Oliver: a Frontier Database™.
Frontier models run on GPUs, and now the database they ask does too. On one GPU, Oliver answers over 6,000 questions a second across 100 million rows, up to 2,529 times faster than asking them one at a time.
Achilles and the tortoise
Zeno of Elea set Achilles a race he could never win. The tortoise starts ahead. By the time Achilles reaches the spot where it was, it has crawled a little farther, and by the time he reaches that spot, it has moved again. There is always one more gap to close, so, Zeno argued, the fastest runner in Greece never catches a tortoise.
He does, of course. The gaps shrink so quickly that all of them together add up to a short, finite stretch of time, and Achilles is past the tortoise before anyone in the crowd notices there was a paradox.
An agent working with a database runs Zeno's race. It asks a question, waits for the answer, and the answer raises three more. Each question is a fresh trip across the data, and by the time the database gets back, the agent has moved on and more data has arrived. A database that answers one question at a time is Achilles on Zeno's terms, forever closing the last gap.
The way out is the same one that settles the paradox. Stop paying for each gap separately. Take every question at once, cross the data once, and the whole endless series of trips comes to one short one.
Agents ask faster than databases answer
Frontier models run on accelerators and work in batches. The databases they query were mostly designed for one person asking one question on a CPU. An agent investigating an incident generates dozens to hundreds of questions, a detection pack checks every rule against every event, and a fleet check looks at every series at once. Asked one at a time, each of those questions pays for its own trip through the data.
Oliver is a Frontier Database™. It is built for the way frontier AI asks questions, and it runs on the hardware frontier AI runs on.
Every question in one pass
A question reaches Oliver as a program rather than a string, so Oliver can take hundreds of them together, work out what they share, and answer all of them in a single pass over the data. On a live tenant, 500 questions over one hour of data took 61.6 seconds asked one at a time and 0.57 seconds asked together. On CPU alone, answering together runs 13 to 262 times faster than answering one at a time.
All of that is on CPU. Then we moved the pass onto a GPU.
The pass moves onto the GPU
A single pass over a lot of data is exactly the work a GPU is built for: one small operation applied to an enormous number of rows at the same time. We put 100 million rows into the memory of one NVIDIA L40S and asked it thousands of questions at a time.
The baseline in every one of those comparisons is Oliver itself, answering one question at a time on the same machine. That is the engine that beat ClickHouse by a median of 3.06×. The GPU runs up to 2,529 times faster than it.
For a sense of scale: a thousand pattern-matching questions take the GPU about a sixth of a second. Asked one at a time, the same thousand take close to seven minutes.
The data stays in GPU memory
Moving data onto a GPU is the slow part, so Oliver does it once. The 100-million-row window loads into GPU memory and stays there between requests. The next thousand questions start straight away, against data that is already sitting where the work happens.
A whole rule pack in one call
Detection rules are cheap to write and expensive to run, so teams ration them. On one L40S, a pack of 500 detection rules runs at about 1,300 questions a second, 5.1 times what the same machine manages on its CPU. Every rule can check every event, every time, instead of the dozen someone had time to schedule.
A whole fleet in under two seconds
Anomaly detection has the same shape at a larger size. We screened 15 million time series of 128 points each, 1.92 billion points in all, for anomalies on the GPU in 1.66 seconds. That covers every metric in a very large fleet, checked in one go rather than a sample of them on a rota.
The same answers on any GPU
A fast answer is only worth having if it is the right one. Every GPU result is checked against what Oliver's CPU engine returns for the same question, and they agree, with sums carried at 128-bit precision. The tier runs on NVIDIA GPUs and on Apple silicon, so it is not tied to one vendor's hardware.
Running it in your own cloud
The GPU tier runs in Oliver deployed inside your own cloud account, on your GPUs, next to your data. Tell us what you want to ask of it, and we will set it up with you.