Benchmarks show a ~2.7× advantage over a current-generation 96-core AMD EPYC 9655P — on CPU-only complex analytical SQL.
NVIDIA’s next-generation Vera CPUs are slated to reach full production soon. As an NVIDIA Preferred Partner, Kinetica was granted early access to put it through its paces.
We decided to run the open-source CoffeeBench suite and compared it head-to-head with a current-generation 96-core AMD EPYC 9655P.
The result: Vera won every single query. Every run. Every time.
For customers, this means complex analytical queries return fast enough to become part of interactive applications, operational decision-making, and AI-agent workflows — without having to take a coffee break in between queries.
This test used Kinetica’s CPU-native query engine with no GPU acceleration. The same Kinetica engine and workload ran on both systems.
1. Methodology
CoffeeBench is an open-source, one-click benchmark designed around the types of questions a retail analytics team might actually ask — rolling trends, rankings, margin calculations, and quarter-over-quarter comparisons — against a ~720M-row sales dataset, joined across time-varying product pricing and store location dimensions. Each query runs three times in sequence: cold, warm, and fully warmed.
Some of the queries include non-trivial asks like:
- Rolling 7-, 30-, and 90-day sales averages
- Month-over-month product rankings
- Discount trends by season and geography
- Quarter-over-quarter profit and margin decomposition with year-over-year comparisons
2. Results
Vera maintained roughly the same 2.6–2.7× advantage from cold through fully warmed execution.
Vera beat EPYC on all 51 individual query/run measurements — not just the average.
3. Standout Queries
A few results from the run are worth calling out on their own:
Query 16 — The Biggest Time Savings in the Suite
This query ranks the top-3 product categories by sales for every state and season, built on a three-way join across the sales, product, and location tables. It’s the heaviest query in the benchmark on every platform, and where Vera’s advantage shows up in real time, not just ratios: 18.1s total on Vera versus 48.8s on the EPYC 9655P — over 30 seconds saved on a single query. Even on the fully-warmed run, after every cache is hot, Vera still finishes 2.7× faster than EPYC.
Query 10 — Where Warming Up Helps Vera Most
A rolling 90-day distinct-order count per city, using an unbounded window function — normally one of the more punishing patterns for a query engine. Vera’s cold-start advantage over EPYC is already solid at 2.46×, but it actually widens to 2.86× once the cache is fully warmed. EPYC’s caching gains flatten out here; Vera’s keeps compounding.
Query 1 — The Sharpest Single-Run Gap
A rolling 7-day sales average by city — about as ordinary a retail analytics query as they come. On the very first, cold execution, Vera finished in 1.13s. EPYC took 4.35s. That’s a 3.86× gap, the largest ratio anywhere in the dataset, on a query type most teams run constantly.
4. What Does 2.7× Faster Actually Mean?
Benchmark ratios are useful, but seconds are sometimes more meaningful. That difference matters when analytics moves beyond traditional BI. A human analyst may tolerate waiting 30 or 60 seconds for a complex report. An application making decisions in real time — or an AI agent issuing multiple analytical queries as part of a reasoning loop — has a much tighter latency budget.
| More interactive analytics | Complex analysis becomes something users can explore iteratively rather than submit and wait for. |
| More responsive AI | Agents can retrieve and analyze current operational data repeatedly without database latency becoming the bottleneck. |
| Higher workload density | Faster queries free compute resources sooner, creating the potential to serve more analytical work on the same infrastructure. |
| Faster operational decisions | Streaming and historical data can be analyzed together quickly enough to influence what happens next. |
The goal isn’t simply to make yesterday’s dashboard load faster. It’s to make complex analytics fast enough to sit inside the operational decision loop.
5. Why Vera Is Interesting for Kinetica
One reason Vera is particularly interesting for Kinetica is memory bandwidth.
Vera combines 88 custom NVIDIA Olympus cores and 176 threads with up to 1.5 TB of LPDDR5X memory and up to 1.2 TB/s of memory bandwidth.
That maps naturally to how Kinetica executes analytical workloads. Kinetica is a memory-first, columnar engine designed to vectorize analytical operations across CPU cores while minimizing unnecessary data movement through data locality, columnar access, and chunk skipping.
CoffeeBench puts those characteristics to work. Rather than testing simple scans or aggregates, its 17 queries use multi-way joins and window functions to calculate rolling sales averages, product rankings, discount trends, and profit and margin analysis.
Higher memory bandwidth gives a data-intensive engine like Kinetica more bandwidth to keep CPU cores fed as they process large amounts of data in parallel. The benchmark does not isolate memory bandwidth as the sole cause of Vera’s advantage, but the results show how well Kinetica can take advantage of the capabilities offered by this new CPU architecture.
Run CoffeeBench yourself: github.com/kineticadb/coffeeshop
Want to see how Kinetica performs on your own complex SQL workloads? Try it free on Kinetica Cloud or download the Developer Edition.
Benchmark environment: single-node Kinetica CPU-native database engine, 17-query analytical workload over a joined fact/dimension retail dataset, three sequential runs per query (cold, warm, fully warmed). Vera figures reflect early-access, pre-launch hardware.