13900KF vs 13700KF vs 7950x vs 7700x vs 5800x3d

[Admin]

5800x3d has been cut in price
This happened right after the raptor lake launch. It's not around the same price as the 13600K which makes it justifiable if you already own an AM4 board and just want a better CPU and only play games.

https://pcpartpicker.com/product/CZ3gXL/amd-ryzen-7-5800x3d-34-ghz-8-core-processor-100-100000651wof

The 5800x3d does have the advantage of working pretty well with cheap DDR4 memory and having and being relatively power efficient. Still if you do productivity it's probably worth the extra cost to do the switch to raptor lake (you could just sell your old AM4 system).

 

[Admin]

Gaming performance at 1440P with a 4090
This is very much relevant since a lot of people are going to play at 1440p rather than just 1080p. Of course raptor lake is faster here but the difference in fps is a lot smaller than it would have been with a better GPU like the 5090 or an overclocked 4090ti (upcoming).

As we see here the 13700K when tuned is a very strong gaming CPU. You don't benefit from buying the i9 when it comes to gaming since very few games utilize more than 16 threads and even in rare games that does that the 13700K also has hyper-threading that offers additional MT headroom.

 

[Admin]

MT comparisons
While intel did destroy AMD in ST AMD does ok in MT. AMD is more expensive but you also get less power-consumption at stock.

When you push both hard via heavy overclocking (you need a custom-loop for that) the 13900K does beat AMD with a comfortable margin but when you use more typical cooling AMD might even win. Here it's basically a draw.

13900KF @ 6.2 Ghz

13900K @ 6Ghz

7950x @ 5.6 Ghz

As we see 13900KF can clock 600mhz higher all p-core than 7950x (custom loop) which contributes to intel being 10% faster in cinebench r23 MT. If intel had replaced the 16 e-cores with 8 additional p-cores the difference would have been even larger. AMD doesn't have anything coming close to the raptor-lake p-cores in performance (under heavy overclock).

 

[Admin]

https://www.techspot.com/review/2554-intel-core-i7-13700k/

 

[Admin]

Asus prime z790-p
The DDR5 version of this board looks great and reviewers (for linux) found it to work just fine.

https://www.phoronix.com/review/intel-core-i9-13900k

Now we have 2 customer reviews on newegg and it's bad

Here the issue seems to be that 2 of the slots have bad signal integrity making them basically unusable when the goal is high performance.

https://www.newegg.com/asus-rog-strix-z790-f-gaming-wifi/p/N82E16813119599

https://www.newegg.com/asus-rog-strix-z790-e-gaming-wifi/p/N82E16813119598

https://www.newegg.com/asus-rog-maximus-z790-hero/p/N82E16813119597

This is for asus prime z790-a wifi but it doesn't mean that the motherboard is bad. The issue might instead be the CPU memory controller (might need extra voltage).

This is for the asus tuf gaming z790-plus wifi D4. Not really a relevant motherboard but still a bad sign for asus z790 boards.

Notice that me providing links to newegg is not an endorsement of that store. You do want to look for better stores to buy from. There have been scandals of them shipping motherboards with bent pins.

Here is someone having good experience with the prime z790-p DDR5:

This of course doesn't tell us much but it's a minor positive sign.

 

[Admin]

The cheap option that might work: gigabyte z790 UD AX/AC
Gigabyte has a worse reputation than asus due to them screwing up badly in the past but if you want a cheap board with official support for DDR5-7600 it might be the only viable option.

https://www.gigabyte.com/Motherboard/Z790-UD-AX-rev-10#kf

https://www.gigabyte.com/Motherboard/Z790-UD-AC-rev-10#kf

 

[Admin]

MSI PRO Z790-P
While this is a budget board from msi it does almost everything you want from a motherboard. The main issue besides the fact that you cannot buy it yet is that it doesn't support more than 7000 MT/s officially. That however does not mean that DDR5 7200 isn't going to work, we just don't know yet.

It doesn't really make sense for nvidia to validate high DDR5 speed on their cheap models when they have more expensive boards to sell.

https://www.newegg.com/p/N82E16813144571

240$ and released the 4:th of november 2022.

There hasn't been any reviews on it yet but the previous z690-a was the best value option but has the issue of having even worse official DDR5 support.

 

[Admin]

Why you shouldn't 'invest' into an AM5 platform now
First off all buying hardware isn't actually an investment if you don't buy it to earn more money back. If you are buying something that will just depreciate in value without making you any money it isn't actually an investment. It's just you spending money for consumption.

One general issue with zen4 is that it comes with what used to be HEDT pricing without delivering HEDT features such as quad channel ram or 32+ pci express lanes directly from the CPU. You are still limited to at most 16 cores and you cannot use 128 GiB of DDR5 without dropping down to just 4000 MT/s which is well below what good DDR4 supports.

Raptor lake support 40% higher memory bandwidth when 128 GiB DDR5 is used.

Some people have suggested buying a low-end zen4 CPU now in order to upgrade later to a better CPU on the same platform but that really doesn't make sense. It makes much more sense to wait until there is actually a CPU you want to buy and then have much better motherboard options to choose from.

The current am4 motherboard chipset seems to be bottlenecked by a single gen4 x4 connection, AMD and motherboard vendors have been awfully quiet about this which is explained by them having something to hide (same with pluton).

anandtech.com/show/17585/amd-zen-4-ryzen-9-7950x-and-ryzen-5-7600x-review-retaking-the-high-end/4

While zen5 is very likely to work with AM5 you are probably going to miss out if you use an old X670E motherboard rather than the newer X770E or whatever. You might also realize that you want to use intel instead and then you end up having to not buy your preferred products or give up on your AM5 'investment'.

Since zen4 has a bad memory controller there is no way to test the signal integrity of AM5 boards when it comes to potential DDR5 overclocking (for zen5, etc) since the CPU imc limit you long before signal integrity of the motherboard becomes a factor.

One downside with using old hardware besides probably missing out on performance/features is that you then need to disassemble your old system, doesn't mean that you shouldn't do it but often it makes more sense just to build something new, you might want a second computer for example. If you for example buy a cheap DDR4 board to re-use your old DDR4 you will need to take that DDR4 away from your old system which makes that computer unusable until you find replacement ram, upgrading to DDR5 later is an option but then you also need to replace the motherboard so you might aswell buy good DDR5 right now (it isn't to expensive) and get that extra performance right away.

 

[Admin]

msi pro z790-p seems like an improved z690-a
They have tweaked the VRM-cooling and upgraded an NVME slot and pci express slot from gen3 to gen4

This means that z790-p is probably a fine motherboard given that the z690-a basically did everything needed of of a z690 motherboard. The main issue the z690-a had was sub-par vrm-cooling and this issue seems to have been fixed with the z790-p. The new VRM heatsink looks like massive overkill:

https://youtu.be/QdSG-0KPnDA?t=414

 

[Admin]

asus prime wifi ddr5 Z790-p vs z690-p
When looking closely at the board we can spot minor differences, they didn't just replace the chipset (and the pci express from chipset upgrades).

 

[Admin]

I was right about 13900KF > 13900K
Looking at the top cinebench R23 overclocks that didn't utilize liquid nitrogen we see that KF CPUs are indeed reaching higher overclocks. This probably means that intel is already saving many of the best chips for 13900KS while not letting anyone buy that now. Chips where the iGPU isn't working cannot be saved for 13900KS and will instead have to be sold as 13900KF even if it clocks really well.

hwbot.org/benchmark/cinebench_-_r23_multi_core_with_benchmate/rankings?start=0#start=0#interval=20

Best 13900K overclock is to around 6100 MHz while there are 2 KF overclocks reaching 6200 Mhz (excluding liquid nitrogen overclocks).

 

[Admin]

Efficiency in typical usage
While zen4 is a bit more efficient in heavy MT at stock settings raptor lake actually pulls ahead in efficiency in laods that are more lightly threaded.

As we see the 7950x draws more power when working on a single-thread while delivering significantly worse performance.

Unfortunately the 13900K isn't as effective in multithreaded loads due to 2 e-cores being weaker than a single p-cores forcing clocks and voltages to be pushed up further to match AMD in MT loads.

 

[Admin]

Why sapphire rapids might crush AMD
AMD is currently charging obscene margins on their threadripper pro lineup. They also broke their promise they made when it comes to supporting X399 indicating that we really shouldn't trust them when it comes to future platform support.

Looks like we are going to see a monolithic version with up to 34 cores in addition to a chiplet version with up to 64 cores

arstechnica.com/information-technology/2022/11/intels-oft-delayed-sapphire-rapids-xeon-cpus-are-finally-coming-in-early-2023/

https://www.angstronomics.com/p/monolithic-sapphire-rapids

The monolithic die will have mesh cache instead of ring-cache which will harm performance in some applications (like gaming).

A monolithic 32-core sapphire rapids CPU would utterly crush the 32-core zen4 AMD counterpart that isn't even launched yet. The intel raptor lake p-cores are vastly superior to the zen4 cores and when you add monolithic design and avx-512 support (due to only having p-cores) to that it will be a one-sided massacre.

AMD has planned to release a 96-core zen4 server CPU which might be able to beat the 60-core chiplet version currently planned by intel but even here the new competition will push AMD towards lower prices since people could otherwise just go for a dual CPU intel system for the additional cores (and being able to cool easier).

Intel does seem to try avoiding the chiplet approach as far as they can indicating that the chiplet approach come with a loss in performance/efficiency (like with AMD CPUs). Another factor is that they actually get more cores/wafer with the monolithic chips due to the interconnects required.

Doing some multiplication, this wafer has 2220 cores printed, while the monolithic design has 2312 cores on the wafer. The chiplet design has lower core density as the chip connection interface takes area as well.

The 64-core AMD CPU will just use 4 chiplets compared to the 9 used by AMD and intel will not use long interconnects potentially enabling much better latency between the cores (reducing the performance cost chiplets come with).

One big issue with the zen chiplet architecture is that the infinity fabric clock doesn't match the clockspeed of DDR5 which adds even more latency when reading and writing to RAM. Another big issue is that the dies are not even connected directly to each other, instead everything will have to go through the big io-die further adding latency and heat.

 

[Admin]

Why HEDT raptor lake might become the superior gaming option
The HEDT raptor lake CPUs will have significant benefits over their LGA1700 counterparts even for things like gaming

0. You get more p-cores
1. Much better memory bandwidth (at least double the number of channels)
2. More total cache
3. AVX-512 support (very valuable for emulation).

The 2 major potential downsides are higher price (probably much more expensive per p-core despite having no e-cores) and lack of ring-cache but having less efficient cache might not be enough to negative all the advantages offered by the upcoming HEDT platform.

Currently with raptor lake the much weaker e-cores will be used when a game loads more than 8 threads and this is far from ideal, while having p-cores is still much better than having no extra core at all (forcing the game to use HT instead) it's will still result in significantly worse performance than just having 16 p-cores.

 

[Admin]

Meteor lake might not be worth waiting for
The number of p-cores will be reduced to just 6 and they will use a separate die for the io-section, both of these design decision will result in worse performance.

hardwaretimes.com/intel-14th-gen-specs-only-6-p-cores-on-meteor-lake-desktop-cpus-2nm-arrow-lake-s-with-8p8e-cores-rumor/

We might however see intel offering something much better for the HEDT segment. The reduction of the number of p-cores to just 6 does signal that intel are going to have an HEDT platform for enthusiasts instead. Intel might end up still making their higher performance HEDT monolithic (except for over 42 cores) since people in that segment prioritize performance over cost.

The performance loss from not being monolithic will be smaller since all cores and cache will still be on the same die, there will still be higher memory latency though. If the memory controller is on the CPU-die there might not even be any real performance loss (there would be added pci express latency but that might not really matter).

Unclear to me if the area below the CPU tile is a memory controller on the same die as the CPU (would be the best-case scenario for meteor lake). The worst-case scenario is data having to travel throught the SOC tile and then to the IOE tile before it can be sent to the DDR5.

 

[Admin]

DDR5 speed is very important
You really do not want to cheap out getting cheap DDR5 or any DDR4 to raptor lake.

This is also why we need HEDT, doubling the memory bandwidth by going to quad channel DDR5 would provide a massive performance uplift.