that 3.67x faster.

Vera now runs the storage data path, hitting 3.29x higher compression throughput than x86 CPUs.

The chip sits inside the BlueField-4 STX Storage Processor, which handles everything done to data before a GPU ever touches it.

That work used to be quiet, because a controller serving ordinary applications could compress and encrypt on spare cores without anyone noticing.

Agentic AI ended the quiet part, since a single agent step can fire several storage operations and thousands of agents run at once.

来源:Rohan Paul · x.com

Rohan Paul · @rohanpaul_ai · X·2026-08-05 19:34·4分钟前
AI 导读

NVIDIA 宣称其自研 Arm CPU Vera 在存储数据路径上比传统 x86 芯片快 3.67 倍,压缩吞吐量高 3.29 倍。该芯片内置于 BlueField-4 STX 存储处理器,在 GPU 接触数据前完成加密、校验、压缩等操作。NVIDIA 称此举意在挑战 Intel 和 AMD 长期主导的存储处理器市场,因 Agentic AI 使存储操作量激增。

NVIDIA is going after a job Intel and AMD have owned for decades, and claims a 3.67x speed lead.

NVIDIA is claiming its own Arm CPU does the routine work storage systems perform, encrypting, checksumming, compressing, rebuilding, up to 3.67x faster than the x86 chip that has always done that job.

Data has to be locked, squeezed, and checked before it lands on a drive, and NVIDIA now does that 3.67x faster.

Vera now runs the storage data path, hitting 3.29x higher compression throughput than x86 CPUs.

The chip sits inside the BlueField-4 STX Storage Processor, which handles everything done to data before a GPU ever touches it.

That work used to be quiet, because a controller serving ordinary applications could compress and encrypt on spare cores without anyone noticing.

Agentic AI ended the quiet part, since a single agent step can fire several storage operations and thousands of agents run at once.

来源:Rohan Paul· x.com