Kioxia introduced the GP1 in early August as the first SSD to clear ten million IOPS, and evaluation samples are due before the end of the year. The drive meant to reach a hundred million was scheduled for 2027 and has now moved to 2028, held there by two things that have to land first.
A hundred million IOPS waits on PCIe 7.0
The GP1 reaches 10.3 million IOPS at a 512-byte block size using second-generation XL-Flash over PCIe 6.0, in 800GB models built from SLC XL-Flash at one bit per cell and 1,600GB models using the two-bit MLC variant, in E1.S and E3.S form factors, according to ComputerBase's account of Kioxia's roadmap. The tenfold successor needs third-generation XL-Flash and a move to PCIe 7.0, and 2028 sits better against the PCI Express roadmap than 2027 did, which is most of the explanation for the slip.
Kioxia's own figures for XL-Flash Gen 3 are a tripling of read performance over Gen 2 and a 150 percent gain in write performance, with power efficiency up 40 percent reading and 80 percent writing. Those come off a slide with no stated baseline configuration, so treat them as projections rather than measurements, and the same applies to the flash-level numbers Kioxia expects from Gen 3: read latency below 5 microseconds, and endurance of 150,000 to 250,000 write cycles, both far outside what mainstream TLC NAND does. How it gets there has not been said. XL-Flash has so far bought its advantage over conventional NAND with shortened word lines and bit lines and many more planes per die, which is a different axis from the vertical layer count consumer NAND competes on.
None of this is desktop hardware. These are enterprise drives in enterprise form factors, developed as Kioxia's contribution to NVIDIA's Storage Next work, where the drive exchanges data with a GPU directly and is meant to complement HBM rather than replace it, and liquid cooling is already part of the conversation for the GP1 and other PCIe 6.0 SSDs. The 512-byte block size behind the IOPS figure is also worth holding onto: it is the number that makes a small-transfer workload look its best, and it is not the block size any desktop benchmark reports.
A 750W supply with a 450W ceiling on its Gen 5 connector
The specification label on the Phanteks AMP GH 750W lists 62A and 744W on the 12V rail, then a separate line stating 450W for the 12V-2×6 output, and the connector housing on the cable carries the same 450W marking, as Tom's Hardware found in its review. That is a hard limit on what the single native Gen 5 connector will carry, and it is printed where a builder will see it rather than buried in a datasheet.
The rest of the unit tests better than its sub-$90 street price suggests. The 12V rail moved 0.4 percent across the 20 to 100 percent load span, ripple peaked at 38 mV against the 120 mV ATX limit, efficiency averaged 91.0 percent at 230 VAC, and the fan stays stopped to nearly 300W at room ambient. It has Japanese capacitors and an eight-year warranty behind a sleeve-bearing Yate Loon fan, which is the part of the bill of materials that does not match the warranty. In a 45 degree hotbox the fanless window narrows to roughly 220W and the unit reaches 50.2 dB(A) at sustained full output, which is loud but was reached without a shutdown, outside its own 40 degree rating.
On a 750W supply the connector cap is mostly academic, because a card that wants more than 450W over one plug does not belong on a 750W unit anyway. It matters on the upgrade path, and it is the reason to read the label rather than the box: on an ATX 3.1 unit the connector rating and the supply rating are two different numbers.
Noctua will keep making mounts for 2005 coolers
Noctua plans to carry on producing socket mounting kits for coolers as far back as the original NH-U12 and NH-U9 from 2005, which can still be fitted to current AM5 and LGA1851 processors, chief product officer Jakob Dellinger told Club386 in an interview. He put a condition on it that is worth reading carefully: the support continues "as long as the key mechanical parameters of modern desktop CPUs don't shift in a way that prevents us from ensuring that legacy compatibility". Dellinger also accepts the obvious point that the company would sell more coolers without the kits, and describes the policy as a loyalty play in an audience that builds machines for other people.
A tower cooler has no wear-out mechanism apart from its fan, and fans are replaceable, so mount availability is the only thing that decides whether a heatsink survives a platform change. Anyone weighing a premium air cooler against a cheaper one can put that in the arithmetic, with the condition attached: the commitment holds only while socket geometry stays close to what it is now.