The answer shouldn't depend on which GPU you rented

A developer built a deterministic inference engine, frozen engine v27, that reproduced all 136 sampled answers token-for-token across nine GPU types spanning four NVIDIA architecture generations, from an H100 NVL to an RTX 3090, while standard vLLM 0.32 matched 0 of 136 on five of those cards. The work targets 8-bit (w8a8) serving of DeepSeek-R1-Distill-Qwen-7B, where the team cites outside research (Yuan et al., arXiv 2506.09501) finding up to 9% accuracy variation and 9,000-token response-length differences from GPU count, type and batch size. Seven Intel and AMD CPU models also matched a shortened 32-token signed replay bit for bit.

Run the same AI model with the same prompt and the same random seed on two different GPUs. A standard serving setup can give you two different answers. In our test, it did. That isn’t a bug in one product. It’s a property of floating-point math. Add the same numbers in a different order and the last bits can come out different, and different chips and software paths can add in different orders. In a reasoning model those tiny differences can compound over thousands of tokens until the model writes a different solution. Change the card and you can change the answer. This is documented. Researchers from Rice University, the University of Minnesota Twin Cities and Adobe found that a reasoning model like DeepSeek-R1-Distill-Qwen-7B, the model we used, “can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size” bfloat16, greedy decoding; Yuan et al., arXiv 2506.09501 . If a benchmark score moves when the GPU changes, nobody can reproduce it on different hardware. We measured it ourselves. On November 9, 2026 we ran a standard setup, vLLM 0.32 in normal mode, and compared each GPU with vLLM’s own H100 run. The H100 NVL and the H200 reproduced 135 of 136 answers. The A100, RTX A6000, A40, RTX 4090 and RTX 3090 reproduced 0 of 136 . The tested vLLM int8 configuration rejected SM120 on both Blackwell cards, so those two have no vLLM match count. Different doesn’t mean wrong. It means a rerun on a different card can’t confirm the original output. Our engine reproduced 136 of 136 on all nine GPU types we tested. What we did We froze one job on frozen engine v27: DeepSeek-R1-Distill-Qwen-7B in 8-bit w8a8 , 17 problems from AIME 2024 and MATH-500, eight sampled answers per problem at temperature 0.6 and top-p 0.95, up to 16,384 new tokens per answer. We recorded the run on an H100 NVL and signed it. Then we replayed the frozen job on nine rented GPU types: H100 NVL, H200, A100 SXM 80GB, RTX PRO 6000 Blackwell Server Edition, RTX A6000, A40, RTX 5090, RTX 4090 and RTX 3090. That’s four NVIDIA architecture generations Ampere, Ada Lovelace, Hopper and Blackwell , data-center cards and gaming cards, rented on RunPod at $0.22 to $5.29 an hour November 9, 2026 rates . Every one matched the signed reference: - all 136 sampled answers , token for token - all 546,114 generated tokens per GPU - every recorded per-step logit fingerprint - a separate 92-case Qwen2-7B check, 92 of 92 on every card On all nine GPU types, there were zero differences from the reference. Seven Intel and AMD CPU models also matched: AMD EPYC 9654, 7742, 9355, 7352 and 7452, Intel Xeon Platinum 8568Y+ and Xeon Gold 6342. Each one replayed a short signed version of the job, capped at 32 tokens, and two small Qwen checks, bit for bit. To be precise: the CPUs ran the short replay, not the full workload. Why 8-bit matters Bit-exact results in full 32-bit floating point have been shown before, and the same outside study found 32-bit runs near-perfectly reproducible, 16-bit moderately variable and bfloat16 substantially unstable. Much of production AI runs at 16-bit or 8-bit to save memory and money, so that’s where portability has to work. 8-bit alone doesn’t make results portable. The same w8a8 model file in the standard vLLM setup matched 0 of 136 on five of the nine cards. This test used that publicly available 8-bit file, not a 32-bit research setup. We didn’t get there by switching off acceleration. Why it matters Portability becomes buying power. Here’s the same exact job at November 9, 2026 RunPod rental rates: $0.244 per million output tokens on an H100 NVL, $0.082 on an RTX 5090, $0.064 on an RTX 3090. If the answer doesn’t change across the GPUs you’ve tested, you can pick hardware by price and availability instead of by whichever card your validation happened to run on. Rental-only math. It excludes licensing, storage, transfers, provisioning and idle time. Anyone can check it. Another party can replay the frozen job on a tested GPU type and compare it token by token with the signed record. A signature binds a record to a key; the replay is what shows the computation was reproduced. Audits and science get a fixed point. An auditor, a regulator or a peer reviewer can inspect the public records and replay the frozen job on another tested machine, instead of taking a results table on faith. That’s a different kind of result. Debugging gets honest. On the GPUs we tested, swapping the card didn’t change a single token. So when an output changes, you can rule out the GPU swap and look at the input, the model or the code. Speed, plainly We’re not the fastest engine on a single card today. In the November 9, 2026 test, vLLM 0.32’s normal mode was faster and cheaper per output token on every GPU where its tested configuration ran. Luxi still delivered 2,946.2 output tokens per second on one H100 NVL 128 concurrent answers and 2,347.5 on an RTX 5090 rented at $0.69 an hour 64 concurrent answers , and every token matched the signed reference. The tested vLLM int8 configuration rejected SM120 on both Blackwell cards. Ours ran on both. We’re working to close the speed gap. Meanwhile, a fast answer that changes with the hardware can’t be checked on different hardware. Don’t take my word for it - The signed answer records from all nine GPUs are public , published unchanged. - The replay demo is a free download Linux x86 64, closed binary . It is a new build of the same v27 source and does not inherit the November 9 nine-GPU result. It replays a short frozen job on a supported NVIDIA GPU or a CPU and checks it against the signed November 9 record. On a supported NVIDIA GPU it also reruns the full 136-answer suite. On November 10, 2026 we tested release 1.0.1 on an RTX A6000, and release 1.0.0, which ships the same engine files, on an RTX 5090 and an A40: 136 of 136 on each. - It fails when it should. A changed token, fingerprint, prompt, seed or temperature, a dropped token, a re-signed forgery and a wrong trusted key were all rejected. A GPU request never silently falls back to the CPU. If you have published signed records like these for an 8-bit reasoning model across GPU types, post them. We’ll read them. The limits, because this is science This is an internal measurement, not independent validation. It covers one cloud RunPod , NVIDIA GPUs, one 7B reasoning model plus a separate Qwen gate, and one measured run per performance setting. We make no AMD or Apple GPU claim for this engine yet. Identical outputs show the computation was reproduced, not that an answer is correct. One engine underneath LuxiEdge is one core engine, patent allowed, grant pending, with seven commercial products built on it. On two published satellite grids October 10, 2026 , the same approach kept orbit states byte-identical across a Mac, ARM Linux, emulated x86 and an NVIDIA H100 near-earth build . On the public Oasis PiWind catastrophe model, stock oasislmf 2.5.8 gave 10 different results in 10 runs across 2 machines; Luxi gave 1 October 12, 2026 . I call it exactation: exact computation that travels with the job, at production precision, on every machine we’ve tested so far. If your answers have to survive a change of hardware, bring us your model and workload. We’ll confirm the supported scope, then run a private GPU-swap evaluation: freeze the job, change the GPU, compare. Evidence, download and evaluation: luxiedge.com/inference https://luxiedge.com/inference replay-demo Signed records: LuxiDemo evidence on GitHub https://github.com/RegularJoe-CEO/LuxiDemo/tree/ebcc3f7777549a10ef69a9319733362a36880641/evidence/r1-distill-7b-gpu-swap-2026-11-09 Private GPU-swap evaluation: luxiedge.com/contact https://luxiedge.com/contact gpu-swap-evaluation