Advice request: 9950X vs 285k

Hi. I’m looking to upgrade my current i7-3770 machine to something a bit more modern. I’m having a really hard time choosing between the Ryzen 9950X and the Core 9 Ultra 285k, so I was hoping maybe someone with more recent PC building experience could help me out.

Some background on my requirements:

  • I’m looking at these higher-tier CPUs despite probably 80% of my usage being pretty undemanding web browsing and word processing. However, a lot of the software I use regularly is compiled from source and I do devops-type stuff as a hobby, so I’m looking to improve compile times, run multiple virtual machines, run fuzzing campaigns and improve the viability of playing with computationally-intensive experiments.
  • Power (and, given the above, idle and near-idle power in particular) is a key concern for me. I live in a locale that gets pretty warm during the summer, and I’m a little scared of going back to something like the Phenom II X6 which just made my living space unbearably hot.
  • I don’t really care about video game performance, so that’s not a reason for me to rule out the 285k or for 3D VCache to be an appealing trade-off. Further, I don’t use Windows, so I’m not bothered by whatever performance issues the 285k has over there at the moment.
  • I’m also not bothered by Intel’s refusal to commit to LGA 1851. If it’s another twelve years before my next CPU upgrade, AM5 won’t provide a better upgrade path.

Here’s where I’m at trying to evaluate the two:

9950X pro 9950X con 285k pro 285k con
slight edge in geomean productivity benchmark scores, but a mixed bag
16 full-fat cores only 8 P-cores
2-SMT, 32 threads no SMT, 24 threads
only 16 cores 24 cores
full width AVX-512 – looks like fun to program with! no AVX-512, E-core AVX double-pumped, AVX seems to cause throttling?
AM5 supports ECC memory – probably good for the mdraid array no ECC support on consumer chipsets, no workstation boards currently available
AM5 seems very picky about memory configuration seems comparatively less picky – simpler to get right and maybe simplifies future memory upgrades?
XMP/EXPO makes idle power consumption worse?
More efficient – default power cap at 200W Less efficient – default power cap at 250W
Worse idle power Better idle power – at least 20W lower?
Cheaper, by about $200 AUD

This leaves me with several questions:

  • I don’t know what to make of the 16 vs 8 P-cores, 32 vs 24 threads, 16 vs 24 cores comparisons.
  • I like to leave my computer on overnight, in which case the 285k wouldn’t just be idle but should be in an S0ix state. Would this meaningfully improve the idle power edge?
  • Is 20-30W lower idle power actually a meaningful difference in terms of heat output?
  • Would it make sense to have a more conservative power limit for summer, or would the performance impact be severe enough to warrant looking at another processor?
  • How do these chips compare power-wise when not idle or pegged, but when doing lighter, bursty tasks like web browsing?
  • Are there any other power gotchas I should be aware of? For instance, I was surprised to learn that apparently Samsung NVMe drives have issues with ASPM.

I know this is a long post, but if you got this far, thanks! I look forward to hearing about your experiences or any advice you have to offer.

To me it appears you have answered your own dilemma. The 9950X has lower power consumption, more threads, and is capable of high single-threaded boosts for office work/general web browsing. However, I’ll address each of your questions directly.

  1. The 9950X has a homogenous core configuration, which would make setting up VMs a lot more straightforward (I’m no expert, but that’s my hunch) than the 285K’s hetergenous makeup. You certainly wouldn’t have the possibility of running into issues with scheduling VMs on different cores and core types if you went with the 9950X. You also have more logical cores (i.e. threads) with the 9950X, and are more powerful than the E-cores on the 285K.
  2. I’m not an expert on sleep states, but I would wager that it doesn’t matter which processor has the better idle temperature. Choosing one processor over the other because it has 20 W different idle power consumption seems a little silly considering the benefits of actually using the processor.
  3. No. If you really cared about heat output, pick a lower-performing processor with significantly less power consumption to begin with.
  4. There a lot of tools you can use to change the power consumption of a processor; first and foremost: the BIOS settings. You can look up the differences between how AMD and Intel handle power limiting, but if heat is really a huge concern you could disable the maximum power consumption in favor of a lower setting. You won’t be losing that much performance by doing so, as trying to eek out 100% of the performance from a processor compared to using 95% of its performance increases the power budget by a near-exponential amount.
  5. I would recommend looking at Gamers Nexus’ video regarding each CPU: here is 285K benchmarks and here is the 9950X benchmarks. You can get a good idea of the power consumption of each of these CPUs from these videos.
  6. I have heard that this issue has been mostly resolved, but I’m not entirely sure. You can watch this video by JayZTwoCents, which he did a follow up recently but I can’t seem to find it.

Hope this helps resolve your dilemma.

2 Likes

Intel been pushing on this P-E core combo, through recent processor gens
Earlier premise was E-Cores being lower base+boost clocks / single threaded / MUCH lower power draw
[making E-Cores plotted for backend tasks – OS Scheduler permitting]
But now with them P-Cores, also not bothering with hyperthread support
P-Cores newest supposed advantage, being the applied individual L2 Cache [per core]

AMD maintaining the common core layout, will allow easier/fairer resource divvying

If you’re concerned for power draw(s), do assess the BIOSes [ea. mainboard you’re interested in]
They should have pliability, whether some preset toggles or full-on custom adjuster page
Thermals would also be pending, on your case/cooler plan(s)

TL;DR, absent decision points which swing clearly one way or the other it doesn’t really matter which you pick.

Effects of core to core differences are pretty workload dependent. Main effect I see of Intel’s choices is strengthening of workload preferences for Lion Cove+Skymont or Zen 5. For the numerically intensive compute we mainly do Zen 5’s gains are unreal—no other arch update I’ve worked with in the past three decades is even vaguely close—but Arrow tends to bench a little better on scalar tasks like compiles. AMD’s dual decode and supporting investments do a lot for hyperthreading (not uncommon I see ~35% improvement over Zen 4) but if a workload likes Skymont then the E-cores’ll win.

Given the lack of data to answer your power questions you’d likely have to buy both and measure. Broadly, though, it looks like Arrow gets less from N3B than Granite Ridge does from N4P and its idle power advantage’s plausibly attributable to Foveros. Rather than let our builds idle overnight we sleep them at the end of the day and use a wake task in the morning. Easiest sizeable power savings available and it’s arch independent.

It’s really not. Mostly what’s happened is forum enthusiast folk have switched to mostly building Ryzen, which centers DDR5 discussion on AMD. Arrow leapfrogs Granite Ridge in some details but perceptions of Intel being better often derive from 1DPC 1R gear 4 headline clocks. Little performance attention’s given to those but it wouldn’t surprise me if they geomean a bit worse than lower gear 2 clocks and, as working builds are usually 1DPC 2R or 2DPC 2R, it’s of limited practical relevance anyways.

No. PPT’s decoupled from efficiency as desktop parts’ power limits are usually well into the flat part of the power-performance curve. Also, the default PPT for 170 W TDP AMD is 230 W, not 200 W.

That said, lowering power caps increases efficiency by excluding operation in the little to no return parts of the power-perf curve. 105 or 120 W eco mode for the 9950X, for example. AMD pushed eco mode mainly for Zen 4, so most of the data available’s for 7700, 7900, and 7950X but the back and forth over 9600X and 9700X’s yielded Zen 5 data as well. IIRC there’s also a couple reviews with small amounts of 9950X eco mode data.

Not sure if there’s open data for Arrow power scaling but N3B and N4P are pretty similar, so probably there’s not a large divergence in behavior.

Upclocking increases IMC and DRAM power regardless of architecture. V²f applies to the 285K as well.

Usually not. Exceptions tend to be things like fan stop on the CPU cooler.

Usually not, unless the cooling solution’s badly underspec. If you’re expecting operation at 35-40+ C ambient temperatures it’s more of a concern, though.

Most common one I hit is 3.5s refusing to spin down. GPU min power tracks with display resolution and refresh rates, so can get substantial with multiple fancy displays.

Thank you for your detailed response and corrections! I find this very helpful information. If you’ll humour me, I have a couple of follow ups.

I was worried this might be the case, but I’m not sure I’m up for spending so much money just to compare the two. It’s why I feel, despite the disadvantages, that I might be leaning Arrow Lake: the power budget can always be toned down, but if idle power turns out to be a problem then there’s no real work around.

Hrm. What does this mean practically? All the power plots I’ve seen for 9950X under load have a solid line at 200W.

This is exactly the problem. I perhaps should have been more specific when I said “it gets warm”: there are two or three weeks in January every year with 40-45C days. Two weeks ago I swapped out an old Radeon HD 6950 for an RX 7800 XT and it dropped indoor ambient temps four degrees – which may not sound like a lot, but sleeping in 28C ambient is much more comfortable than sleeping in 32C ambient!

12 years between hardware refreshes??

I’d consider building a budget and running lower end hardware with a more sane lifecycle.

We typically run machines 6 years total in the enterprise:
3 in prod roles
3 in backup roles

1 Like

I’ve the same basic problem. We’ve enough AVX[-512]y stuff to call for Zen 5 no problem. But the carbon consequences apparently associated with AMD not being on CoWoS really aren’t great. So I try to educate for sleeping or shutting down machines when not in use. Predictably some users are pretty zealous, others neither remember nor care, but I wouldn’t really agree there’s no workaround for idle power. More involved and less effective workarounds, yeah.

Given Intel’s current position I’m lower confidence than I’d like to be as to Arrow’s ability to avoid repeating Raptor-like issues. N3B should be solid and N5P+N6 clearly are but that’s not the same as Intel’s use of them being solid. While I agree with your point about refresh cycles and socket life, I’m also not sure I trust Intel’s willingness or ability to invest in LGA1851 quality and sustainment. Particularly if it’s a single gen socket.

So I’m kinda wait and see. Probably that’ll mean skipping Arrow entirely and seeing if Intel ends up in a more convincing position with 18A. At the moment I’m concerned by how badly Intel needs 18A to get to a better financial state.

I don’t know why Phoronix data tops out at 200 W. Might be a test harness bug, might be the mobo and BIOS used given Zen 5 launch reviews are on beta versions. HWBusters got 240 W steady for EPS+ATX draw, suggesting Powenetics v2 picks up a few watts AMD doesn’t consider package power, Gamers Nexus got 223 W, and TechPowerUp 220. Realistically, to get entirely clear answers you’d have to measure the workloads you’re interested in and integrate the total power used.

FWIW, I don’t have a 9950X (yet?) but the workloads I run have no trouble parking a 9900X at its stock 162 W PPT 24x7 for 3-5 days. HWiNFO usually says 161.6 W package power or something within a couple hundred mW of that.

We get 40-45 C in the summer as well. Usually towards 45 we’re asked to start shedding compute load. If you have stuff that can’t be put off for a few days or shuffled into a more favorable part of the diurnal cycle then power limiting for January’s probably desirable. I’d try to plan HVAC climate mitigations, though, as cooler isn’t happening anytime soon.

I haven’t eco moded the 9900X yet but dropping a 7950X from 230 W PPT to 145 W (105 W eco mode)'s only a ~6% lower bench. Zen 5 pulls more than Zen 4 so I’d expect a larger drop with the 9950X. Maybe like 12-15%, which is still a net win on a 37% power cut.

For us it’s more like 10. ~5 years active, ~5 years as backup/overflow/somebody just needs a machine for something. I’ve been trying to pull that in by moving to building in house and waterfalling rather than buying prebuilts but hardware refresh is just starting to transition to more of { new CPU+DDR+mobo } + { move dGPU + drives } + { reuse case + PSU } model.

I have not used AM5 but in AM4 with both Ryzen 3950X and 5950X, idle power draw was incredibly low with Eco Mode enabled (from BIOS). I typically had about ~20W total CPU power draw at idle, with most of the CPU cores being put to sleep and the rest of the power coming from the SoC and other parts of the CPU. I have not read up on the Eco Mode in AM5 but I would hope it is similar. The idle power with Eco Mode on my 5950X is so good that many times, I have contemplated swapping it with a lower-tier CPU instead, and just could not justify it. I pretty much only run my AM4 system with Eco Mode and never had an issue with it.

It is, but the core sleeping you’re describing is default behavior unrelated to eco mode. As context, AMD APUs and Raptor idle ~4 W and Arrow’s ~13 W. So, er, while dual chiplet Zen 3 at ~20 and dual chiplet Zen 5 at ~35 W are low in comparison to workstation and server parts, they’re energy hogs in a desktop context.

I’d have to check back on Zen 4 but IIRC the 7950X I did idles ~30 W at DDR5-4800. At least on the board I built 7950X eco mode does some cool factory undervolt things farther out on the load curve but the voltage cut’s pretty modest.

1 Like

I’ve found it strangely difficult to find numbers like these reported, so this helps quite a bit!


The more I think about it, the more I think my decision hinges on whether 20-30W is a negligible amount of power in the context of heating a room. I kind of only just realised that’s something you can calculate!

Trying to properly model heating loads seems like a deeper rabbit hole than I want to entertain, but in the case where it’s hotter outdoors than indoors I think it simplifies to calculating sensible heat. If it’s hotter outside for eight hours, then over that time period a 20W source contributes to raising the temperature of 50-100kg of air about 6-11C. Obviously at some point heating loads come back into it and an equilibrium will be reached, but I find it’s illustrative nonetheless.

I think that largely seals the deal for me in favour of Arrow Lake.

1 Like