Two GPU stories from the last two weeks are worth separating from the usual pricing noise. One shows how a driver regression actually gets found and fixed, and where the driver's responsibility ends. The other is the first real technical spec for a memory type that could eventually loosen the capacity pressure behind today's GDDR and HBM costs.
A driver regression closes, one crash stays open
NVIDIA's GeForce 610.88 WHQL release fixes a bug the driver itself introduced: Halo: Campaign Evolved was crashing on RTX 50-series cards under the R610 branch (610.62 and 610.74), and 610.88 removes it. The same build resolves a Path of Exile 2 issue where long DirectX 12 sessions developed periodic pauses, and quietly patches shutdown crashes in Strange Brigade and Zombie Army 4: Dead War tied to the NVIDIA App overlay. On the feature side, it adds DLSS 4 with Multi Frame Generation for Halo: Campaign Evolved and DLSS Super Resolution plus Reflex for the Gears of War: E-Day multiplayer beta, ahead of the full game's October 6 launch.
What 610.88 does not touch is the Battlefield 6 Season 4 crash that had been hitting RTX 50 owners on water-heavy maps and after swapping weapon attachments, first appearing after installing 610.74. EA and DICE told affected players to roll back to driver 596.49 while they investigated, and TweakTown reported that the eventual fix shipped from the game side rather than NVIDIA's: Battlefield update 1.4.1.5 addressed the interactions causing the crash. It's a useful case for telling a driver bug from a game bug before reinstalling anything: check which side of that boundary a crash report actually sits on, since the fix (and the update to install) depends on it.
A NAND-based alternative to HBM gets its first real spec
Sandisk and SK hynix published the first formal High Bandwidth Flash specification through the Open Compute Project, with Google DeepMind and Tenstorrent named as ecosystem partners. The spec defines three bandwidth tiers running from roughly 400 GB/s up to 3 TB/s, with individual 8-high or 16-high 3D NAND stacks reaching up to 512GB, against the roughly 64GB ceiling on a single HBM4 stack. That capacity gap is the entire pitch: HBF targets AI inference workloads, particularly mixture-of-experts models, that need a large pool of memory close to the compute die more than they need HBM-class latency on every access.
None of this reaches a graphics card soon, or possibly ever in a consumer form. The spec targets data-center packaging, and Sandisk and SK hynix frame it as enterprise-only for now, with the UCIe chiplet integration and heterogeneous-packaging costs still without a consumer path. It's worth tracking mainly as a signal of where relief for the memory-capacity pressure behind current GDDR and HBM pricing might eventually come from, even while that relief stays architecturally distant.
Sources: