Nvidia's Vera CPU More Than Doubles Xeon Memory Bandwidth in StorageReview Tests but Trails EPYC on Encryption

StorageReview measured 2,073GB/s of STREAM Triad bandwidth on a dual-socket server built around Nvidia's Vera CPU, more than double the 914GB/s it recorded on a Xeon 6978P system. On OpenSSL AES-256-GCM across all threads, the same Vera machine reached 1,257GB/s against 3,965GB/s for AMD's EPYC 9965.

The review, by Kevin OBrien and published November 10, 2026, is the first independent look at the part since Nvidia announced it on June 1, 2026 as "now in full production". It is also a useful correction to the single number Nvidia led with.

What Vera is, and what Nvidia claimed

Each Vera socket carries 88 Olympus cores and 176 threads through what Nvidia calls Spatial Multithreading. Cache is 64KB of L1 instruction and 96KB of L1 data per core, 2MB of private L2 per core and 164MB of shared L3. Memory is LPDDR5X on eight SOCAMM2 modules per socket, rated at 9,600MT/s and 1.2TB/s per socket. TDP is configurable from 250W to 450W.

Nvidia's own claim is "1.8x faster task completion compared with x86 CPUs", a vendor-supplied figure with no baseline given in the release. The company also cited Phoronix as finding the fastest overall performance across agentic workloads such as code compilation, Python, Java and database processing. Customers named as exploring the part include NYSE, Anthropic, OpenAI and SpaceXAI; the release says Anthropic and others are evaluating, which is not the same as buying.

The number that matters here is not the headline speedup but the memory system. A server CPU that feeds agent workloads is judged by how quickly it moves data, because compilation, database scans and in-memory caches are bandwidth-bound long before they are core-bound.

What StorageReview measured

The lab ran Vera in a dual-socket system with 1.5TB of memory (16 modules of 96GB Micron SOCAMM2) against four Dell PowerEdge machines: an R7725 with EPYC 9965 (192 cores per socket) and EPYC 9175F (16 cores per socket), an R770AP with Xeon 6978P (120 cores), an R770 with Xeon 6787P (86 cores) and an R760 with Xeon Platinum 8592+ (64 cores).

1. Nvidia Vera (2 sockets)2073 GB/s
2. Xeon 6978P914 GB/s
3. EPYC 9175F651 GB/s
4. EPYC 9965634 GB/s
5. Xeon 8592+482 GB/s
6. Xeon 6787P479 GB/s

Source: StorageReview, Nvidia Vera CPU review, published 2026-11-10, accessed 2026-11-11

The bandwidth lead is real. It also has a seam: a single Vera socket delivered 1,055GB/s to its own memory but 677GB/s to the other socket's, and tinymembench latency with 2MB pages was 99.2ns local against 212.7ns remote. The EPYC 9175F in the same lab showed 114.3ns local and 199.9ns remote.

The application results are mixed, which is what a bandwidth-led design predicts.

Test	Vera	Best x86 result listed
OpenFOAM large mesh	1,854s	Xeon 6978P, 2,227s
HammerDB TPROC-C, PostgreSQL 18.6, one socket	3.31M NOPM	EPYC 9965, 2.62M
HammerDB TPROC-C, MySQL 26.7.0, one socket	2.03M NOPM	EPYC 9965, 2.51M
GROMACS benchPEP, SMT off	1.77 ns/day	EPYC 9965, 5.36
OpenSSL AES-256-GCM, all threads	1,257GB/s	EPYC 9965, 3,965GB/s

On the large OpenFOAM mesh, Vera finished in 1,854 seconds, 17 percent ahead of the Xeon 6978P and well ahead of the EPYC 9965's 2,678 seconds. On GROMACS benchPEP it managed 1.77 nanoseconds of simulation per day, about a third of the EPYC 9965. Single-thread AES-256-GCM came in at 6.2GB/s, against 19.3GB/s for the EPYC 9175F.

Where the lead disappears

The encryption result deserves attention because it is where Arm cores in this class have historically been weakest, and Vera has not closed it. A buyer terminating TLS or encrypting storage at line rate would see EPYC 9965 deliver roughly three times the throughput across all threads.

MySQL is the other soft spot. Vera was ahead on PostgreSQL 18.6 and behind on MySQL 26.7.0 in the same HammerDB workload, a gap of about 21 percent against the EPYC 9965. StorageReview's tables give no explanation, and the lab did not publish power draw alongside them, so performance per watt cannot be computed from the review.

Price is the largest missing variable. The review lists none, and Nvidia's release gives none.

Availability and what would change the read

Dell's PowerEdge R8822 has shipped since October, with an M9882 Vera CPU rack due in the first quarter of 2027, according to the review. HPE's ProLiant Compute DL394 Gen12 arrives this fall, and Supermicro lists a 256-CPU Vera rack as available to quote. Oracle Cloud Infrastructure has Vera instances generally available, with CoreWeave, Nebius and Lambda expected to follow.

The question for a buyer is whether the workload is bandwidth-bound. For in-memory analytics and large-mesh simulation the review favours Vera; for encryption-heavy or MySQL-heavy fleets it does not. Nvidia's NUMA penalty, 677GB/s remote against 1,055GB/s local, also means software that ignores memory placement will leave a third of the advantage unused.

What would change my read is a price. At the 450W top of the TDP range, two sockets are a 900W CPU budget before memory and accelerators, and an EPYC 9965 system that costs half as much would win on delivered work per dollar even where Vera wins on raw time. Neither Nvidia nor Dell has published a list price for the R8822 configuration StorageReview described.

Related: AMD's counter-move at the top of the EPYC range is covered in our look at the 256-core EPYC 9996, and Nvidia's other robotics-and-agents push is in the GR00T 2 launch.