News & Commentary

The GeForce Recap: No New Driver, and GeForce NOW’s Back-to-School Pitch

Week ending August 3: the GeForce channel stayed quiet after 610.88, so NVIDIA's news came from the corporate blog instead, from cloud streaming to the CPU it uses to design its own chips.

Saya5 minute readNews & Commentary

The GeForce News feed published nothing between July 27 and August 3, so its newest items are still the 610.88 driver and the Halo: Campaign Evolved launch from July 28, both covered in the previous edition. What follows came off the corporate blog instead: cloud streaming, an edge-AI overview, a chip-design note, and a security alliance. All of it is NVIDIA writing about NVIDIA, so the figures are company claims until someone independent measures them.

No new driver, and nothing to do about it

No Game Ready release and no hotfix landed this week. 610.88 from July 28 is still the current R610 branch driver, and its release notes fix crashes in Halo: Campaign Evolved on RTX 50 Series cards running R610, the intermittent long pauses in Path of Exile 2 after extended DirectX 12 sessions, a crash on exit in Strange Brigade and Zombie Army 4 with the NVIDIA App overlay enabled, and frame pacing in certain DirectX 11 games when Smooth Motion runs together with G-SYNC.

That last one is worth noticing. Smooth Motion plus G-SYNC is a common configuration, and pacing problems there surface as stutter that never appears in an average frame rate. Beyond that there is nothing to act on: Game Ready releases cluster around game launches, so a quiet week usually means the next big title is still out.

GeForce NOW rents a GPU to students, and skips the network question

GeForce NOW's back-to-school pitch aims the Ultimate tier at anyone carrying a Chromebook, Mac or Surface, streaming what NVIDIA calls 5080-class performance with DLSS, ray tracing and Reflex. Eight games joined the library that week, headlined by Halo: Campaign Evolved, with Mistfall Hunter on July 29 and Corsair Cove on Game Pass July 31. The day pass is the low-commitment entry, its cost applying toward a first membership.

The decision underneath is real: a thin laptop either gets supplemented with a second machine or rented capacity, and streaming genuinely removes install size, patch time, VRAM ceiling, and sustained thermals under load. What it does not remove is the network. Latency and jitter decide whether an Ultimate session feels local, and average bandwidth is the least useful number in that equation. Reflex trims latency inside the render pipeline and does nothing about the round trip to the data center. The catalog is also a subset rather than all of Steam, since the model streams games already owned on supported stores.

One oddity in NVIDIA's own list: it records Call of Duty: Black Ops 6 as a new release on Ubisoft Connect, a strange storefront attribution for that title. Confirm the entitlement before counting on it.

Vera is designing the chips that come after it

NVIDIA has deployed its Vera CPU across the electronic design automation workflows it uses to build its next generation of CPUs and GPUs. Vera is 88 custom Olympus cores with an LPDDR5X memory subsystem and a second-generation Scalable Coherent Fabric. Working with Cadence and Synopsys, NVIDIA reports up to 1.5x higher performance on selected workloads for Cadence Jasper, a formal verification platform, and Synopsys VCS, a functional verification and simulation tool it says ran on the same number of cores.

The qualifiers carry the weight there: no baseline CPU named, no absolute runtimes, core parity stated for VCS rather than both tools, and "up to" on "selected workloads" is as narrow as a performance claim gets. Directional evidence, not a benchmark. It still matters for GeForce hardware, because verification and logic simulation are among the last heavily CPU-bound stages of chip design and they gate how many design iterations fit into a schedule. Faster regression turnaround is one of the few levers that actually shortens the path from RTL to silicon. The named successor is Rosa on the Rigel core, no date attached.

The other two, briefly

The Jetson post announces nothing: no new hardware, no pricing, no dates, just a promotional overview of modules already shipping. Its figures are 67 TOPS for Orin Nano Super, 275 TOPS for AGX Orin, and up to 2070 FP4 teraflops with 128GB of memory for AGX Thor, which should not be read as a progression. The Thor number is FP4 while the Orin numbers are bare TOPS, so they are different measurements. Choosing an edge accelerator on one throughput headline, without the precision, sparsity assumption, and memory bandwidth behind it, is how a project lands a module that benchmarks well and then stalls on the actual model. The useful part is the tooling it names, Jetson Device Skills and Jetson BSP Skills, aimed at the setup work that eats the first week of an edge project.

The Open Secure AI Alliance launched July 27 with more than 70 partners, among them Cisco, CrowdStrike, Hugging Face, IBM, Microsoft, Mozilla, Red Hat, and the Linux Foundation, building on the Akrites initiative and OpenSSF work. Deliverables exist rather than only intent: NVIDIA's Object-Oriented Agent framework is on GitHub, Hugging Face offered the Safetensors format to the PyTorch Foundation, HPE is contributing to SPIFFE/SPIRE for agent identity, IBM and Red Hat brought signed-patch supply-chain work, Microsoft a multi-model scanning harness. NVIDIA's case for open models leans on the recent Hugging Face incident, where by its telling an open-weight model run on Hugging Face's own infrastructure let the team work through more than 17,000 actions and contain the intrusion. None of it changes anything on a gaming PC, but the direction is worth noting: the company pushing local inference on RTX cards is also arguing publicly for open weights and inspectable agent harnesses.

Nothing announced this week changes a hardware decision. The item to watch carries over: NVIDIA put the improved DLSS 4.5 Ray Reconstruction model in an August window, and August has started. For what the hardware costs meanwhile, Market Watch tracks it week to week.

Sources: