Compute
Storage
For consumers
Workloads
Why Hivenet
Sign in to Hivenet
Access your files and manage your storage account.
Store
Access your files and manage your storage account.
Compute
Launch instances and manage your compute resources.
Store and Compute still use separate accounts today. We’re working on a unified sign-in.
Just need to send a file?
Use Send directlyNew to Hivenet? Get Started
Get started with Hivenet
Choose what you want to use first.
Store
Back up photos and files across your devices.
Compute
Launch self-serve GPU compute for demanding work.
Store and Compute still use separate accounts today. We’re working on a unified account experience.
Need Hivenet for business?
Talk to salesNeed to send a file?
Open SendAlready have an account? Sign in
Contact Hivenet
Choose the path that fits your question.
General contact
Questions, support, partnerships, press, or anything else not tied to a sales conversation.
Talk to sales
Ask about Compute, storage, private AI, or a larger business deployment.
Need help with setup, billing, or troubleshooting? Start with Support for the fastest answer.
Performance claims need numbers.
Explore Hivenet benchmark results for GPU virtualization, AI inference, model performance, OCR workloads, and API latency. Each benchmark shows what was tested, where it ran, which hardware was used, and where the result does not apply.
GPU VM vs bare metal
NCCL AllReduce
RTX 5090
RTX 4090
Foundational model inference
OCR throughput
API latency
Methodology-first reporting
GPU VM vs bare metal: matched within run-to-run variance.
Customers ask whether running serious GPU work in a VM slows the workload down. Hivenet tested NCCL AllReduce on a single host with 8× NVIDIA GeForce RTX 5090 GPUs. The Compute with Hivenet VM matched the bare-metal baseline within normal run-to-run variance.
Bare metal
19.25 GB/s
NCCL AllReduce bandwidth on the 8× RTX 5090 bare-metal baseline.
Compute with Hivenet VM
19.34 GB/s
NCCL AllReduce bandwidth on the same benchmark inside a VM.
Result
+0.5%
The VM result was slightly higher, but the delta sits inside normal variance. The useful conclusion is that the VM matched bare-metal bandwidth on this test.
Benchmark areas we track.
Hivenet groups benchmark work by workload type so teams can compare the results that matter to their own use case.
GPU virtualization
VM vs bare metal
NCCL AllReduce on a single 8× RTX 5090 host, measuring whether the VM introduces a measurable single-host multi-GPU communication penalty.
Developer GPU workloads
RTX 4090 vs A100
Comparison content for 7B–8B model workloads, RAG, development, and cost-conscious inference.
Foundational model inference
Throughput and latency under load
Serving benchmarks using realistic prompt shapes, concurrency sweeps, TTFT, TPOT, inter-token latency, end-to-end latency, and throughput.
Production model performance
Model-specific performance reports
Production-oriented reports for model-specific workloads, including SLO gates, throughput curves, tail latency, and recommended concurrency settings.
OCR workloads
Falcon-OCR quality and throughput
OCR pipeline testing across layout detection, region cropping, and OCR quality/throughput for document workloads.
API performance
API latency benchmarks
Automated endpoint latency measurements across environments, tracking average, P90, P95, and P99 latency against baselines and previous versions.
What every benchmark shows.
Benchmarks are useful when the setup is clear enough to repeat, question, or compare. Hivenet benchmark pages make the test conditions visible.
Benchmark field
What it covers
Workload
What the benchmark tested
Hardware
GPU, CPU, memory, storage, network, host count
Environment
VM, bare metal, container, region, driver, CUDA, framework
Model or data
Model name, precision, dataset, prompt shape, file size, or workload input
Load profile
Concurrency, requests per second, batch size, message size, duration
Metrics
Throughput, latency, TTFT, TPOT, bandwidth, error rate, cost basis
Comparison baseline
What Hivenet is compared against and why
Result
Main result and how to read it
Limits
What the benchmark does not prove
Date
When the test ran and whether results have changed since
Swipe left to see more
What the GPU virtualization result proves, and what it does not.
The VM vs bare metal benchmark is useful because NCCL AllReduce exposes problems that softer single-GPU tests can miss. If GPU passthrough is incomplete, topology is exposed poorly, or the communication path is inefficient, multi-GPU communication often shows it quickly.
This result supports
This result does not prove
The tested VM did not show a measurable NCCL AllReduce bandwidth penalty
Every workload will perform identically to bare metal
Tuned PCIe passthrough can expose the single-host GPU path cleanly
Cross-host distributed training will behave the same
Compute with Hivenet can be credible for single-host multi-GPU work
Storage, networking, CPU, data loading, or framework choices will never bottleneck
Swipe left to see more
Cross-host distributed workloads need their own test. Once a job spans multiple hosts, networking becomes a major factor.
How Hivenet measures inference workloads.
Inference performance depends on the model, precision, prompt size, output length, concurrency, serving engine, and latency target. Hivenet inference benchmarks separate latency floor, throughput ceiling, and production concurrency recommendations.
TTFT
Time to first token. Important for perceived responsiveness and interactive workloads.
TPOT
Time per output token. Useful for generation speed after the first token.
End-to-end latency
Total request time from submission to completed output.
Tokens per second
Throughput measure for comparing serving setups under defined load.
Error rate
Shows whether a setup remains stable as concurrency rises.
Performance only matters when the economics hold.
A faster setup is useful when it also fits the workload, budget, and operating model. Hivenet benchmark pages connect technical results to pricing basis where possible, so teams can compare performance and cost together.
Price basis
Benchmarks should state the instance, endpoint tier, or storage price used for the comparison.
Workload fit
A strong result for one model, batch size, or prompt shape does not automatically apply to another workload.
Runtime behavior
Short tests, long jobs, steady inference, bursty traffic, and batch workflows can produce different cost-performance outcomes.
Platform path
Compare GPU/CPU rental, Inference API, S3 storage, and Private AI based on how much of the stack your team wants to operate.
Enterprise-grade infrastructure needs measurable proof.
Hivenet runs AI, compute, and storage workloads on Policloud-backed infrastructure designed for reliable performance, cost visibility, and regional deployment. Benchmarks show how that infrastructure behaves under specific workloads.
Policloud-backed capacity
Modular infrastructure gives Hivenet a practical way to place capacity closer to energy, region, and workload demand.
Reliable workload paths
Hivenet connects infrastructure, software, APIs, billing, access patterns, and product workflows into a platform teams can test and operate.
Standard tools
Benchmarks use practical toolchains and metrics so teams can compare results against workflows they understand.
Clear limits
Each benchmark should explain where the result applies and where a workload needs its own test.
Your workload still needs its own test.
A benchmark can show whether Hivenet is a strong fit for a class of workload. It cannot replace testing your own model, data pipeline, latency target, and production constraints.
Training and fine-tuning
Test step time, GPU utilization, communication time, data loading behavior, precision, batch size, sequence length, and optimizer settings.
Inference
Test latency, throughput, concurrency behavior, cold starts, prompt shape, output length, and model quality on real traffic.
Storage-heavy workloads
Test file size, access pattern, egress, object count, listing behavior, throughput, and integration with the tools your team uses.
APIs
Test average latency, tail latency, error rates, rate-limit behavior, request IDs, and performance changes across releases.
Choose the next step.
Compute with Hivenet
Explore cloud compute paths for GPU and CPU workloads, programmable infrastructure, and team-ready workflows.
GPU/CPU rental
Rent RTX 4090, RTX 5090, or vCPU instances for AI, ML, rendering, notebooks, and development workloads.
Hivenet Inference API
Use OpenAI-compatible managed endpoints for foundational models.
AI workloads
Route production inference, RAG, fine-tuning, model hosting, and Private AI workloads to the right Hivenet path.
S3-compatible storage
Store datasets, backups, media, archives, and AI pipeline files with free egress and familiar tools.
FAQ
Common benchmark questions
Bring your workload. Test it properly.
Share the model, data path, target latency, throughput needs, region requirements, and current infrastructure. We'll help you decide what to benchmark and which Hivenet path fits.
