An RTX 3090 Ti test put BF16 matrix math at 157 teraflops, up from 87 and 1/2. On two RTX 3060s running quantized chat models, token generation did not change.
Both results came from the same GPU software change. Faster math helped in some reported workloads, but the quantized chat test shows why a large gain in matrix math does not automatically mean faster replies.
Faster GPU Math, Different Workloads
The project maintainer says the change lets consumer RTX 30 cards start FP16 and BF16 instructions at the rate of their professional equivalents. Nvidia’s published figures show a performance gap between cards using the same GA102 chip. Those figures do not establish why the gap exists; the maintainer attributes it to product segmentation.
Other users reported gains. On an RTX 3070 Ti laptop chip, language-model prefill rose 44%, while ResNet 50 training rose 14%. An RTX 3090 Ti user reported roughly 56 to 59% faster BF16 image generation. Eight-bit versions of the same models barely changed.
The quantized chat test looked different. On two RTX 3060s, prefill rose between 4/10 of a percent and 3 and 1/2. Token generation stayed where it was. One possible explanation is that generating tokens can spend much of its time waiting for data to move from GPU memory. Faster tensor math has little to do during that wait.
That makes the distinction between prefill and token generation important here. Even within the same chat test, one moved slightly and the other did not. These speed reports were not independent lab tests, so they establish reported results for these workloads, not a general speed gain. One user also reported a driver failure after loading the module in a virtual machine; the card recovered after a reboot, and a proposed fix worked for that user.
Two Model Ideas, Two Missing Measurements
Levenshtein sends a question through the same Oro model layers up to four times, reusing the same weights. Its score rose from 63% on the first pass to 76 on the fourth. The extra passes helped the reported score, but more passes can also break an answer, and the score barely changed by the last pass.
A proposed early stopper would quit when another pass is unlikely to help. The paper reports 15 to 24% less compute with little effect on accuracy. That saving was calculated afterward from saved outputs, though. Live early stopping has not been built, so the reported saving is not yet a measurement of a running stopper.
K3 raises a different measurement question. It has about one and a half terabytes of weights to place across Cerebras wafers, and one layout estimate rose from 34 wafers to 59. Its 24 full attention layers each get a wafer for KV cache, which holds earlier conversation information. Its 69 linear attention layers keep no KV cache.
There is a measured speed result nearby, but it belongs to Kimi K 2.6, not K3. Artificial Analysis clocked Kimi K 2.6 at 981 tokens a second in May, with its weights spread across wafers. K3 itself has not been measured. The unanswered question for its layout is how much of the estimated 28 GB of KV cache on each full attention wafer one conversation would use.
These are different kinds of unfinished result. Levenshtein has a score from repeated passes and a compute saving calculated from saved outputs, but no live stopper. K3 has a wafer layout estimate, while its speed and per-conversation KV-cache use remain unmeasured.
What the TV Tests Could See
A 2024 study observed automatic content recognition, or ACR, on LG and Samsung TVs showing broadcast channels and HDMI inputs. The TVs made fingerprints from screen captures and sent them to servers to match against known shows.
The documented capture rates were much more frequent than the network traffic researchers saw. LG’s documentation says it captures a frame every 10 milliseconds; researchers saw network traffic about once every 15 seconds. Samsung’s documentation says one frame every 500 milliseconds; researchers saw fingerprints sent about once a minute. A capture rate and an observed sending rate are separate measurements.
The researchers saw no ACR traffic inside Netflix or YouTube. In the US, ACR remained on inside the TV makers’ own streaming app. Those observations describe what the researchers saw in these tests, including where traffic appeared and where it did not.
The opt-out result was specific and concrete. In the LG and Samsung tests, opting out stopped contact with every ACR domain the researchers identified. They saw no replacement ACR domains.
The open measurements are just as specific: independent reproduction of the GPU workload gains, live early stopping for Levenshtein, and K3’s KV-cache use per conversation.