Can someone please explain CCX to me?

[horrorshow]

This is quite confusing....

I've been keeping tabs on the R5 1400X and R3 1200x in particular.

As these cores, true cores?

I could care less about threads etc.

In theory, shouldn't an R3 1200x 4 core oc'ed to 3.8-4.0ghz be competitive to an i5?? Within reason obviously..

(I know CCX has something to do with caching, but the more I read, the more it sounds like AMD is pulling a fast one yet again....)

 

[Bandalo]

They are true cores.

They group the cores together with their own chuck of shared cache into a little block, then put these blocks together to make the CPU.

The problem comes when a task is being shared across cores in different CCX blocks. Say core 1 is in one CCX and core 5 is in the other. If both of these cores are working on the same task, they don't share the same block of cache memory. So extra clock cycles have to be used to sync up their cache memory so they can work together.

Intel doesn't do this the same way, and in general all the cores have access to the same shared cache, so there's no cross-core cache penalty.

Intel Broadwell (10 core version):

 

[horrorshow]

Why.... would they do that??

Clearly there's an IPC penalty....

 

[Tsumi]

Not this true core nonsense again…

But with regards to CCX, it is the connection between the two modules in Ryzen. The modules are at the L3 cache level rather than the core (L1) level of Bulldozer.

Why.... would they do that??

Clearly there's an IPC penalty....

There is no IPC penalty. There is a multithreaded scaling penalty in non-Ryzen-aware programs.

 

[Bandalo]

Why.... would they do that??

Clearly there's an IPC penalty....

There's not really an IPC penalty per se, there's only a penalty sharing tasks across cores in different CCXs when you have cache sync issues.

And as to why, well, the cache controllers and cache management inside an Intel CPU is REALLY freaking complicated. It's a very hard problem to keep everything in sync, avoid cache misses, etc, etc. The more cores you add, the more complicated it gets.

AMD's solution isn't bad...it makes it easy to scale from 2 core to huge chips, since you're basically "cut and pasting" CCX blocks. It does have some drawbacks, but it's not a terrible solution. Intel's is more elegant and probably faster, but it's a heck of a lot more expensive as well.

 

[horrorshow]

Soooooo.....

Based on speculation, would an R3 1200X oc'ed to 3.8/4.0 be a significant upgrade over my old-as-dirt i5-750?

Or, should I just throw my money at Intel, bleed my wallet, and get a 7600K/Z270 platform?

 

[Bandalo]

Soooooo.....

Based on speculation, would an R3 1200X oc'ed to 3.8/4.0 be a significant upgrade over my old-as-dirt i5-750?

Or, should I just throw my money at Intel, bleed my wallet, and get a 7600K/Z270 platform?

No f'ing clue, since there's way too many factors to consider. Wait a week and there'll be a dozen benchmarks out you can use to compare for games/apps you use.

I personally like Intel right now, but I think AMD is going the right way. When Ryzen 2 hits, I may jump on the bandwagon.

 

[Tsumi]

An R3-1200x at 4 ghz would approximately be equivalent to a first gen i5 quad-core at 4.5+ ghz. The micro-op queue Intel introduced with Sandy Bridge and now implemented by AMD was a significant boost in IPC from Nehalem.

 

[horrorshow]

An R3-1200x at 4 ghz would approximately be equivalent to a first gen i5 quad-core at 4.5+ ghz. The micro-op queue Intel introduced with Sandy Bridge and now implemented by AMD was a significant boost in IPC from Nehalem.

Sounds like I might be holding onto this old-ass CPU a bit longer..

(Unless [H] benches prove otherwise)

*still playing "the waiting game"*

 

[Topweasel]

An R3-1200x at 4 ghz would approximately be equivalent to a first gen i5 quad-core at 4.5+ ghz. The micro-op queue Intel introduced with Sandy Bridge and now implemented by AMD was a significant boost in IPC from Nehalem.

Well I don't know IPC wise it's probably closer 3rd Gen, in terms of IB so like a 3500k. I also don't think what we are seeing in games is a ccx penalty (it could be a part but, I have other theories). But it's really not just about straight up performance, a Quad core at Sandybridge level or higher, including Ryzen is just going to feel like a side grade. The only reason to really upgrade is dying parts or better board featureset. Otherwise you won't really get a good feeling for the "15-20%" faster it's gotten over the last 5 years without doing stuff that loads all the cores to 100%. So while the 750 is outside the envelope I pointed out, I wouldn't do it unless you were looking for a reason to swap. That said a quadcore for i3 prices and i5 performance that supports NVME on sub $100 boards is a great to minimize the minor performance increase portion and allow funds to go after the parts that will feel better like memory, vid card, and hard drive.

 

[Tsumi]

Everything I have read shows IPC being more or less at Haswell levels, which is about 5-10% above IB levels. I believe Ryzen would be a good upgrade for first gen Core i3/i5/i7 unless you had one that overclocked well above 4 ghz.

 

[Topweasel]

Everything I have read shows IPC being more or less at Haswell levels, which is about 5-10% above IB levels. I believe Ryzen would be a good upgrade for first gen Core i3/i5/i7 unless you had one that overclocked well above 4 ghz.

You know what I missed the .5GHz part of that. I thought you were saying it was on par with Clarksdale clock for clock. I was just trying to be conservative.

 

[DuronBurgerMan]

The IPC, according to Userbench, was around 9% less than Kaby Lake/Skylake (or a little bit less, even, than that). So IPC is closer to Broadwell than Haswell. But this is overall.

In throughput tasks, Ryzen's IPC is very close to Kaby Lake (but, of course, it suffers from a clock speed penalty). In latency-sensitive tasks, it appears to be a bit slower, even, than Broadwell. Probably more comparable to Haswell there. The catch to all of this, however, is that latency-sensitive tasks may benefit from a little Ryzen awareness to avoid cross-CCX traffic when not necessary. So Ryzen may see improvement in that area as time goes on. The potential, at least, is there, if AMD can be properly supportive of developers (that's a big if, though).

On average, I'd put the IPC roughly in the Broadwell ballpark.