Skip to content

Repository files navigation

kernelmeter

PyPI CI Python License: MIT

which gpu is worth renting, computed from rooflines

Try that right now, no install and no GPU required:

uvx kernelmeter compare 4090 a100-80gb h100-sxm --ai 0.33

Or skip the terminal entirely: nuemaan.github.io/kernelmeter runs the comparison in your browser, and the settings live in the URL so you can send someone the exact tradeoff you mean.

Small tools for two questions: is my GPU kernel actually good, and which GPU is actually worth paying for? All in one package with zero required dependencies.

  • kernelmeter info prints every device attribute your GPU driver knows, plus the card's theoretical peak bandwidth and FP32 throughput. No CUDA toolkit, no torch, no kernel launch. It reads straight from libcuda, which is part of the NVIDIA driver.
  • kernelmeter bench times your kernel, checks the output against a reference, and scores it against the roofline: the best your card could possibly do for that kernel's mix of math and memory traffic. 240 GB/s means nothing on its own; "76% of attainable" tells you how much room is left. While the kernel runs it also samples the real clocks, power and temperature through NVML, and re-scores against the ceiling the card actually held.
  • kernelmeter roofline draws your card's roofline in the terminal and shows where a kernel sits on it.
  • kernelmeter occupancy answers "why is my occupancy 50%?" from block size, registers and shared memory, and shows which block sizes fix it.
  • kernelmeter ceiling measures what the card really delivers (STREAM-style bandwidth tests plus a big FP32 matmul), because spec sheet numbers are never fully reachable.
  • kernelmeter compare 4090 h100-sxm --ai 0.33 tells you which card is actually faster for your kernel, and per rental dollar. Works with no GPU at all: there is a built-in database of 40 cards, NVIDIA and AMD, whose specs are unit-tested against the vendor sheets.
  • kernelmeter llm 70b --quant q4 --gpus 4090 a100-80gb answers the question everyone actually has: does it fit, and what is the hard ceiling on tokens per second. Same roofline math, zero benchmarks.
  • kernelmeter report writes a single-file HTML report card for your GPU that you can share, attach to an issue, or keep as a record.

Install

pip install kernelmeter           # info only, no dependencies
pip install "kernelmeter[bench]"  # adds torch for the bench harness

Or from source:

git clone https://github.com/nuemaan/kernelmeter
cd kernelmeter
pip install -e ".[bench]"

No GPU handy? Start here

Deciding what to rent is a roofline question, and you can answer it from any laptop. Say your kernel does 0.33 flop per byte (a fused elementwise op) and you're choosing between renting cards:

kernelmeter compare 4090 a100-80gb h100-sxm --ai 0.33 --cost 4090=0.35,a100-80gb=1.19,h100-sxm=2.69
card            bw GB/s  fp32 TF  fp16 TC  ridge  @0.33 TF  vs rtx-4090   $/hr  TF per $
----------------------------------------------------------------------------------------
rtx-4090           1008     82.6      330   81.9      0.33        1.00x   0.35      0.95
a100-80gb          2039     19.5      312    9.6      0.67        2.02x   1.19      0.57
h100-sxm           3347     66.9     1071   20.0      1.10        3.32x   2.69      0.41

at this intensity every card is memory-bound: bandwidth is what you're buying

Read the last two columns: the H100 is 3.3x faster in absolute terms, but for this kernel the 4090 delivers more than twice the throughput per rental dollar. The overlaid rooflines print below the table so you can see why: left of every ridge point, only the bandwidth line matters.

kernelmeter gpus lists the built-in database: NVIDIA from the T4 to the RTX 5090 plus A100/H100 and the workstation cards, and AMD from the MI100 to the MI300X plus RX 6900 XT through RX 9070 XT. Every entry stores physical parameters (SMs or CUs, clocks, bus width, per-pin rate) and derives its peaks through per-architecture rate tables, and the test suite asserts each derived number against the vendor spec sheet, so a wrong entry fails CI. kernelmeter roofline --gpu 4090 (or --gpu mi300x) draws any card's roofline the same way, and cross-vendor comparisons just work:

kernelmeter llm 70b --quant q4 --gpus mi300x h100-sxm --cost mi300x=2.49,h100-sxm=2.69

puts a 70B on a single MI300X (192 GB, ~131 t/s decode ceiling) against a single H100 (fits at 80 GB, ~82 t/s), priced per token.

Will it run a 70B, and how fast at best?

LLM decode is the memory-bound case of the same roofline: generating one token reads every weight once, so the hard ceiling is bandwidth divided by weight bytes. Prefill is the compute-bound case, about 2 FLOPs per parameter per token. That means honest upper bounds need no benchmark at all:

kernelmeter llm 70b --quant q4 --gpus 4090 a100-80gb h100-sxm --cost a100-80gb=1.19,h100-sxm=2.69
70b model at q4 (~0.58 bytes/param): 40.6 GB of weights

card                         vram  fits  decode t/s  prefill t/s   $/hr  t/s per $
----------------------------------------------------------------------------------
rtx-4090                     24GB    no           -            -      -          -
a100-80gb                    80GB   yes          50         2228   1.19       42.2
h100-sxm                     80GB   yes          82         7647   2.69       30.6

these are roofline ceilings, not predictions: well-tuned stacks land at 50-85%
of them, none land above. kv cache and activations need room on top of the
weights (2 GB per gpu assumed here).
geforce/titan prefill uses the fp32-accumulate tensor rate (half the fp16 peak).

The honesty note is the point. If someone quotes you 120 tok/s for a 70B q4 on one A100, the ceiling says that's physically impossible; if your own stack gets 20, the ceiling says you're leaving half on the table. Quant sizes use effective bytes per parameter (gguf k-quants carry scales, so q4 is ~0.58, not 0.5); pass --bytes-per-param for an exact figure. With no --gpus it estimates for the GPU in your machine, using its real memory size for the fit check.

It also handles the setups people actually run:

  • --num-gpus 2 models the classic budget rig. Two 3090s hold a 70B q4 that one can't, at a ~46 t/s ceiling, and at rental prices they come out around 2.5x the tokens per dollar of an A100 pair.
  • --active-params 37b handles MoE models: fit needs the full weights, but each token only reads the active experts, which is why a 671B DeepSeek decodes faster than a dense 70B on the same hardware.
  • --batch 32 shows throughput ceilings: weight reads amortize across concurrent streams until the compute roof takes over, and the table splits total t/s from per-stream t/s. The crossover batch size is the ridge point again, just wearing different clothes.
  • --per-watt adds a tokens-per-watt column from the cards' TDP.

GeForce and Titan prefill ceilings use the fp32-accumulate tensor rate, half the marketing fp16 number, because that is what inference stacks actually do. Datacenter cards run full rate either way.

Querying your GPU

kernelmeter info

Output from a Tesla T4:

CUDA driver version : 13.0

Device 0: Tesla T4 (14.6 GiB)
  compute capability        : 7.5
  theoretical mem bandwidth : 320.1 GB/s
  theoretical FP32 peak     : 8.14 TFLOP/s
  theoretical fp16 tensor   : 65.13 TFLOP/s (dense)
  architecture (nvml)       : Turing, 2560 CUDA cores
  pcie link (nvml)          : gen1/3 x8/16
  memory in use (nvml)      : 450 / 15360 MiB
  ecc (nvml)                : on
  vbios (nvml)              : 90.04.96.00.02

  attribute                                        value
  ------------------------------------------------ ------------
  max_threads_per_block                            1024
  max_block_dim_x                                  1024
  max_shared_memory_per_block                      49152
  warp_size                                        32
  clock_rate_khz                                   1590000
  ...                                              (147 attributes total)

The attribute table is read straight from the driver via cuDeviceGetAttribute, the same values Nsight Compute shows as device__attribute_*, but you don't need to profile a kernel to see them. Every id is probed live, so the output matches the machine you run it on; ids newer than the bundled name table show up as attribute_<id>.

The (nvml) lines come from a second source: NVML, the library behind nvidia-smi, also shipped with the driver. They surface facts the driver attribute enum doesn't have (architecture name, real CUDA core count, PCIe link, live memory use, ECC, VBIOS) and are skipped silently if NVML isn't present. (The gen1/3 x8/16 above is the live link: an idle T4 drops to a lower PCIe state and ramps up under load.) Add --json for machine-readable output; the NVML block lands under devices[].nvml.

On an AMD machine the same command goes through the HIP runtime and rocm-smi instead. Real output from an MI300X:

HIP runtime version : 70051831

Device 0: AMD Instinct MI300X VF (191.7 GiB)
  architecture              : cdna3
  theoretical mem bandwidth : 5324.8 GB/s
  theoretical FP32 peak     : 163.43 TFLOP/s
  theoretical fp16 matrix   : 1307.44 TFLOP/s (dense)
  live (rocm-smi): sclk 172 MHz, mclk 900 MHz, 147/750W, vram 286/196288 MiB

The attribute table below it uses hipDeviceAttribute_t names, generated from AMD's headers the same way the CUDA table tracks the driver enum. One quirk found on hardware: the HIP runtime reports HBM3's quarter-rate memory clock, so CDNA3 bandwidth uses four transfers per clock where everything else uses the usual two.

Benchmarking a kernel

Three steps.

1. Write your kernel in a file and decorate it. Anything callable from Python works: Triton kernels, custom CUDA extensions, torch.compile output, CuPy. Here is a complete file you can copy:

# mybench.py
import torch
import kernelmeter as km

N = 1 << 26  # work on big inputs so you measure memory, not cache

def make_args():
    return (torch.randn(N, device="cuda"), torch.randn(N, device="cuda"))

@km.benchmark(
    "my_add",
    args=make_args,                 # builds fresh inputs for the run
    ref=torch.add,                  # trusted implementation to compare with
    bytes_per_call=lambda x, y: 3 * x.numel() * x.element_size(),
)
def my_add(x, y):
    return x + y                    # <- replace with your kernel

bytes_per_call is how much memory the algorithm has to move (here: read x, read y, write the result). The tool divides it by measured time to get your effective bandwidth.

2. Run it.

kernelmeter bench mybench.py

3. Read the result. From a T4, with the add written as a Triton kernel:

kernel                    median ms      GB/s   TFLOP/s  bound    %roof   vs ref  correct
------------------------------------------------------------------------------------------
my_add                       3.2725     246.1         -    mem    76.9%    1.03x     PASS
  • correct - your output matched the reference. If this says FAIL, nothing else on the line matters.
  • bound - whether the memory system (mem) or the ALUs (comp) limit this kernel, decided by its arithmetic intensity (flops per byte).
  • %roof - how close you are to the best this card could possibly do for that intensity. This is the score to improve. Above ~80% there is little left to win.
  • vs ref - speedup over the reference implementation.

Pass flops_per_call too and the roofline model places your kernel precisely; pass peak_tflops=... if your kernel runs on tensor cores so it gets judged against the right ceiling (kernelmeter info prints the derived fp16/tf32 tensor peaks for your card). Raw %peak bw and %fp32 numbers are always in the --json output.

When NVML is available, a second table follows with what the card was doing during each measurement:

telemetry                    sm MHz   mem MHz   temp   power  %roof@clk
-----------------------------------------------------------------------
my_add                    1062/1590      5000    42C   53.1W      76.9%

%roof@clk is the same roofline score, but against the ceiling at the clocks the card actually held. If %roof looks bad but %roof@clk is high, your kernel is fine: the card is thermal or power limited, and no amount of kernel work will change that. A real example, cuBLAS fp32 matmul on a 70 W T4:

kernel                    median ms      GB/s   TFLOP/s  bound   %roof   vs ref  correct
----------------------------------------------------------------------------------------
fp32_matmul                 32.0354       6.3      4.29   comp   52.7%        -        -

telemetry                    sm MHz   mem MHz   temp   power  %roof@clk
-----------------------------------------------------------------------
fp32_matmul                877/1590      5000    46C   70.4W      95.5%

53% of peak looks like a kernel problem. The telemetry shows it is not: the card hit its 70 W power limit and dropped to 877 MHz, and at those clocks the kernel was at 95.5% of what the silicon could deliver. cuBLAS was never the problem.

Timing uses CUDA events with warmup, and the L2 cache is flushed between iterations so small workloads can't fake huge bandwidth numbers from cache hits. Pass --no-flush-l2 if you want cache-hot numbers.

The examples folder has ready-to-run starting points: two Triton kernels (vector add, fused softmax) and a compute-bound matmul.

Seeing the roofline

kernelmeter roofline --ai 0.33        # mark a kernel at 0.33 flop/byte
Device 0: Tesla T4
  peak bandwidth : 320.1 GB/s
  peak compute   : 8.14 TFLOP/s (fp32)
  ridge point    : 25.4 flop/byte

8.14 TF/s |                                      **x*****************
          |                                   ***
          |                               ****
          |                            ***
          |                        ****
          |                    ****
          |                 ***
          |             ****
          |          ***
          |      *o**
          |  ****
          |**
          +----------------------------------------------------------
           2^-3            2^0            2^3             2^6          flop/byte

at 0.33 flop/byte the kernel is memory-bound; attainable: 0.11 TFLOP/s

The o is your kernel, the x is the ridge point. Left of the ridge, more FLOPs are free: the memory traffic is the bill you are paying anyway. That is the whole argument for kernel fusion, in one picture. No GPU around? --peak-bw and --peak-tflops let you draw any card. --tensor swaps in the fp16 tensor-core roof, which moves the ridge point far to the left; that picture explains why tensor-core kernels are almost always memory-bound.

Why is my occupancy low?

Feed it what ptxas -v or Nsight Compute tells you about your kernel:

kernelmeter occupancy --block 256 --regs 64 --smem 8192 --cc 8.6
occupancy for compute capability 8.6
  block=256 regs/thread=64 smem/block=8192

  occupancy    : 66.7% (32/48 warps per SM)
  blocks per SM: 4
  limited by   : registers

  block size      64    128    192    256    384    512    768   1024
  occupancy      46%    67%    62%    67%    50%    67%    50%    67%

It names the resource that is capping you and sweeps block sizes so you can see if a different launch shape helps. Works with no GPU present: pass --cc for any architecture from 7.0 (Volta) to 12.x (Blackwell).

What can the card really do?

Theoretical peaks assume the max boost clock, which the card cannot hold. Measure the real ceilings once and judge your kernels against those:

kernelmeter ceiling

This runs the four STREAM kernels (copy, scale, add, triad) and a large TF32-disabled matmul. On the same T4:

test            median ms      GB/s   TFLOP/s  % of theoretical
---------------------------------------------------------------
copy               1.1495     233.5         -             73.0%
scale              1.1674     230.0         -             71.8%
add                1.6903     238.2         -             74.4%
triad              1.6878     238.6         -             74.5%
fp32 matmul        3.5563         -      4.83             59.3%

measured bandwidth ceiling: 238.6 GB/s (use this as the honest 100%
for memory-bound kernels)

This reframes the bench results above: the vector add that scored "76.9% of theoretical" was moving 246 GB/s on a card whose memory system tops out at 238.6 GB/s in practice. It was already saturated. Without the measured ceiling you would have kept optimizing a finished kernel.

Catching regressions

kernelmeter bench mykernels.py --save baseline.json
# ...edit your kernels...
kernelmeter bench mykernels.py --compare baseline.json

The compare run prints a delta column per kernel and exits non-zero if anything got more than 5% slower, so it slots straight into CI.

A report you can share

kernelmeter report                 # from the local device
kernelmeter report --gpu 4090      # from the card database

Writes a single self-contained HTML file: the card's peak numbers as tiles, an SVG roofline, the NVML facts and the launch limits. No javascript, no external assets, dark theme. Attach it to a bug report, drop it in your team wiki, or keep one per machine in your cluster docs.

A workflow that works

If you are learning CUDA (say, working through the PMPP book) and wondering whether your kernels are any good:

  1. Run kernelmeter info and kernelmeter ceiling once. Now you know your card's real limits.
  2. Benchmark your kernel with bytes_per_call and flops_per_call set. The bound column tells you which resource you are fighting.
  3. %roof under ~60%? If the kernel is memory-bound, check occupancy first: too few warps in flight cannot hide memory latency. Then open Nsight Compute. Now you know what you are looking for, instead of staring at forty unfamiliar counters.
  4. %roof above ~80%? Stop optimizing this kernel. The next win is algorithmic (fuse it with a neighbor, move less data), and the roofline chart shows why: left of the ridge, FLOPs are free.

Caveats

  • Theoretical peaks are computed from the max boost clock the driver reports. Sustained clocks under load are lower; the telemetry table and kernelmeter ceiling both show what you can actually reach.
  • The tensor-core peaks are dense rates with fp16 accumulate. GeForce cards run tensor cores at half rate when accumulating in fp32, and sparse rates are double; pass peak_tflops=... when those apply.
  • On AMD, info and the llm local-device path run through the HIP runtime and rocm-smi (validated on an MI300X); bench and ceiling should work through torch's ROCm build, whose cuda API is HIP, but haven't been run on AMD hardware yet. The fp16 column for AMD is matrix-core throughput on CDNA and RDNA3+, packed vector math on RDNA2.
  • The occupancy command implements the standard calculator model. Real occupancy can differ (launch bounds, driver decisions); confirm with Nsight Compute when it matters.
  • The attribute name table tracks the CUDA 13.x driver enum. Values are always read live from your driver, and ids newer than the table show up as attribute_<id> rather than being dropped. PRs that extend the table when new toolkits land are welcome.

Development

pip install -e ".[dev]"
pytest

The tests fake the driver, so they run anywhere, no GPU needed. CI runs them on plain GitHub runners. For an end-to-end check on a real GPU there is a Modal script: modal run scripts/modal_gpu_test.py. The numbers in this README come from that script on a T4.

Releases are tag-driven: bump the version in pyproject.toml, add a CHANGELOG.md entry, push a v* tag. CI tests, builds and publishes to PyPI through trusted publishing.

License

MIT

About

Query every CUDA device attribute without profiling, and benchmark kernels against your hardware's theoretical peak.

Topics

Resources

Contributing

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages