Nvidia shows off its next-generation Kyber rack-scale solution to be powered by Rubin Ultra GPUs with four compute chiplets and 1 TB of HBM4E memory per package.
At GTC 2026, Nvidia revealed the Groq 3 accelerator and Groq LPX rack as part of the Vera Rubin platform. These SRAM-packed, inference-focused chips deliver large amounts of memory bandwidth to help Rubin deliver low-latency interactions with AI models spanning trillions of parameters and million-token contexts.
Nvidia announced more details about its new 88-core Vera data center CPUs, claiming impressive 50% performance gains over standard CPUs, fueled by a 1.5X increase in IPC from its Olympus cores. The firm also unveiled its new Vera CPU Rack architecture, which brings 256 liquid-cooled CPUs into one rack for CPU-centric workloads.
A single-fan RTX 4070 Ti Super had been in the works at Zephyr, a Chinese vendor, for a while, and it was close to completion, with even thermal testing data publicly released. Unfortunately, the memory crisis has gotten to Zephyr as well, and it has cancelled the project, choosing to instead develop an RTX 4070 Super instead.
An RTX 5070 Ti with user-applied liquid metal died because the TIM leaked out everywhere and shorted multiple components, eventually killing the core as well. Despite being part of a "repair" video, there's nothing really here to fix, as most of the important ICs would need to be replaced or at least reballed.
A new LLVM patch has added V_FMA_F32, a 3-operand fused multiply-add (FMA instruction and introduced the VOPD3 instruction format for RDNA 5. Both of these changes should make it easier for compilers to use dual issue execution, working around the strict pairing rules that would otherwise limit max FP32 throughput in certain workloads.
At GDC 2026, Nvidia held a presentation somehow aimed at gamers and not data center clients. Still, it was sprinkled with AI pat-on-the-backs, with the company touting that its future gaming GPUs will offer 1,000,000 better path tracing performance. And the current-gen Blackwell family is apparently already 100,000x better due to dedicated Tensor and RT cores.
Nvidia says it is permissible for ByteDance to use AI clusters outside of China to develop its AI prowess as long as these clusters are built in compliance with the U.S. export controls.
The G100 lineup from Lisuan Tech now has an official release date, with the LX 7G106 being the gaming SKU with 12 GB of VRAM and a purported RTX 4060-like performance. The "LX" series of professional GPUs, likely using the 7G105 silicon, were also revealed with three SKUs offering up to 24 GB of ECC VRAM.
XeSS 3.0 has been available inside Intel's graphics driver suite for some time now, but the chipmaker has finally released the accompanying SDK for game developers, albeit only in closed-source form.
Asus’ special one-off ROG Astral RTX 5090 Real Gold Edition has been a great investment, and is currently worth an estimated $830,000 based on its scrap gold value alone.
Nvidia is reportedly working on a new RTX 5050 with 9GB of GDDR7 VRAM, up from 8GB GDDR6 on the existing model. Moving from 20 Gbps chips to 28 Gbps chips, the memory interface is also said to be reduced to 96-bit, from 128-bit on the original. Moreover, an RTX 5060 with a cut-down GB205 die is also said to be in the works, with otherwise same specs.
Users report that Nvidia's latest 595.71 driver is creating artificial voltage limits on many RTX 40- and 50-series GPUs, causing some products to lose as much as 200MHz in overclocking headroom.
登录
