[REVISION HISTORY]

NVIDIA AI hardware efficiency push

Updated 4 times since CLSTR started tracking revisions of this situation.

What changed

2026-08-09 03:06 UTC → 2026-08-11 02:51 UTC · added ~~removed~~

In mid‑July 2026 NVIDIA unveiled the Blackwell NVL72 platform, claiming up to a 25‑fold performance‑per‑watt gain for leading open‑source models and a 5‑fold monthly improvement on DeepSeek V4. A week later the company launched the Vera CPU, an Arm‑based server processor built around 88 custom Olympus cores, 176 threads, up to 1.5 TB of LPDDR5X memory and 1.2 TB/s NVLink‑C2C bandwidth to GPUs. NVIDIA said Vera runs agentic‑AI tasks 1.8× faster than top x86 CPUs while cutting latency by roughly 40 % and operating at 45 °C with liquid cooling. The Ruby AI server, using the same cooling loop, was presented as a rack‑space‑saving, energy‑efficient solution. On July 26 NVIDIA confirmed Vera Rubin chips were in production and announced a reduction in rack‑scale memory module size to curb HBM4 costs, cutting total CPU memory from about 55 TB to 28 TB. On August 3 NVIDIA announced that Vera is integrated into the BlueField‑4 STX AI‑native storage platform, delivering up to 3.3× faster compression, 1.43× faster AES‑128 encryption, and more than three‑fold acceleration for data‑recovery operations such as Reed‑Solomon reconstruction (3.26×) and CRC32C checks (3.67×). These gains target the CPU bottleneck in high‑concurrency AI storage, promising lower power use while handling massive data streams securely. Two days later the company addressed online speculation that a new SKU represented a performance downgrade. NVIDIA clarified that the 96 GB memory modules constitute a different configuration, not a reduction in capability, and that the Vera CPU continues to support the full 1.5 TB memory option. ~~The clarification reinforces NVIDIA’s focus on efficiency and spec consistency across its AI‑hardware portfolio.~~ On August 8, NVIDIA addressed market rumors regarding potential memory downgrades for its upcoming Vera and Rubin Ultra AI computing platforms. On August 11, NVIDIA faced rising component costs for DRAM and NAND flash memory, driving price increases for Blackwell mid-range graphics cards.

Versions

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.