Is way forward for Python numerical computation?
Late final yr, NVIDIA made a big announcement concerning the way forward for Python-based numerical computing. I wouldn’t be stunned for those who missed it. In spite of everything, each different announcement from each AI firm, then and now, appears mega-important.
That announcement launched the cuNumeric library, a drop-in alternative for the ever-present NumPy library constructed on prime of the Legate framework.
Who’re Nvidia?
Most individuals will most likely know Nvidia from their ultra-fast chips that energy computer systems and knowledge centres everywhere in the world. You may additionally be aware of Nvidia’s charismatic, leather-based jacket-loving CEO, Jensen Huang, who appears to pop up on the stage of each AI convention lately.
What many individuals don’t know is that Nvidia additionally designs and creates modern machine architectures and related software program. One in all its most prized merchandise is the Compute Unified Gadget Structure (CUDA). CUDA is NVIDIA’s proprietary parallel-computing platform and programming mannequin. Since its launch in 2007, it has advanced right into a complete ecosystem comprising drivers, runtime, compilers, math libraries, debugging and profiling instruments, and container pictures. The result’s a neatly tuned {hardware} and software program loop that retains NVIDIA GPUs on the centre of recent high-performance and AI workloads.
What’s Legate?
Legate is an NVIDIA-led open-source runtime layer that permits you to run acquainted Python data-science libraries (NumPy, cuNumeric, Pandas-style APIs, sparse linear-algebra kernels, …) on multi-core CPUs, single or multi-GPU nodes, and even multi-node clusters with out altering your Python code. It interprets high-level array operations right into a graph of fine-grained duties and palms that graph to the C++ Legion runtime, which schedules the duties, partitions the information, and strikes tiles between CPUs, GPUs and community hyperlinks for you.
In a nutshell, Legate lets acquainted single-node Python libraries scale transparently to multi-GPU, multi-node machines.
What’s cuNumeric?
cuNumeric is a drop-in alternative for NumPy whose array operations are executed by Legate’s process engine and accelerated on one or many NVIDIA GPUs (or, if no GPU is current, on all CPU cores). In observe, you put in it and want solely change one import line to start out utilizing it instead of your common NumPy code. For instance …
# previous
import numpy as np
...
...
# new
import cupynumeric as np # the whole lot else stays the identical
...
...… and run your script on the terminal with the legate command.
Behind the scenes, cuNumeric converts every NumPy name you make, for instance, np.sin, np.linalg.svd, fancy indexing, broadcasting, reductions, and so on, into Legate duties. These duties will,
- Partition your arrays into tiles sized to suit GPU reminiscence.
- Schedule every tile on one of the best out there machine (GPU or CPU).
- Overlap compute with communication when the workload spans a number of GPUs or nodes.
- Spill tiles to NVMe/SSD routinely when your dataset outruns GPU RAM.
As a result of the API of cuNumeric mirrors NumPy’s practically 1-for-1, current scientific or data-science code can scale from a laptop computer to a multi-GPU cluster and not using a rewrite.
Efficiency advantages
So, this all appears nice, proper? But it surely solely is smart if it leads to tangible efficiency enhancements over utilizing NumPy, and Nvidia is making some sturdy claims that that is the case. As knowledge scientists, machine studying engineers and knowledge engineers usually use NumPy so much, we will respect that this generally is a essential facet of the programs we write and keep.
Now, I don’t have a cluster of GPUs or a supercomputer to check this on, however my desktop PC does have an Nvidia GeForce RTX 4070 GPU, and we’re going to make use of that to check out a few of Nvidia’s claims.
(base) tom@tpr-desktop:~$ nvidia-smi
Solar Jun 15 15:26:36 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 565.75 Driver Model: 566.24 CUDA Model: 12.7 |
|-----------------------------------------+------------------------+----------------------+
| GPU Title Persistence-M | Bus-Id Disp.A | Risky Uncorr. ECC |
| Fan Temp Perf Pwr:Utilization/Cap | Reminiscence-Utilization | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 4070 Ti On | 00000000:01:00.0 On | N/A |
| 32% 29C P8 9W / 285W | 1345MiB / 12282MiB | 2% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Kind Course of title GPU Reminiscence |
| ID ID Utilization |
|=========================================================================================|
| No working processes discovered |
+-----------------------------------------------------------------------------------------+I’ll set up cuNumeric and NumPy on my PC to conduct comparative assessments. This can assist us assess whether or not Nvidia’s claims are correct and perceive the efficiency variations between the 2 libraries.
Establishing a improvement surroundings.
As all the time, I prefer to arrange a separate improvement surroundings to run my assessments. That method, nothing I do in that surroundings will have an effect on any of my different initiatives. On the time of writing, cuNumeric will not be out there to put in on Home windows, so I’ll be utilizing WSL2 Ubuntu for Home windows as an alternative.
I’ll be utilizing Miniconda to arrange my surroundings, however be happy to make use of whichever software you’re snug with.
$ conda create cunumeric-env python=3.10 -c conda-forge
$ conda activate cunumeric-env
$ conda set up -c conda-forge -c legate cupynumeric
$ conda set up -c conda-forge ucx cuda-cudart cuda-version=12Code instance 1 — A easy matrix multiplication
Matrix multiplication is the bread and butter of mathematical operations that underpin so many AI programs, so it is smart to strive that operation out first.
Observe that in all my examples, I’ll run the NumPy and cuNumeric code snippets 5 occasions in a row and common the time taken for every. I additionally carry out a “warm-up step on the GPU earlier than the timing run to consider overheads akin to just-in-time (JIT) compilation.
import time
import gc
import argparse
import sys
def benchmark_numpy(n, runs):
"""Runs the matrix multiplication benchmark utilizing normal NumPy on the CPU."""
import numpy as np
print(f"--- NumPy (CPU) Benchmark ---")
print(f"Multiplying two {n}×{n} matrices ({runs} runs)n")
# 1. Generate knowledge ONCE earlier than the timing loop.
print(f"Producing two {n}x{n} random matrices on CPU...")
A = np.random.rand(n, n).astype(np.float32)
B = np.random.rand(n, n).astype(np.float32)
# 2. Carry out one untimed warm-up run.
print("Performing warm-up run...")
_ = np.matmul(A, B)
print("Heat-up full.n")
# 3. Carry out the timed runs.
occasions = []
for i in vary(runs):
begin = time.time()
# The operation being timed. The @ operator is a handy
# shorthand for np.matmul.
C = A @ B
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.4f}s")
del C # Clear up the outcome matrix
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"nNumPy common: {avg:.4f}sn")
return avg
def benchmark_cunumeric(n, runs):
"""Runs the matrix multiplication benchmark utilizing cuNumeric on the GPU."""
import cupynumeric as cn
import numpy as np # Import numpy for the canonical sync
print(f"--- cuNumeric (GPU) Benchmark ---")
print(f"Multiplying two {n}×{n} matrices ({runs} runs)n")
# 1. Generate knowledge ONCE on the GPU earlier than the timing loop.
print(f"Producing two {n}x{n} random matrices on GPU...")
A = cn.random.rand(n, n).astype(np.float32)
B = cn.random.rand(n, n).astype(np.float32)
# 2. Carry out a vital untimed warm-up run for JIT compilation.
print("Performing warm-up run...")
C_warmup = cn.matmul(A, B)
# The perfect observe for synchronization: drive a replica again to the CPU.
_ = np.array(C_warmup)
print("Heat-up full.n")
# 3. Carry out the timed runs.
occasions = []
for i in vary(runs):
begin = time.time()
# Launch the operation on the GPU
C = A @ B
# Synchronize by changing the outcome to a host-side NumPy array.
np.array(C)
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.4f}s")
del C
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"ncuNumeric common: {avg:.4f}sn")
return avg
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Benchmark matrix multiplication on NumPy (CPU) vs. cuNumeric (GPU)."
)
parser.add_argument(
"-n", "--n", sort=int, default=3000, assist="Matrix dimension (n x n)"
)
parser.add_argument(
"-r", "--runs", sort=int, default=5, assist="Variety of timing runs"
)
parser.add_argument(
"--cunumeric", motion="store_true", assist="Run the cuNumeric (GPU) model"
)
args, unknown = parser.parse_known_args()
# The dispatcher logic
if args.cunumeric or "--cunumeric" in unknown:
benchmark_cunumeric(args.n, args.runs)
else:
benchmark_numpy(args.n, args.runs)Operating the NumPy aspect of issues makes use of the common python example1.py command line syntax. For working utilizing Legate, the syntax is extra complicated. What it does is disable Legate’s automated configuration after which launch the example1.py script below Legate with one CPU, one GPU, and 0 OpenMP threads utilizing the cuNumeric backend.
Right here is the output.
(cunumeric-env) tom@tpr-desktop:~$ python example1.py
--- NumPy (CPU) Benchmark ---
Multiplying two 3000×3000 matrices (5 runs)
Producing two 3000x3000 random matrices on CPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 0.0976s
Run 2: time = 0.0987s
Run 3: time = 0.0957s
Run 4: time = 0.1063s
Run 5: time = 0.0989s
NumPy common: 0.0994s
(cunumeric-env) tom@tpr-desktop:~$ LEGATE_AUTO_CONFIG=0 legate --cpus 1 --gpus 1 --omps 0 example1.py --cunu
meric
[0 - 7f2e8fcc8480] 0.000000 {5}{module_config}: Module numa can't detect sources.
[0 - 7f2e8fcc8480] 0.000000 {4}{topology}: cannot open /sys/units/system/node/
[0 - 7f2e8fcc8480] 0.000049 {4}{threads}: reservation ('GPU ctxsync 0x55cd5fd34530') can't be glad
--- cuNumeric (GPU) Benchmark ---
Multiplying two 3000×3000 matrices (5 runs)
Producing two 3000x3000 random matrices on GPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 0.0113s
Run 2: time = 0.0089s
Run 3: time = 0.0086s
Run 4: time = 0.0090s
Run 5: time = 0.0087s
cuNumeric common: 0.0093sProperly, that’s a powerful begin. cuNumeric is registering a 10x speedup over NumPy.
The warnings that Legate is outputting will be ignored. These are informational, indicating Legate couldn’t discover particulars concerning the machine’s CPU/reminiscence format (NUMA) or sufficient CPU cores to handle the GPU.
Code instance 2 — Logistic regression
Logistic regression is a foundational software in knowledge science as a result of it supplies a easy, interpretable approach to mannequin and predict binary outcomes (sure/no, move/fail, click on/no-click). On this instance, we’ll measure how lengthy it takes to coach a easy binary classifier on artificial knowledge. For every of the 5 runs, it first generates N samples with D options (X), and a corresponding random 0/1 label vector (Y). It initialises the burden vector w to zeros, then performs 500 iterations of batch gradient descent: computing the linear predictions z = X.dot(w), making use of the sigmoid p = 1/(1+exp(–z)), computing the gradient grad = X.T.dot(p – y) / N, and updating the weights with w -= 0.1 * grad. The script information the elapsed time for every run, cleans up reminiscence, and eventually prints the typical coaching time.
import time
import gc
import argparse
import sys
# --- Reusable Coaching Perform ---
# By placing the coaching loop in its personal operate, we keep away from code duplication.
# The `np` argument permits us to move in both the numpy or cupynumeric module.
def train_logistic_regression(np, X, y, iters, alpha):
"""Performs a set variety of gradient descent iterations."""
# Guarantee w begins on the right machine (CPU or GPU)
w = np.zeros(X.form[1])
for _ in vary(iters):
z = X.dot(w)
p = 1.0 / (1.0 + np.exp(-z))
grad = X.T.dot(p - y) / X.form[0]
w -= alpha * grad
return w
def benchmark_numpy(n_samples, n_features, iters, alpha):
"""Runs the logistic regression benchmark utilizing normal NumPy on the CPU."""
import numpy as np
print(f"--- NumPy (CPU) Benchmark ---")
print(f"Coaching on {n_samples} samples, {n_features} options for {iters} iterationsn")
# 1. Generate knowledge ONCE earlier than the timing loop.
print("Producing random dataset on CPU...")
X = np.random.rand(n_samples, n_features)
y = (np.random.rand(n_samples) > 0.5).astype(np.float64)
# 2. Carry out one untimed warm-up run.
print("Performing warm-up run...")
_ = train_logistic_regression(np, X, y, iters, alpha)
print("Heat-up full.n")
# 3. Carry out the timed runs.
occasions = []
for i in vary(args.runs):
begin = time.time()
# The operation being timed
_ = train_logistic_regression(np, X, y, iters, alpha)
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.3f}s")
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"nNumPy common: {avg:.3f}sn")
return avg
def benchmark_cunumeric(n_samples, n_features, iters, alpha):
"""Runs the logistic regression benchmark utilizing cuNumeric on the GPU."""
import cupynumeric as cn
import numpy as np # Additionally import numpy for the canonical synchronization
print(f"--- cuNumeric (GPU) Benchmark ---")
print(f"Coaching on {n_samples} samples, {n_features} options for {iters} iterationsn")
# 1. Generate knowledge ONCE on the GPU earlier than the timing loop.
print("Producing random dataset on GPU...")
X = cn.random.rand(n_samples, n_features)
y = (cn.random.rand(n_samples) > 0.5).astype(np.float64)
# 2. Carry out a vital untimed warm-up run for JIT compilation.
print("Performing warm-up run...")
w_warmup = train_logistic_regression(cn, X, y, iters, alpha)
# The perfect observe for synchronization: drive a replica again to the CPU.
_ = np.array(w_warmup)
print("Heat-up full.n")
# 3. Carry out the timed runs.
occasions = []
for i in vary(args.runs):
begin = time.time()
# Launch the operation on the GPU
w = train_logistic_regression(cn, X, y, iters, alpha)
# Synchronize by changing the ultimate outcome again to a NumPy array.
np.array(w)
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.3f}s")
del w
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"ncuNumeric common: {avg:.3f}sn")
return avg
if __name__ == "__main__":
# A extra sturdy argument parsing setup
parser = argparse.ArgumentParser(
description="Benchmark logistic regression on NumPy (CPU) vs. cuNumeric (GPU)."
)
# Hyperparameters for the mannequin
parser.add_argument(
"-n", "--n_samples", sort=int, default=2_000_000, assist="Variety of knowledge samples"
)
parser.add_argument(
"-d", "--n_features", sort=int, default=10, assist="Variety of options"
)
parser.add_argument(
"-i", "--iters", sort=int, default=500, assist="Variety of gradient descent iterations"
)
parser.add_argument(
"-a", "--alpha", sort=float, default=0.1, assist="Studying fee"
)
# Benchmark management
parser.add_argument(
"-r", "--runs", sort=int, default=5, assist="Variety of timing runs"
)
parser.add_argument(
"--cunumeric", motion="store_true", assist="Run the cuNumeric (GPU) model"
)
args, unknown = parser.parse_known_args()
# Dispatcher logic
if args.cunumeric or "--cunumeric" in unknown:
benchmark_cunumeric(args.n_samples, args.n_features, args.iters, args.alpha)
else:
benchmark_numpy(args.n_samples, args.n_features, args.iters, args.alpha)And the outputs.
(cunumeric-env) tom@tpr-desktop:~$ python example2.py
--- NumPy (CPU) Benchmark ---
Coaching on 2000000 samples, 10 options for 500 iterations
Producing random dataset on CPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 12.292s
Run 2: time = 11.830s
Run 3: time = 11.903s
Run 4: time = 12.843s
Run 5: time = 11.964s
NumPy common: 12.166s
(cunumeric-env) tom@tpr-desktop:~$ LEGATE_AUTO_CONFIG=0 legate --cpus 1 --gpus 1 --omps 0 example2.py --cunu
meric
[0 - 7f04b535c480] 0.000000 {5}{module_config}: Module numa can't detect sources.
[0 - 7f04b535c480] 0.000000 {4}{topology}: cannot open /sys/units/system/node/
[0 - 7f04b535c480] 0.001149 {4}{threads}: reservation ('GPU ctxsync 0x55fb037cf140') can't be glad
--- cuNumeric (GPU) Benchmark ---
Coaching on 2000000 samples, 10 options for 500 iterations
Producing random dataset on GPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 1.964s
Run 2: time = 1.957s
Run 3: time = 1.968s
Run 4: time = 1.955s
Run 5: time = 1.960s
cuNumeric common: 1.961sNot fairly as spectacular as our first instance, however a 5x to 6x speedup on an already quick NumPy program is to not be sniffed at.
Code instance 3 — fixing linear equations
This script benchmarks how lengthy it takes to resolve a dense 3000×3000 linear algebra equation system. This can be a elementary operation in linear algebra used to resolve the equation of sort Ax = b, the place A is a big grid of numbers (a 3000×3000 matrix on this case), and b is an inventory of numbers (a vector).
The objective is to search out the unknown checklist of numbers x that makes the equation true. This can be a computationally intensive process that’s on the coronary heart of many scientific simulations, engineering issues, monetary fashions, and even some AI algorithms.
import time
import gc
import argparse
import sys # Import sys to verify arguments
# Observe: The library imports (numpy and cupynumeric) are actually carried out *inside*
# their respective capabilities to maintain them separate and keep away from import errors.
def benchmark_numpy(n, runs):
"""Runs the linear clear up benchmark utilizing normal NumPy on the CPU."""
import numpy as np
print(f"--- NumPy (CPU) Benchmark ---")
print(f"Fixing {n}×{n} A x = b ({runs} runs)n")
# 1. Generate knowledge ONCE earlier than the timing loop.
print("Producing random system on CPU...")
A = np.random.randn(n, n).astype(np.float32)
b = np.random.randn(n).astype(np.float32)
# 2. Carry out one untimed warm-up run. That is good observe even for
# the CPU to make sure caches are heat and any one-time setup is finished.
print("Performing warm-up run...")
_ = np.linalg.clear up(A, b)
print("Heat-up full.n")
# 3. Carry out the timed runs.
occasions = []
for i in vary(runs):
begin = time.time()
# The operation being timed
x = np.linalg.clear up(A, b)
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.6f}s")
# Clear up the outcome to be protected with reminiscence
del x
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"nNumPy common: {avg:.6f}sn")
return avg
def benchmark_cunumeric(n, runs):
"""Runs the linear clear up benchmark utilizing cuNumeric on the GPU."""
import cupynumeric as cn
import numpy as np # Additionally import numpy for the canonical synchronization
print(f"--- cuNumeric (GPU) Benchmark ---")
print(f"Fixing {n}×{n} A x = b ({runs} runs)n")
# 1. Generate knowledge ONCE on the GPU earlier than the timing loop.
# This ensures we're not timing the information switch in our major loop.
print("Producing random system on GPU...")
A = cn.random.randn(n, n).astype(np.float32)
b = cn.random.randn(n).astype(np.float32)
# 2. Carry out a vital untimed warm-up run. This handles JIT
# compilation and different one-time GPU setup prices.
print("Performing warm-up run...")
x_warmup = cn.linalg.clear up(A, b)
# The perfect observe for synchronization: drive a replica again to the CPU.
_ = np.array(x_warmup)
print("Heat-up full.n")
# 3. Carry out the timed runs.
occasions = []
for i in vary(runs):
begin = time.time()
# Launch the operation on the GPU
x = cn.linalg.clear up(A, b)
# Synchronize by changing the outcome to a host-side NumPy array.
# That is assured to dam till the GPU has completed.
np.array(x)
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.6f}s")
# Clear up the GPU array outcome
del x
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"ncuNumeric common: {avg:.6f}sn")
return avg
if __name__ == "__main__":
# A extra sturdy argument parsing setup
parser = argparse.ArgumentParser(
description="Benchmark linear clear up on NumPy (CPU) vs. cuNumeric (GPU)."
)
parser.add_argument(
"-n", "--n", sort=int, default=3000, assist="Matrix dimension (n x n)"
)
parser.add_argument(
"-r", "--runs", sort=int, default=5, assist="Variety of timing runs"
)
# Use parse_known_args() to deal with potential additional arguments from Legate
args, unknown = parser.parse_known_args()
# The dispatcher logic: verify if "--cunumeric" is within the command line
# This can be a easy and efficient approach to change between modes.
if "--cunumeric" in sys.argv or "--cunumeric" in unknown:
benchmark_cunumeric(args.n, args.runs)
else:
benchmark_numpy(args.n, args.runs)The outputs.
(cunumeric-env) tom@tpr-desktop:~$ python example4.py
--- NumPy (CPU) Benchmark ---
Fixing 3000×3000 A x = b (5 runs)
Producing random system on CPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 0.133075s
Run 2: time = 0.126129s
Run 3: time = 0.135849s
Run 4: time = 0.137383s
Run 5: time = 0.138805s
NumPy common: 0.134248s
(cunumeric-env) tom@tpr-desktop:~$ LEGATE_AUTO_CONFIG=0 legate --cpus 1 --gpus 1 --omps 0 example4.py --cunumeric
[0 - 7f29f42ce480] 0.000000 {5}{module_config}: Module numa can't detect sources.
[0 - 7f29f42ce480] 0.000000 {4}{topology}: cannot open /sys/units/system/node/
[0 - 7f29f42ce480] 0.000053 {4}{threads}: reservation ('GPU ctxsync 0x562e88c28700') can't be glad
--- cuNumeric (GPU) Benchmark ---
Fixing 3000×3000 A x = b (5 runs)
Producing random system on GPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 0.009685s
Run 2: time = 0.010043s
Run 3: time = 0.009966s
Run 4: time = 0.009739s
Run 5: time = 0.009383s
cuNumeric common: 0.009763sThat could be a super outcome. The Nvidia cuNumeric run is 100x quicker than the NumPy run.
Code instance 4 — Sorting
Sorting is such a elementary a part of the whole lot that occurs in computing, and fashionable computer systems are so quick that almost all builders don’t even give it some thought. However let’s see how a lot of a distinction utilizing cuNumeric could make to this ubiquitous operation. We’ll type a big (30,000,000) 1D array of numbers
# benchmark_sort.py
import time
import sys
import gc
# Array dimension
n = 30_000_000 # 30 million parts
def benchmark_numpy():
import numpy as np
print(f"Sorting an array of {n} parts with NumPy (5 runs)n")
occasions = []
for i in vary(5):
knowledge = np.random.randn(n).astype(np.float32)
begin = time.time()
_ = np.type(knowledge)
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.6f}s")
del knowledge
gc.gather()
avg = sum(occasions) / len(occasions)
print(f"nNumPy common: {avg:.6f}sn")
def benchmark_cunumeric():
import cupynumeric as np
print(f"Sorting an array of {n} parts with cuNumeric (5 runs)n")
occasions = []
for i in vary(5):
knowledge = np.random.randn(n).astype(np.float32)
begin = time.time()
_ = np.type(knowledge)
# Power GPU sync
_ = np.linalg.norm(np.zeros(()))
finish = time.time()
period = finish - begin
occasions.append(period)
print(f"Run {i+1}: time = {period:.6f}s")
del knowledge
gc.gather()
_ = np.linalg.norm(np.zeros(()))
avg = sum(occasions) / len(occasions)
print(f"ncuNumeric common: {avg:.6f}sn")
if __name__ == "__main__":
if "--cunumeric" in sys.argv:
benchmark_cunumeric()
else:
benchmark_numpy()The outputs.
(cunumeric-env) tom@tpr-desktop:~$ python example5.py
--- NumPy (CPU) Benchmark ---
Sorting an array of 30000000 parts (5 runs)
Creating random array on CPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 0.588777s
Run 2: time = 0.586813s
Run 3: time = 0.586745s
Run 4: time = 0.586525s
Run 5: time = 0.583783s
NumPy common: 0.586529s
-----------------------------
(cunumeric-env) tom@tpr-desktop:~$ LEGATE_AUTO_CONFIG=0 legate --cpus 1 --gpus 1 --omps 0 example5.py --cunumeric
[0 - 7fd9e4615480] 0.000000 {5}{module_config}: Module numa can't detect sources.
[0 - 7fd9e4615480] 0.000000 {4}{topology}: cannot open /sys/units/system/node/
[0 - 7fd9e4615480] 0.000082 {4}{threads}: reservation ('GPU ctxsync 0x564489232fd0') can't be glad
--- cuNumeric (GPU) Benchmark ---
Sorting an array of 30000000 parts (5 runs)
Creating random array on GPU...
Performing warm-up run...
Heat-up full.
Run 1: time = 0.010857s
Run 2: time = 0.007927s
Run 3: time = 0.007921s
Run 4: time = 0.008240s
Run 5: time = 0.007810s
cuNumeric common: 0.008551s
-------------------------------Yet one more vastly spectacular efficiency from cuNumeric and Legate.
Abstract
This text launched cuNumeric, an NVIDIA library designed as a high-performance, drop-in alternative for NumPy. The important thing takeaway is that knowledge scientists can speed up their current Python code on NVIDIA GPUs with minimal effort, typically by merely altering a single import line and working the script with the ‘legate’ command.
Two major parts energy the know-how:
- Legate: An open-source runtime layer from NVIDIA that routinely interprets high-level Python operations into duties. It intelligently manages distributing these duties throughout single or a number of GPUs, dealing with knowledge partitioning, reminiscence administration (even spilling to disk if wanted), and optimising communication.
- cuNumeric: The user-facing library that mirrors the NumPy API. If you make a name like np.matmul(), cuNumeric converts it right into a process for the Legate engine to execute on the GPU.
I used to be capable of validate Nvidia’s efficiency claims by working 4 benchmark assessments on my desktop PC (with an NVIDIA RTX 4070 Ti GPU), evaluating normal NumPy on the CPU in opposition to cuNumeric on the GPU.
The outcomes exhibit important efficiency positive aspects for cuNumeric:
- Matrix Multiplication: ~10x quicker than NumPy.
- Logistic Regression Coaching: ~6x quicker.
- Fixing Linear Equations: An enormous 100x+ speedup.
- Sorting a Giant Array: One other large enchancment, working roughly 70x quicker.
In conclusion, I confirmed that cuNumeric efficiently delivers on its promise, making the immense computational energy of GPUs accessible to the broader Python knowledge science neighborhood with out requiring a steep studying curve or a whole code rewrite.
For extra data and hyperlinks to associated sources, take a look at the unique Nvidia announcement on cuNumeric right here.
