6,832,000 IOPS of 4K random reads at 28 GB/s, out of one slot. That is the line I look at first in StorageReview's review of the Apex X21, "The Roaring Twenty-Ones", published May 1, 2023. I run our benchmarking lab, so reading someone else's numbers for our card is a strange exercise. I want them to be high, and I also want to know whether they make sense. These ones do, and three of their observations match things we measured ourselves this year on different hardware.

Everything below that comes from the review is StorageReview's data, on their platform, with their drives. I have not rerun their tests. Where I bring in our own numbers, I say which cards and which hosts they came from.

Their setup

StorageReview built the test box around an AMD Ryzen 9 7950X, a 16-core desktop CPU, on an MSI X670E board with 64 GB of DDR5 and integrated graphics. The OS was Debian. The card carried 21 Samsung 980 Pro 1 TB drives, which are consumer Gen4 M.2 drives.

They ran two tools. Vdbench gave them workloads they can compare against the enterprise devices they have reviewed. fio with io_uring gave them the peak I/O and bandwidth the card could pass. Before each fio test they formatted every SSD and then filled it completely. I like that detail. Unwritten flash does not read like written flash, and in our lab it read faster on one drive model and slower on another, so a benchmark on fresh drives can mislead in either direction. A filled drive is the honest starting point.

What they measured

The fio results, all on filled drives:

Workload IOPS Bandwidth
4K random read 6.832M 28 GB/s
4K random write 5.274M 21.6 GB/s
8K random read 3.454M 28.3 GB/s
8K random write 3.499M 28.7 GB/s
512-byte random read 14.6M 7.477 GB/s

They also report that the same 512-byte read test reached 23M IOPS at 11.8 GB/s before the drives were filled, and they kept the filled figure as the result. That is the right call.

Vdbench peaks came in at 27.2 GB/s for 64K sequential read and 23.3 GB/s for 64K sequential write. Their SQL workload peaked at 3.46M IOPS, or 27.1 GB/s. VDI boot reached 1.99M IOPS at 27.5 GB/s.

Their verdict was that "The X21 offers unparalleled performance to the point that this one component could be a productivity game-changer" for a certain set of professionals, and that buyers who build a suitable platform "will not be disappointed with what this card brings to the table." They named data scientists and creative professionals as the people the card fits.

If those numbers fit your build, request a multi-unit quote. The X21 Gen4 and X16 Gen4 are in stock and ready to ship.

Observation one: the 4K reads hit the link

Multiply it out. 6.832M reads of 4,096 bytes each is 27.98 GB/s, which matches the 28 GB/s StorageReview printed. A PCIe Gen4 x16 link carries about 31.5 GB/s per direction before protocol overhead, so their 4K reads used about 89 percent of the raw link rate. The 8K reads landed at 28.3 GB/s, the 8K writes at 28.7 GB/s, and VDI boot at 27.5 GB/s. Four different workloads stop within 1.2 GB/s of each other. When unrelated workloads converge on one bandwidth figure, the shared pipe is the limit. StorageReview said so themselves, writing that the card's main limit is the Gen4 x16 slot cap at around 28 GB/s.

We saw the same pattern in 2026 on Gen5. That campaign used two Apex X16 Gen5 cards and 32 consumer Gen5 drives, not the X21, and we have not run that campaign on the Gen4 cards. On the Threadripper workstation, with polled io_uring and spec_rstack_overflow=off, 4K random reads stopped at 25.6M IOPS at 24 threads and 25.9M at 32 and 48 threads. That is about 12.9M per Gen5 x16 uplink, or 53 GB/s per card at 4 KiB per read, against about 63 GB/s of raw link rate per Gen5 x16. The 32 drives, each measured on its own, sum to about 49M 4K IOPS, so the drives had headroom the link could not pass.

The cause is the same in both cases. At 4K, a single x16 uplink runs out of bandwidth before a full card of drives runs out of IOPS. Past that point, more 4K IOPS comes from more cards in more slots.

Observation two: smaller blocks, more IOPS

StorageReview's 512-byte read result of 14.6M IOPS is about 2.1 times their 4K result, and it moved only 7.477 GB/s across the link, about a quarter of what the 4K test moved. Their pre-fill run hit 23M IOPS at 11.8 GB/s. Each 512-byte read carries one eighth of the data of a 4K read, and StorageReview wrote that at the smaller block size they started to hit the limits of the 980 Pro drives.

Our Gen5 numbers have the same shape. The setup that stopped at about 26M 4K IOPS reached about 47M IOPS at 512 bytes, roughly 1.8 times the 4K figure, and stayed there from 24 to 48 threads. That is only about 12 GB/s of data per Gen5 x16 port. Our analysis treats that plateau as a limit on how many transfers each port completes, about 23.5M per second, rather than on how many bytes it carries. In both labs the 4K result measures the uplink's bandwidth, and the 512-byte result measures something else. In their test that was the drives, and in ours it was the per-port transfer rate. If you see one number quoted without its block size, ask for it.

Observation three: the slot and the CPU matter as much as the card

StorageReview noted that an Intel-based consumer system connected to the card at x8 instead of x16, and its BIOS offered no setting to force the width. They also wrote that the CPU in that system began to struggle in the multi-million IOPS tests.

We found both effects. An x16 card in a slot wired with fewer lanes trains at that narrower width, and at x8 that halves the uplink. On the Threadripper workstation board we test with, a Gen4 x16 card trained at x8 in the board's Gen5 x8 slot and at x4 in its Gen4 x4 slot. You can check what your card negotiated with sudo lspci -vv -s <bus:dev.fn> | grep -E "LnkCap|LnkSta" and compare the speed and width on the two lines.

The host CPU sets a ceiling too. With the same two X16 Gen5 cards and the same drives, the stock interrupt-driven Linux NVMe driver delivered 13.3M 4K random read IOPS on the Threadripper workstation and 6.5M on the Xeon workstation. All our Threadripper figures were taken with spec_rstack_overflow=off, which turns off an AMD speculative-execution mitigation and is a security trade-off. On the Xeon, the 6.5M figure stayed flat from 8 drives to 32, because the host's interrupt completion path ran out before the drives did. That 6.5M is close to the 6.8M a Gen4 x16 link carried in StorageReview's test, which is why a Gen4 card is a sensible match for hosts with a small-block ceiling like that.

What this means for a build

StorageReview's results come from one platform with one drive model. Change the drives, the CPU, the slot or the kernel settings and the numbers change with them. Their numbers line up with what a Gen4 x16 link can carry, and the limits they ran into are the same slot and CPU limits we measure.

If your host has Gen4 slots, a Gen5 card links at Gen4 there and gains no link bandwidth over a Gen4 card. The X21 Gen4 puts 21 M.2 drives behind one x16 link, and the X16 Gen4 puts 16 there. Both use a PCIe switch, so the slot needs no bifurcation support, and the drives show up as ordinary NVMe devices. Both cards are in stock and ready to ship.

Request Multi-Unit Quote · Apex X21 Gen4 product page

Related:

  • Which workstation and which slot
  • Where the IOPS go
  • Linux setup checklist