Research & evidence / PRISM benchmarks

Performance,
with the evidence.

Recorded portfolio optimization results. Explore the workload, inspect the timings, and read the numerical quality contract behind each comparison.

Explore the results Download data

Generated, seeded common factor-QP workloads. Construction + solve after GPU warm-up. These observations do not measure client workflow latency or realized investment returns.

Study 01 / Generated common factor-QP

Explore the recorded timings.

Choose a workflow and asset count. Every value comes from the same 24-observation record.

Assets per portfolio
Link to this view

Direct indexing · 100K assets · 25.50× speedup

Selected observation

Direct indexing / 100K assets

PRISM GPU0.327seconds
CPU reference8.329seconds
Measured speedup25.50×
Numerical qualityPassedObjective + feasibility gates

Direct indexing

Construction + solve · seconds

Select an asset count

PRISM GPUCPU reference
PRISM GPU: 0.074 sCPU reference: 1.595 s
PRISM GPU: 0.199 sCPU reference: 3.980 s
PRISM GPU: 0.327 sCPU reference: 8.329 s

One linear time scale across all three asset counts. Shorter bars mean less elapsed time.

Across the eight workflows

Speedup at 100K assets · select to inspect

0–70×
Quality contract

Objective gap ≤ 10−3 · Feasibility error ≤ 2 × 10−6 · Native convergence

Methodology

What is timed

Instance construction and solve after one excluded GPU kernel warm-up.

What is compared

PRISM GPU and a tuned eight-thread frontier commercial CPU reference.

What it establishes

Numerical quality for generated common-QP formulations. Full tax-lot, SOC and executed-ledger parity need separate acceptance tests.

Full measurement protocol

Common factor-QP on generated, seeded portfolio data. Eight workflow profiles were confirmed sequentially in isolation at 20K, 50K and 100K assets. All 24 cases passed: objective gap ≤ 1e−3, feasibility error ≤ 2e−6 and native convergence.

Times include instance construction and solve after one excluded GPU kernel warm-up, compared with a tuned eight-thread frontier commercial CPU reference. Each isolated cell is a recorded observation; it is not client workflow latency or a distribution of repeated measurements. The concurrent qualification campaign is excluded from these latency charts.

These profile names describe common-QP formulations. They do not establish full tax-lot, SOC or executed-ledger equivalence, realized tax savings, or universal speedups for every portfolio model.

Source: quality-capped recomputation v1, isolated large-scale confirmation. Download the plotted observations (CSV)

Observation register

All 24 recorded observations

Select a workflow to inspect its recorded observation.

Download full CSV

Scroll the table horizontally to inspect all columns.

Construction + solve time in seconds. Speedup = CPU reference time ÷ PRISM GPU time. All 24 recorded observations are available in the CSV.
WorkflowAssetsPRISM GPUsecondsCPU referencesecondsSpeedupQuality gate
20,0000.0744721.59530021.42×Passed
20,0000.0614031.56350025.46×Passed
20,0000.0503911.58395531.43×Passed
20,0000.0546701.53178728.02×Passed
20,0000.0249651.51106860.53×Passed
20,0000.0626631.60664225.64×Passed
20,0000.0737391.56453121.22×Passed
20,0000.0529971.45117227.38×Passed
50,0000.1987433.97978920.02×Passed
50,0000.1456563.97849527.31×Passed
50,0000.1387424.20322130.30×Passed
50,0000.1486093.74336225.19×Passed
50,0000.0836183.78739845.29×Passed
50,0000.1505874.15917427.62×Passed
50,0000.2021404.51575922.34×Passed
50,0000.1505914.60404830.57×Passed
100,0000.3266338.32940825.50×Passed
100,0000.3135208.36161126.67×Passed
100,0000.3289479.63102529.28×Passed
100,0000.3091088.10149026.21×Passed
100,0000.1657138.01709648.38×Passed
100,0000.3233938.48203826.23×Passed
100,0000.4383319.26199221.13×Passed
100,0000.3129918.22314426.27×Passed

Study 02 / Batched factor-QP

From one account
to a book of accounts.

All five measured workloads. This study uses a four-worker CPU QP reference, a different comparison from the single-account study.

1 account

500 assets per account

PRISM GPU 7.006 msCPU reference 5.808 ms
0.83×Reference faster

32 accounts

500 assets per account

PRISM GPU 7.255 msCPU reference 37.787 ms
5.21×Speedup

128 accounts

500 assets per account

PRISM GPU 7.419 msCPU reference 141.303 ms
19.05×Speedup

512 accounts

500 assets per account

PRISM GPU 56.689 msCPU reference 558.420 ms
9.85×Speedup

128 accounts

2,000 assets per account

PRISM GPU 103.416 msCPU reference 1,847.000 ms
17.86×Speedup

The CPU reference is faster for the single 500-asset account. GPU speedups apply to the stated batches; they are not a universal latency claim.

Batch protocol and exact values

Generated batched factor-QP. Five workloads, two seeds and three repeats in randomized confirmation. All measured cells passed their quality gate. The comparison here is PRISM GPU versus a four-worker CPU QP reference; it is a different reference lane from the frontier commercial solver in the single-account study above.

The register shows all five workloads, with the full timing record below. At one 500-asset account PRISM measured 7.006 ms versus 5.808 ms for the reference (0.83×); the GPU speedup is specific to the stated batched workloads.

Reported batch elapsed time, milliseconds.
Accounts × assetsPRISM GPU (ms)4-worker CPU (ms)Speedup
1 × 5007.0065.8080.83×
32 × 5007.25537.7875.21×
128 × 5007.419141.30319.05×
512 × 50056.689558.4209.85×
128 × 2,000103.4161847.00017.86×

Interpretation / Acceptance criteria

Define the decision.
Then define the test.

Numerical agreement and executed economic parity answer different questions. Choose the acceptance contract for the decision your platform needs.

Numerical quality

Objective & feasibility

The charts above report observations that meet the stated numerical gates for their model and reference lane.

Executed-ledger parity

Economic agreement

Full tax-lot, SOC, rounding and ledger outcomes require their own frozen acceptance gate. A numerical pass alone does not establish interchangeable trades.

Executed-ledger parity and historical results

The prior strict full tax-lot/SOC/ledger configuration established executed-objective parity at 1K assets and lots: 1.480 s vs 3.587 s (2.424×), with a $0.001496 executed difference inside the frozen $0.01 gate. At 5K, $0.018935 exceeded that gate. Capacity at 500K does not establish strict-parity speed superiority.

The newer quality-capped full-D0 results measure numerical quality under a different acceptance contract. They do not establish interchangeable rounded trades or economic parity and are excluded from the marketing charts on this page.

Separate historical real-data studies, hosted service measurements and policy simulations have their own model and timing scopes. They are not merged into the generated common-QP results. Explore the evidence library

Public data / Generated common-QP study

Inspect the
recorded results.

The downloadable table contains 24 observations: eight generated common factor-QP workflows at 20,000, 50,000, and 100,000 assets. Each row records PRISM GPU time, commercial CPU reference time, speedup, and whether the numerical quality gates passed.

All 24 observations passed an objective-gap gate of ≤ 10−3 and a feasibility-error gate of ≤ 2 × 10−6. Median speedups at the three asset counts were 26.51×, 27.47×, and 26.25×. Timings cover construction and solve after an excluded GPU warm-up.

These are generated numerical workloads. They do not establish realized investment returns or full tax-lot, SOC, or ledger parity. The full table and measurement protocol above provide the context needed to interpret each observation.

Download the 24-row results (CSV) Find the original research

Your book / Your acceptance gate

A benchmark built
around your workflow.

An 8-week paid matched-workload pilot, with the model, baseline and acceptance gate agreed before kickoff.

Benchmark PRISM on your book