P4801X, Optane, and TDP

[IntelUser2000]

The above is the leak for P4801X, the M.2 version of P4800X Optane drive. This slide gives us lots of info to work on. It seems to me writes are directly responsible for the difference in power consumption. Dare I say its 2W per 150K IOPS? If we use this rudimentary metric, the controller's portion of power usage is a little over 8W. We can also see that the performance greatly varies based on capacity.

A potential positive from the information-
Based on the fact that the 800P can do 145K IOPS at 3.75W, its plausible the controller in the 90x series of Optane SSDs take up overwhelming majority of the power used. This itself is good for the 3D XPoint media.

Negative-
However, there's also the negative. The controller needs lots of power at high performance. This explains why the sequential performance of Optane drives are lacking. Since the 800P achieved more than 2x the write performance(over Optane Memory) while barely increasing power used, something dramatic changed in a year. Similar improvements are attributable to power reduction for the P4801X drive.

Despite being rated the same for performance, the 960GB version of the 905P has higher TDP and idle ratings compared to the 480GB version. The reviews out there all use 960GB, never the 480GB. The 960GB 905P not only has higher TDP compared to the 480GB 900P, it likely uses a more efficient controller. So Intel extracted extra performance using higher power + better controller. I don't think Intel wants to draw attention to the fact that Optane may be power limited. The 480GB version of the 905P may not perform better than the 900P due to lower power.

If you look at the long-term graph of the 905P from the Tomshardware review, it seems to be throttling using a technique similar to PWM. The 900P throttles too, but with a lower performance level.

Micron's POV?
This might also explain why Micron didn't come out with the QuantX drive. Their QuantX series were aimed at far higher performance levels saturating the PCIe 3.0 x8 bus and achieving nearly 2 million IOPS(or nearly 8GB/s) on mixed random read/writes! The U.2 version with presumably the x4 bus was reaching 900K IOPS in the same metric. The mighty P4800X looks like a mice in comparison - it only gets 500K IOPS. The early QuantX models had a fan and looked like a graphics card of the Riva TNT2 generation. At 1.8 million IOPS, it might have used upwards of 35W!

Further analysis

This might have been the reason why we didn't see a much higher performing 800P and M10 drive. Higher power would have prevented usage in laptops.

I also find 900P and P4800X comparison interesting. The P4800X has 4W higher power, and 10% higher random and sequential write performance. Clearly Intel sees lower power as important for client. Aside from unknown firmware optimizations(Client for 900P, Server for P4800X), the 900P can get away with lower TDP because client usage is bursty, even in the very demanding usage scenarios.

It'll be interesting in the coming days to see how much more efficient the controller can get. Also I wonder about the DIMMs. Potentially, the DIMMs are something entirely different. Not having to use block addressing modes and just use byte addressing may improve power efficiency of the controller. 18W is the number thrown around in some early leaks. 4W allows for a potential 800K random write IOPS, and that's probably fine if they can get latency really low as promised.