Get in touch

Send us a message — we reply within 24 hours.

ROLV Primitive© · 3 Patents Pending · rolv.ai

How ROLV Primitive© Works

The core insight

In modern AI inference, large portions of each weight matrix are mathematically zero for any given input — not approximately zero, exactly zero. Standard libraries compute every row regardless. ROLV Primitive© identifies the non-zero structure at load time and restricts all computation to live elements only. The result is bit-identical to full computation. No approximation. No accuracy trade-off.

Three-phase operation

Phase 1 — Analysis (once per model load): The weight matrix is analysed to determine its non-zero structure. This analysis is deterministic and hardware-independent. Build time is measured in milliseconds and amortised across thousands of inference calls.

Phase 2 — Compute (every inference call): Only the identified live elements participate in computation. The operation uses the same underlying BLAS/tensor-core primitive as the standard dense library — but on a matrix proportional to the non-zero fraction only. At 90% sparsity, the active matrix is 10% the size. The reduction in compute is exact and proportional.

Phase 3 — Assembly: Results are placed into the correct positions in the full output tensor. The final output is identical to running the full dense or sparse operation. This is verified with SHA-256 hashing and perturbation testing on every benchmark case.

Why it beats cuSPARSE

Vendor sparse libraries use index-based formats (CSR, COO) designed for scientific matrices with randomly scattered non-zeros. AI weight matrices have a different structure — sparsity is organised by rows or blocks. ROLV exploits this structure to perform a single contiguous dense operation rather than indexed scatter-gather, achieving better cache utilisation and full tensor core throughput.

Correctness guarantee

Every result is verified with: (1) SHA-256 hash of input matrix A, (2) SHA-256 hash of input vector V, (3) SHA-256 hash of reference output, (4) SHA-256 hash of ROLV output. A perturbation test confirms that modifying one weight changes the ROLV output hash — proving real computation on real data. ATOL=0.05 on column-normalised fp64 outputs.

Real model weights · SHA-256 verified · Perturbation PASS every case

Zero pruning applied. Router zeros are architecturally exact. cuBLAS = production baseline (what inference frameworks call). cuSPARSE at ≥70% sparsity. ATOL=0.05 col-normalised fp64. 4 SHA-256 hashes + perturbation test per case.

★ = peak REAL-weight result. ★★ = peak overall (synthetic routing). † = cuSPARSE active submatrix (full matrix exceeds INT_MAX; ROLV handles full matrix natively and still wins). At 90%+ sparsity, production inference uses cuSPARSE not cuBLAS — the vs cuSPARSE column is the primary real-world comparison. At 96.9%, ROLVswitch™ auto-selects cuSPARSE as baseline; ROLV still wins 25–53×. B200: Mixtral-8×7B, Qwen3-30B-A3B, Llama-4-Scout. H200 NVL: Qwen2-57B-A14B, Gemma4-26B-A4B, OLMoE, DeepSeek-V2-Lite, Phi-3.5-MoE, Qwen1.5-MoE-A2.7B, DeepSeek-MoE-16B, Jamba-1.5-Mini. H200 (REAL): Kimi-K2-Instruct, DeepSeek-V3-0324, DeepSeek-V3, DeepSeek-R1, Mixtral-8×22B. H100 (REAL): MiniMax-M2.5 (custom MiniMaxM2 arch — cuSPARSE full matrix viable, ROLV wins 25×). H200 (synthetic routing, real architecture dims): Snowflake-Arctic, Kimi-K2.5, Llama-4-Maverick, Qwen3-235B-A22B, DBRX. Gemma4 top_k used=6 vs config top_k=8; sp 95.3% vs true 93.8%. iter = per-iteration speedup. total = 1000 iters including ROLV build. Energy via pynvml. GPU: batch=2048, 1000 iters, BF16, TF32 ON, 17 MoE models confirmed real weights. CPU: 357/357 PASS. 4 SHA-256 hashes + perturbation PASS every case.

GPU — Pruned / Weight Sparsity

1,684/1,684 total PASS across all GPU benchmarks · AMD MI300X synthetic benchmark: rocBLAS 8.5× (rocSPARSE has known regression at high sparsity — disclosed) available in full benchmark PDF · Full CSV on request

Any hardware. Any OS. Any framework.

ROLV Primitive© operates at the matmul level — the single universal primitive shared by all compute hardware. If your platform supports matrix multiplication, ROLV works on it today. No drivers to update. No firmware. No hardware changes.

The Story of Rolv Eitrem Heggenhougen

A Regular Guy With Ideas That Refuse to Stay Ordinary — and Absolutely No Patience for Slow People

Some people spend their lives following the rules. Rolv Eitrem Heggenhougen spent his life looking past them, around them, under them, and occasionally straight through them at full speed.

Born in Norway and shaped by the American belief in possibility, Rolv built his career in business — founding companies, expanding across continents, and turning unconventional ideas into reality. He was never the stereotypical engineer or academic. His father had been the mathematical prodigy; Rolv was the one who preferred building things, selling things, imagining things. But imagination, as he would later say, is the only limitation to innovation.

And Rolv’s imagination never rested. His patience definitely didn’t.

Early Life: The Making of a Relentless Engine

Rolv’s story didn’t begin in boardrooms or data centers. It began with a paper route in Norway. It continued with hard-labor summer jobs during college — the kind of work that teaches you to push through discomfort, to keep going, to rely on yourself.

He also discovered something important: he was stubborn. Very stubborn. And impatient. Very impatient. If something could be done today, he wanted it done yesterday. If he had an idea, stopping was not an option. This would later become both his superpower and the reason people around him occasionally needed aspirin.

A Quiet Talent for Math — Even If He Pretended Not to Care

After majoring in math and physics in high school in Norway, Rolv came to the United States for college. On arrival, he challenged every math and algebra class — and passed them all by simply taking the placement tests.

He didn’t brag about it. He didn’t even think it was a big deal. But it was a hint: the raw horsepower was always there, idling quietly, waiting for the right idea to hit the accelerator.

Entrepreneurship Across Borders

Rolv’s business career began with European Trading Corporation (ETC) — a computer component trading company he founded and later took public on the Oslo Exchange in 2000. By then, ETC had expanded across Europe and the United States and was even building its own ETC PC line.

Then came PC500 in Norway, Sweden, and Germany — same concept, but selling directly to end users. Fast, scrappy, and very Rolv.

In 2010, WrapMail — his patented email-based marketing platform — went public in the United States. It transformed everyday corporate emails into branded, clickable marketing assets long before “growth hacking” was a thing.

These ventures weren’t accidents. They were the natural output of a mind that refused to sit still.

A Lifetime of Unconventional Ideas

Rolv’s imagination has always run ahead of the world around him. Among his many concepts:

• Giving away new cars wrapped in advertising, turning vehicles into mobile billboards — with targeted ads broadcast through the car’s multimedia system.
• An airplane safety seat that transforms into a protective ball, capable of being dropped from 20,000 feet with the passenger surviving inside.
• Next Minute, a patented real-time marketplace connecting local service providers with nearby customers.
• The ROLV Unit™, a new conceptual framework for measuring computational efficiency.

These ideas weren’t random. They were expressions of a mind that instinctively challenges assumptions and sees possibilities others overlook.

The Bike Ride That Changed Everything

In May 2025, during a bike ride in Fort Lauderdale, another idea arrived — one that would pull him into one of the most technical arenas in modern computing.

He wondered whether matrix operations — the mathematical backbone of AI — could be made dramatically faster. He didn’t know sparse computing. He didn’t know GPU kernels or the pitfalls of zero FLOPs. But he knew how to follow an idea.

He filed a provisional patent. Then continuation-in-part patents. He taught himself Python, Jupyter, CUDA, ROCm, TPU stacks — the entire modern AI compute ecosystem. He learned sparse linear algebra from scratch. He built prototypes. He kept going.

What emerged became ROLV Primitive© — a new approach to sparse computation that was hardware-agnostic, deterministic, and astonishingly fast.

And yes — doing all of that in six months was extraordinary. Ridiculous, even. But very on-brand for Rolv.

The Unlikely Inventor

What makes the story extraordinary is not just the technology — it’s the inventor.

Rolv was not a career programmer. He was not a numerical methods researcher. He was not a GPU architect. He was a business founder, a problem-solver, a helicopter pilot, a marathon runner, a multilingual Norwegian-American who followed ideas wherever they led.

He describes himself as a regular guy with a lot of ideas. But the pattern is unmistakable: He sees possibilities others overlook. He imagines solutions others never consider. He acts on ideas others dismiss. And he moves fast — sometimes too fast for the people around him — but always forward.

ROLV Primitive© · 3 Patents Pending · rolv.ai · ROLV Primitive© · 3 Patents Pending · rolv.ai · [email protected]