NVIDIA L4 24GB
- form
- PCIe add-in card · vendor spec
- memory
- 24 GB
- gpus
- 1
- bandwidth
- 300 GB/s · vendor spec
- runs
- 16
bestmodel.run / cloud anchors
The pool measures cards people own. These two the maintainer rented, paid for once and dated. One session on 2026-08-31, 19 jobs, and the record kept. Every number below carries its own basis, its own run count and its own date, so a reader can place a GPU they own next to a GPU they will never own.
Five things a reader has to hold before an anchor number means anything. Without them an anchor reads like a community cell, and it is not one.
Hardware the community does not own. The maintainer rented both cards for the session on 2026-08-31; the runs are the maintainer's own, dated and hashed, not a community submission and not a claim.
The host image decides the driver and the torch build. The multimodal runs record both: 2.13.0+cu130, driver host-managed. The text runs record the engine, llama.cpp, and leave the stack unstated. Nobody tuned any of it, and the log says which parts of the stack it knows.
Both cards are PCIe add-in boards. The memory bandwidth quoted for each, 300 GB/s and 600 GB/s, is a vendor figure rather than a pool measurement: it is the yardstick for reading the numbers, not one of them. The pool also holds 3 SXM-module rigs, and a number measured on a PCIe board does not transfer to an SXM module without a step in between.
The lowest reading of each multimodal batch is the first one recorded. On the image job that run read 0.58 img/s against a 6.09 img/s median; on the audio job it read 3.4×real against 4.82×real. Every cell here is a median over its runs for that reason, and the spread column shows how wide the batch was.
A rented card is not a card at home, and the log is deliberately silent on the three differences that matter: whether weights sat on a local disk or were fetched per run, whether the machine carried other work at the time, and how the card was attached beyond the vendor form factor. Nothing here is a claim about the rental host. Take an anchor as a bound on a scale, and read it against the community cell on the hardware you actually own.
Each card is one clause of the anchor: what it is, and what it produced. Cells carry the median of their runs, the basis the site ladder gives that count — measured once a cell reaches 3 runs, reported below that — and the date of the run that produced it.
Llama 3.1 8B Instruct GGUF at 4-bit, on llama.cpp, on both cards, three runs each. This is the pair a reader interpolates from: same model, same quantization, same engine, different memory bandwidth.
Decode moved 1.69× while quoted memory bandwidth moved 2.00×. The gain is sub-linear against bandwidth, which is what memory-bound decode looks like when it is not purely proportional to it. Two points on a curve, not the curve: treat the pair as a bracket to interpolate inside, never as a coefficient to multiply by.
The frozen pool snapshot of 2026-09-18 carries 2 of the 5 anchor cells: sd-turbo on NVIDIA L4 24GB, Whisper Large v3 on NVIDIA L4 24GB. The rest of that session stayed in the log. The wall of community cells is at /wall; the machine catalogue is at /hardware.
source: anchor run log 2026-08-31, 19 runs, 2026-08-31T11:46:19Z to 2026-08-31T12:22:35Z · sha256 de50e613cc5c2d8395c7443b2aeb761dfe04c0f340f440a839c905c3bf809526 · measured by the maintainer on hardware rented for the session · hardware strings reduced to the GPU model and the rental host's name dropped
No plan, no price, no provider list, no way to buy compute from this page. bestmodel.run measures hardware people own and publishes the numbers with their basis. The anchors are reference points bought once to calibrate that scale against hardware outside the community pool, and they stay on the site as dated measurements, not as an offering.