Fuck AMD and intel

Admin

Administrator
Moderator
Messages
3,616
#1
Intel refuses to provide more than 8 performance cores unless you go for their super-expensive and outdated alder lake HEDT lineup. Instead they waste space for estrogen cores which is not want you really want on desktop.

raptor-lake-die.png


AMD instead pushes chiplets down your throat despite how much performance you lose from it.

The Z790 motherboards are also overpriced and have poor connectivity even compared to AM5.
 

Admin

Administrator
Moderator
Messages
3,616
#2
Low end desktop raptor lake isn't that great
It used to be the case that high end desktop CPUs were fairly affordable.

i7-3930K had an MSRP of 600$ and then you actually got 40 pci express 3.0 lines and quad channel ram.

Now instead you have to pay 600$ for the low end desktop CPU 14900K which only has 8 performance cores. It's a ripoff since it's 180$ more expensive than the 14700K while only provided 4 more estrogen cores and maybe slightly better overclocking headroom on average.

With raptor lake s you only get 16 gen5 lanes which are all used for the GPU, the issue is that there is currently no GPU actually utilizing gen5.
 

Admin

Administrator
Moderator
Messages
3,616
#3
Sapphire rapids: expensive, poor overclocking headroom, very power hungry
The single-threaded performance is pretty bad due to the CPU being limited to 4800 Mhz by default

13900K in geekbench6 2963 in single-thread performance, 20021 in multi-thread performance.
2495x in geekbench6 2373 in single-thread performance, 18728 in multi-thread performance.

This was disappointing since we were hoping to see great things from 24 performance cores.

Pretty bad for a CPU costing 2130$ on newegg.

It also doesn't clock particularly great, seems like 5200 Mhz is about the best you can get with water-cooling.

https://hwbot.org/benchmark/cinebench_-_r23_multi_core_with_benchmate/rankings?cores=24#start=0#interval=20

So if you want the best single-threaded non-gaming performance you basically have to go for the 14900K/14900KF.
 

Admin

Administrator
Moderator
Messages
3,616
#4
Issues with zen4
While AM5 is a better platform than Z790 by offering significantly better connectivity and isn't end of life according to AMD the zen4 CPUs hasn't been great.

The 7950x lost to the at first cheaper 13900K in gaming and single-threaded productivity.

The 7800x3D does perform well overall in gaming but it only has 8 cores so it's not really that great overall.

The 7950x has a mixed architechture where one chiplet has 3D cache while the other chiplet lacks it. Unfortunately they did not provide a proper hardware scheduler so people neded up having to mess around with windows to not get worse performance than the 7800x3D.

You also only get dual channel ram, it's not really high end desktop.
 

Admin

Administrator
Moderator
Messages
3,616
#5
Why the AMD chiplet approach is awful
Instead of providing one large monolithic die they separate each cluster of 8 cores into a separate die and then have very slow and energy consuming interconnects to communicate.

When sending data to and from memory it has to pass through the IO-die and then finally it can reach the ram, this adds latency (which is awful for gaming) and also limits the bandwidth (due to the zen4 interconnects having less bandwidth than dual channel DDR5).


Unfortunately the infinity fabric only clocks to around 2300 Mhz so forget about the 1:1 thing that people used to do with zen3.

So at best you get 137 GiB/s in read and 77.5 GiB/s in write, this is with 2 CCDs.

Zen4 actually has two Infinity Fabric links per CCD but on Zen4 desktop only one of those links are actually used.

Under load zen4 is still pretty efficient (overall better than raptor lake) but zen4 actually consumes more power during idle (around 10W more) which does add up over time if you like me have your computer on 24/7 (around 10$ per year in electricity cost).
 

Admin

Administrator
Moderator
Messages
3,616
#6
DDR5 overclocking with raptor lake is often a nightmare
Their quality control for their memory controller is abysmal and it's a nightmare to get solid stability since you often run into stability issues that are very hard to find (such as after 48 hours of stress-testing). This is especially a prominent problems with motherboards cheaper than 500$.



This is especially a problem when you try using more than 48 GiB of memory since dual rank memory is around 15% harder on the raptor lake memory controller (so you might only reach 7200 MT/s instead of maybe 8200 with top tier motherboard)

https://vintologi.com/threads/ddr5-overclocking-nightmare.1229/

 

Admin

Administrator
Moderator
Messages
3,616
#7
Meteor lake doesn't look good
Consistently with meteor lakes leaks have indicated only 6 performance cores on low end desktop (successor to LGA1700). The thing that has varied in leaks is the number of estrogen cores going as high as 16 and as low as 8.

https://www.techpowerup.com/313286/intel-core-ultra-meteor-lake-h-partial-lineup-leaked

Intel will opt for a chiplet approach which when reading and writing to ram, there will not however be additional latency when sending data between cores since all cores will still be on the same die (making it better than AMD in that aspect at least).

https://www.pcworld.com/article/2055579/inside-meteor-lake-intels-14th-gen-core-cpu-optimized-for-the-future.html

So what about people who actually want great performance on desktop? well the 14900K will always be an option for around 600$.

Of course AMD should eventually release zen5 but then the price will be very high if their isn't any direct competition from intel. AMD has a history of ripping their costumers of every time instead doesn't compete.
 

Admin

Administrator
Moderator
Messages
3,616
#8
Bad/awful core to core latency when you try to use more than 8 cores
This is the case both for the 13900K and 7950x. Sure it has more than 8 cores but when games try to actually use more than 8 there will be problems.



 

Admin

Administrator
Moderator
Messages
3,616
#9
Intel abandoned AVX-512 in favor of e-cores
It is often claimed that this would not provide a benefit for gaming which is utter nonsense, it is very valuable for emulation which makes it something you definitely want for gaming.

https://www.tomshardware.com/news/ps3-emulator-avx-512-30-percent-performance-boost

Initially you could get AVX-512 support on alder lake by disabling the e-cores and having suitable bios & motherboard but later intel began to fuse that off for no good reason.

https://www.tomshardware.com/news/how-to-pick-up-an-avx-512-supporting-alder-lake-an-easy-way

Of course it would have been possible to engineer the e-cores such that they would be able to handle AVX512 instructions (they will probably do that at some point) but they ended up pushing out the e-cores to consumer before they were ready instead of actually delivering more performance cores (instead raptor lake dropped down to 8 performance cores from 10).
 

Admin

Administrator
Moderator
Messages
3,616
#11
AMD has a history of screwing over costumers who wants more than 16 cores
You may ask "what about threadripper" and AMD tried push people towards the super-expensive threadripper pro instead, this was not due to lack of demand, it was due to them wanting to push people towards much more expensive products.

AMD promised future CPU upgrades to TRX4 breaking earlier promises in order to push people towards much more expensive threadripper pro CPUs.

https://youtu.be/T_r0WRexILo?si=QUFxwbaucpZ3jxCK&t=192

AMD tried charging 3299$ for the 32-core Threadripper Pro 5975WX

https://hothardware.com/news/amd-ryzen-threadripper-pro-5000-wx-cpu-pricing-revealed

Which is a significant jump from the 1999$ the 3970x was launched for (which is also rather expensive for 32 cores).

They also screwed over their their old costumers not releasing new CPUs for TRX3

https://www.notebookcheck.net/AMD-justifies-the-shift-to-sTRX4-socket-for-3rd-gen-Threadrippers-and-promises-long-term-support-but-current-Threadripper-owners-appear-to-be-unconvinced.442727.0.html

AMD has finally plans to return to HEDT now but this is 1 year after launching zen4, you end up having to wait a year and paying more than twice the current price per core (compared to 7950x)

https://www.pcworld.com/article/2109168/amds-monstrous-threadripper-7000-returns-to-desktop-pcs.html

Of course their prices would be even higher if it wasn't for intel launching sapphire rapids earlier (which was a big disappointment).
 

Admin

Administrator
Moderator
Messages
3,616
#12
The e-cores are not actually efficient
It's unclear to me why intel is insisting on using their shitty hybrig approach when they are not even gaining efficiency from it.

When you actually do the math and look at the performance/W we actually see that there isn't really much gain here. 5 e-cores will perform about the same as 2 p-cores in terms of multithreaded performance

https://www.pugetsystems.com/labs/articles/intel-12th-gen-how-do-p-cores-and-e-cores-compare-2289/



https://chipsandcheese.com/2022/01/28/alder-lakes-power-efficiency-a-complicated-picture/

Unfortunately with raptor lake the e-cores do not get their own separate voltage since the all the cores get the same voltage with raptor lake, this prevents the e-cores from actually being utilizeable as actual efficiency cores. Hence "estrogen cores" is a more appropriate term.
 

Admin

Administrator
Moderator
Messages
3,616
#13
AMD bottlenecked their X670 chipsets
Ironically the shitty Z790 platform intel is offering actually has twice the bandwidth to the chipset, how ironic.


Yes that's right, both PROM21 chips used are fed by a single just 4 pci express 4.0 lanes, it would have been much better to go for pci express 5.0 for the connection between the CPU and the PROM21 but it's AMD so of course they opted to not bother to actually deliver greatness.
 

Admin

Administrator
Moderator
Messages
3,616
#14
Proof low latency is important for games
First let's start with latency between the CPU cores and DDR5


Here the tuned DDR5 at 6000 performed significantly better than XMP at 7200 (MT/s).

https://www.youtube.com/watch?v=RTmbYak_8gE

Low latency was also important with DDR4:


We see that AMD CPUs gets a massive performance uplift in games when more cache is added.

1697966197661.png


https://www.tomshardware.com/reviews/amd-ryzen-7-7800x3d-cpu-review/4

We are also seeing the 7700X with PBO outperforming the 7950x. At the same frequency it's basically performing the same:





https://www.sweclockers.com/test/34873-amd-ryzen-9-7950x-och-ryzen-7-7700x-raphael/13

This despite the 7950x having twice the number of cores and twice the total cache.

Note that it will be hard to pinpoint whether the issue with the infinity fabric is mostly added latency or mostly limited bandwidth (preventing effective communication between cores/cache). But if it's mostly a bandwidth issue why didn't AMD put effort into making the bandwith better? planned obsolescence?
 

Admin

Administrator
Moderator
Messages
3,616
#15
Graphic cards
AMD isn't even trying to compete at the high end anymore. The 4090 is still alone at the top with no real competitor in sight.

Intel entered the GPU market fairly recently and their cards has had loads of issues with drivers and performed much worse than older cards even from AMD.

Main issue with the 4090 is not having support for displayport 2.0 or higher (you are limited to 1.4 speeds) which is very much not what you want when you are going to buy a new monitor.

At least with big navi (6900XT, 6950XT, etc) AMD actually got close to beating nvidia, what happened?

The 7900XTX couldn't even beat the 4080, what a disgrace.
 

Admin

Administrator
Moderator
Messages
3,616
#16
What raptor lake could have been
The die only needed to be made marginally smaller to replace the 16 estrogen cores with 6 performance cores.

raptor-lake-14c.png


The multithreaded performance would be about the same (maybe marginally worse).

There would be AVX-512 support.

No issues in any games.

Note that you could actually 16 performance cores in a smaller area by getting rid of the integrated graphics.

raptor-lake-16c.png


One issue with adding more cores however is that then the ring-bus has to run at slower speed to accommodate the extra cores so at some point adding more cores to the same die will stop making sense.

Currently most people opt to pay 30$ more to get integrated graphics, this might be that people actually want integrated graphics (such as in the case their GPU stops working) but it can also just be that people don't know what they are buying so they don't know that they can save 30$ by going for a KF variant instead (in exchange for not getting an iGPU).

And if there is a defect on the iGPU area intel can just sell it as a KF variant and only lose 30$.
 

Admin

Administrator
Moderator
Messages
3,616
#18
Why cannot you cool a 300W CPU with an aircooler?

It's a bit absurd how with the 4090 it can cool 600W just fine without any fancy liquid cooler but with CPUs just cooling off 300W is suddenly a monomental challange.

One big contributing factor is that the stock heatspreaders are simply bad so you need to "delid" your CPU to get decent termal performance which will very much void the warranty and will risk destroying a 600$ CPU.

The 7950x runs around 19° C cooler delidded

There are of course plenty of issues on the blue side as well

 
Top