Liveright Review by Walter Cummins In her latest book, Jill Lepore presents her analysis of how we have reached the creation of the Artificial State where machines set the guidelines and prioritie…

For now, from what I could find so far about running two independent CUDA processes on Windows: * * * ## Short answer to the two questions **1. Is there a better way to share a native-Windows GPU between two independent CUDA processes?** I could not find an off-the-shelf Windows mechanism that gives one process a reliable TTS playback deadline while another continuously submits LLM decoding work. NVIDIA’s MPS documentation lists Linux and QNX support, not native Windows. CUDA stream priorities are scheduling hints, not real-time deadlines or guaranteed immediate preemption. NVIDIA’s Best Practices Guide describes separate CUDA contexts as time-sliced. That makes scheduling interference plausible, but does not by itself identify which device or host-side delay dominates in your test. For the configuration you measured, **keeping both model services loaded while explicitly taking turns on GPU-intensive generation looks like a sound practical baseline**. This does not stop _audio playback_ while the LLM runs: the device can keep playing previously generated PCM. I would not replace this working design with MPS, a single-process integration, or GPU partitioning without a measurable benefit that justifies the additional work. **2. Would turn-taking still win on a different GPU?** Not necessarily. The useful test is whether TTS, _while the LLM is generating_ , can deliver the needed PCM **before the playback buffer runs out** , including the slowest relevant chunks. If it does so reliably, overlap may let the LLM finish the answer sooner. If it misses those deadlines, turn-taking can give a better listening experience even if complete text generation takes longer. GPU model, VRAM size, and _average_ TTS real-time factor alone do not determine the result. The first-to-second PCM interval, cold starts, long answers, and dedicated-memory residency all matter. The paired benchmark already provides a useful reference: the matched `plain` and `turns` runs used identical generated text for each question/seed pair. In the long-sentence condition, predicted playback gaps appeared in **6/6** runs without turn-taking and **0/6** with it. Those are **predicted underruns with the project’s simulated 0.5-second playback buffer** , not a direct microphone or speaker measurement, and not a result on another GPU. The broader improvement from roughly **4.59 seconds to 0.09 seconds** of estimated gaps also involved changes in the TTS engine and LLM placement; it should not all be attributed to scheduling alone. The results and measurement notes keep the two comparisons separate. ## Practical default I would retain the current sentence-level cooperative handoff as the reference behavior for this Windows/8 GB setup. Both services stay loaded; the application yields GPU-heavy generation to TTS when a speech segment needs synthesizing. That directly addresses the observed stuttering without making an OS-level GPU priority guarantee a dependency. The unsuccessful LLM token-rate limits and WDDM priority experiment are already documented, so simply repeating those would add little. If further tuning would be useful, the least expensive informative step is **one short timestamped PCM trace** , not a CUDA rewrite. Record when the first few PCM reads arrive and how much audio each contains, preferably alongside TTS begin/end and LLM pause/resume events. If the trace covers the first predicted gap, it can help distinguish startup buffering, synthesis pacing, and handoff timing. If the result is already smooth enough, this is optional rather than homework. ## A compact decision guide for another GPU Can the required GPU allocations remain resident in dedicated memory? No / uncertain / shared-memory spill observed: Stabilize memory residency first; the scheduling comparison is confounded. Yes: Compare concurrent LLM+TTS against cooperative turn-taking, using the same text, TTS setup, and playback-buffer rule. Does concurrent TTS deliver PCM before the playback buffer empties? Yes, with margin across cold starts and long answers: Concurrent operation is worth considering. Compare text finish time, audible startup, longest gap, and interruption behavior. No, or only particular chunks miss their deadline: Retain turn-taking as the reliability baseline. Find the first missed deadline and distinguish startup, handoff, synthesis, delivery, and memory causes before changing the scheduler. There is a small separate question about the roughly **0.13–0.14-second early estimated gaps** in two final-stage standalone runs. If these are perceptible on the actual device, a first-PCM timeline or a small buffer-sensitivity check could be informative. It is not necessary to establish the value of the existing turn-taking result. The following details are optional background and reproducibility notes for someone testing another GPU or attempting to locate the remaining latency. 1. What Windows/CUDA can and cannot control (click for more details) 2. What the published experiments establish (click for more details) 3. Why the audio buffer, not average GPU utilization, decides continuity (click for more details) 4. A practical comparison on another GPU (click for more details) 5. Small timing traces and measurement boundaries (click for more details) 6. If a CUDA-level cause becomes the question (click for more details) 7. Related reports worth separating by mechanism (click for more details) 8. Conditional edge cases and a reusable takeaway (click for more details)

The board is meeting for the first time this school year, with four student unions across two campuses presenting their yearly priorities to the body. Also being presented is the president's prioritie...

The board is meeting for the first time this school year, with four student unions across two campuses presenting their yearly priorities to the body. Also being presented is the president's prioritie...

Liveright Review by Walter Cummins In her latest book, Jill Lepore presents her analysis of how we have reached the creation of the Artificial State where machines set the guidelines and prioritie…

I used to think changing my mind meant I hadn’t been certain enough the first time. Life, motherhood, and music have taught me otherwise. Growth gives us new perspectives, different prioritie…

SINGAPORE: Singapore will step up efforts to translate research into real-world outcomes under its RIE2030 plan, with a new grant to help develop early-stage technologies from public research so the private sector can invest in and further develop them. The National Research Foundation (NRF) released the full five-year plan on Monday (Aug 31), which will channel a record S$37 billion (US$29 billion) in research, innovation and enterprise spending towards national economic and strategic priorities through 2030. Part of the spending will go to a new Research Translation Grant, which was created to provide a more direct way of working with universities and the Agency for Science, Technology and Research to grow the pipeline of "commercialisable assets", said NRF Permanent Secretary Tan Chorh Chuan. 

NOV8TIV Advisory has launched N8IQ, a platform designed to assess corporate technology and AI risks. The tool offers rapid evaluations of technology maturity, operational risk, and strategic prioritie… · Photo : GovernmentZA · CC BY-ND · Flickr

The Maryland Board of Public works approved an agreement to utilize solar energy in a Western Maryland former coal district, representing a fundamental difference in state and federal energy prioritie...

With just hours to spare before the midnight deadline, Gov. Glenn Youngkin on Monday unveiled his proposed amendments to the revised state budget, signaling his acceptance of many Democratic prioritie...