Best CPU platform

[Bulat Ziganshin]

Really, some thoughts which CPU platform to buy. First, two rules of thumb:

• prices are highest in December and lowest in July, at least here in Russia. In the first half of year, each month waiting make things cheaper
• don't buy new technology right after release, wait 3-9 months till they will fix most of errors

In the mid-segment ($300-1000 for CPU) there are now 4 platforms, in order of cpu+mobo price:

• 8-core Ryzen: best m/t speed per dollar
• 6-core CoffeeLake: best s/t speed (as well as up to 6 threads), best memory latency
• 8/10-core Skylake-X: only platform with AVX512 support
• 16-core ThreadRipper: best m/t speed

So, AMD is good when you want best m/t performance for a given budget. They lose for any other scenarios. Best performance with up to 6 threads, or for latency-limited computations (including compilation and many compression algorithms) - choose CoffeeLake. Want to play with new AVX512 instruction set - Skylake-X.

For developers, Intel CPUs are especially preferable for software optimization because most of our program users are also using Intel CPUs. I.e. we should follow the cloud. In this regard, desktop CoffeeLake is preferable since Skylake-X has its own subtleties, and optimizing for Skylake-X isn't exactly the same as optimizing for desktop Intel CPUs.

Finally, only CoffeeLake includes GPU, very same GPU as employed by vast majority of computers. So, it has extra points as developer computer, since you can look into adding some GPU processing into your program and really test/optimize it for the most popular GPU.

So, today i will choose CoffeeLake, unless there are plans for AVX512. Ideally, just wait until desktop CPUs with AVX512.


Closest future events:

• Zen+ and X470 in March - probably will improve Zen CPU and memory frequencies
• Cheap CoffeeLake mobos based on H370, B350 and other reduced chipsets

As you see, they go in opposite directions, so B350+8700 will be cheaper but slower than Z370+8700K, while Zen+ will became faster.

[encode]

At the end of this year I'll upgrade myself to the i7-8700(K) or Ryzen+ (2600X/2700/2700X). Currenty, the Ryzen price/performance is exceptionally good! My overclocked i7-4790K looks like piece of junk now!
Anyway, I'd like to collect the performance statistics from the data compression field from current newest platforms, as we did in the Ryzen thread!

[Piotr Tarsa]

Ryzen processors based on 2 CCXes (so currently all desktop ones) are interesting because they work like two CPUs in a dual-socket motherboard. This is a problem on one side as the latencies between CCXes are big, but on other side that is an opportunity to check whether our algorithm scale well. If you achieve very good scaling on Ryzen CPU then with high probability your algorithm would also scale well across many CPUs in a single motherboard.

OTOH, there were many reports of unstability of Ryzen processors under heavy load. That makes them less interesting. I hope future Ryzen generations will be very stable.

[Bulat Ziganshin]

at the end of this year, Intel should also release AVX512 desktop cpu and 8-core desktop cpu, but they will be probably different products. Also, we will be pretty close to Zen2 launch. Prices will change too

I personally will prefer avx512 cpu with iGPU, 4+ cores and SMT - this includes everything required for diverse development

[encode]

Somewhat started to building my new PC. Since I need the fastest single core performance I chose Intel.
RAM is Corsair Vengeance LPX 3000 MHz CL15. But can't decide on CPU. Of cource, i7-8700K is the way to go, but the price right now is insane! The Core i5-8600K might be a good enough alternative for my needs.
i9-9900K - not even considered at this price
i7-9700K - no Hyper Threading and in major tests weaker than 8700K
i5-9600K - somehow a weaker overclocker compared to 8600K

[JangoFatXL]

Hi,

You might wanna wait a little bit. AMD wants to present Zen2 aka Ryzen 3000 series on the CES 2019 in just a few weeks. This sound very promising:

https://www.pcbuildersclub.com/en/20...s-at-ces-2019/

If you can't wait, still consider AMD
If you can't, yeah, you already know, what to take: 8600K.

But don't forget, when you want to upgrade in future, Intel has the bad habit to force you to upgrade the motherboard, too. AMD is very customer friendly, it looks, like there will be a 16(!)-cores CPU for the end customer market in an affordable price range.

I think I will upgrade my FX to an 8-cores Ryzen3000 first and later, when the prices drop, to a 16-cores.

[encode]

The most important things to me are the single core performance and the cache size. 16 cores difinitely will be the cool thing for the Cinebench. But I'm not care that much on a multithreaded performance. Just of right now the Intel has a higher performance per core + it is much more stable and mature platform, but at notable higher price unfortunately.

[JangoFatXL]

You are right, Intel has the better single core performance. I just hope AMD will catch up, we will see. And about the price, unfortunately the prices developed into the wrong direction

[encode]

Santa brought me the i3-8100 (due to insane Intel prices)! What a nice little beast! Now I can comfortable wait for i7-8xxx/i7-9xxx price drop...
BTW, with my Noctua NH-U12S, the temps under full load with Prime95 is about 50C!

[SvenBent]

at the end of this year, Intel should also release AVX512 desktop cpu and 8-core desktop cpu, but they will be probably different products. Also, we will be pretty close to Zen2 launch. Prices will change too

I personally will prefer avx512 cpu with iGPU, 4+ cores and SMT - this includes everything required for diverse development

just curios question
how much does avx impact compression/decompression ?

[comp1]

https://www.techspot.com/news/80685-...wing-only.html

https://www.cpubenchmark.net/singleThread.html

[encode]

Check out the memory latency + later, actual street prices

[Piotr Tarsa]

Ryzen 3000 is out and with regards to memory controller the write speed per chiplet is halved (but you can buy 2 chiplet AM4 Ryzens now, so that would restore the full write speed) and main memory access latency is somewhat worse:
https://www.guru3d.com/articles_page...review,21.html
https://techreport.com/review/34672/...pus-reviewed/5

However even the Ryzen 3700x still performs much better than Ryzen 2700x in multithreaded compression tests, it's in fact near Core i9-9900k. Depending on test Ryzen 3700x is faster (on synthetic benchmarks) or slower (on real file sets) than i9-9900k:
https://www.guru3d.com/articles_page...review,12.html
https://www.anandtech.com/show/14605...ing-the-bar/10
https://www.techpowerup.com/review/a...-3700x/12.html

As for compilation speed, Ryzen 3700x is slightly faster than i9-9900k (Phoronix test) or slightly slower (techpowerup test):
https://www.phoronix.com/scan.php?pa...0x-linux&num=7
https://www.techpowerup.com/review/a...7-3700x/7.html
Ryzen 3900x (which has price comparable to i9-9900k) is faster in compilation (slightly on techpowerup, strongly on Phoronix) than i9-9900k.

Skylake-X has generally lower single thread performance than Ryzen 3900x or i9-9900k when AVX-512 is not used. OTOH it has four memory channels and that can help a lot in multi-threaded compression tasks (but I can't find where I saw that).

Intel still doesn't have AVX-512 enabled desktop processor or any 10nm desktop processor. It seems that Intel won't present anything like that this year and I would say even in 2020 the top tier desktop processor from Intel will be another Skylake (non X) refresh on 14nm.

On AMD side it doesn't seem that AMD will ever make memory latency competitive to Intel, at least if they continue with chiplet design on desktops. Charlie from https://www.semiaccurate.com/2019/04...entiate-cores/ is writing about AMD going to "differentiate" cores but as I don't have subscription to SemiAccurate I don't know what he means. If AMD is going to keep monolithic (non-chiplet) design for (at least) APUs then they could improve memory latency in such designs. We'll see.