New Zen microarchitecture details

[hojnikb]

Wow, looks like they are still sticking with 2MB of cache. Very disappointing if true.

 

[jpiniero]

Wow, looks like they are still sticking with 2MB of cache. Very disappointing if true.

I'm sure OEMs are demanding 35W models, and 2 MB instead of 4 may have been a sacrifice to be able to sell 2M models easily at 35W. Plus cheaper.

 

[frozentundra123456]

http://wccftech.com/amd-am4-%C2%B5opga-socket-1331/

Low clocked 2 module part has higher TDP than several higher clocked 4 module parts? That seems really strange.

 

[ShintaiDK]

Low clocked 2 module part has higher TDP than several higher clocked 4 module parts? That seems really strange.

Its just a trashbin for all failed parts one way or the other I guess. No need to waste more money on testing. High voltage may be the key.

 

[JDG1980]

Wow, looks like they are still sticking with 2MB of cache. Very disappointing if true.

Isn't Bristol Ridge the same silicon as Carrizo? I wouldn't expect any increases in cache size; this is a stopgap product based on an architecture that is already EOL.

 

[ShintaiDK]

Exactly the same yes.

 

[NTMBK]

So you are willing to buy a system for future upgrade without knowing how a future CPU will perform? D:

But it went so well with AM3+!

 

[el etro]

Wow, looks like they are still sticking with 2MB of cache. Very disappointing if true.

Lower manufacturing costs, to make Zen cheaper to most consumers.

 

[AtenRa]

Low clocked 2 module part has higher TDP than several higher clocked 4 module parts? That seems really strange.

It can maintain turbo longer than the 35W TDP. It could also keep the same base CPU clocks when iGPU is used without downclocking the CPU part.

 

[nismotigerwvu]

Lower manufacturing costs, to make Zen cheaper to most consumers.

Bristol Ridge isn't Zen, it's Excavator for the desktop. The reason it is relevant is that it's launching the socket that will be used by Zen. As Bristol Ridge is essentially the same silicon as Carrizo, the 2MB L2 cache is a known entity. From what we've seen from Carrizo, these chips will be a reasonable step up from Kaveri, but still no where near anything Intel is selling in most cases.

 

[el etro]

Bristol Ridge isn't Zen, it's Excavator for the desktop. The reason it is relevant is that it's launching the socket that will be used by Zen. As Bristol Ridge is essentially the same silicon as Carrizo, the 2MB L2 cache is a known entity. From what we've seen from Carrizo, these chips will be a reasonable step up from Kaveri, but still no where near anything Intel is selling in most cases.

My bad, didn't get the conversation.

But anyway technical aspect remains valid: 2MB L2 make no significant performance penalty over 4MB L2, and the latency is even better, making excavator cache to perform better.

 

[Zodiark1593]

I think it's just another angle to pick on something. After over 40 years of computer repairs and upgrades, I've discovered that if you're sober and pay attention then things don't get bent.

I don't have a problem with either PGA or LGA myself. The PGA used in mobile parts look like they could bend if you so much as look at them funny, but those are no problem either.

 

[DrMrLordX]

But anyway technical aspect remains valid: 2MB L2 make no significant performance penalty over 4MB L2, and the latency is even better, making excavator cache to perform better.

Overall you're right, but some folks are convinced it's gonna hurt game performance.

 

[ShintaiDK]

Sometimes cache speed matters more than size

 

[jpiniero]

But anyway technical aspect remains valid: 2MB L2 make no significant performance penalty over 4MB L2, and the latency is even better, making excavator cache to perform better.

The L2 would have to be really awful for this to be true.

 

[PG]

Overall you're right, but some folks are convinced it's gonna hurt game performance.

Haven't you seen reviews?

http://www.jagatoc.com/2016/03/hands-on-review-overclocking-athlon-x4-845-carrizo-murah-tanpa-igp/6/

Why is the 860K at stock beating the 845 in games, (see 3DMark and GTA V) but it pretty much loses everywhere else? There is a reason people have this perception. If someone else has a better explanation I'd like to hear it.

 

[itsmydamnation]

The L2 would have to be really awful for this to be true.

the L2 itself isn't awful, but that way CON cores cache is designed is. L2 has really high latency and it has nothing to do with the speed of the arrays but the way it connects to the module.


edit: CON core is resource limited , (two SSE units over 2 cores) ( two alu's a core), there is lot of SSE code in games meaning the other improvements in EX dont have to execution resources to show themselves in games. The differences you see there are easily within the difference of clock speed.

 

[Tuna-Fish]

The L2 would have to be really awful for this to be true.

The problem is actually that the L1 is awfully designed. L1 is write-through, so L2 write throughput limits L1 throughput. L2 throughput is limited by number of possible parallel accesses and by latency, so reducing L2 latency improves L1 throughput.

 

[VirtualLarry]

I'd build a Bristol Ridge system if there was confirmation that I could later plop in a Summit Ridge upgrade. That'd be great.

Ditto!

 

[DrMrLordX]

Haven't you seen reviews?

That's exactly what I meant.

 

[Soulkeeper]

So you are willing to buy a system for future upgrade without knowing how a future CPU will perform? D:

Not exactely, but it should be a certainty that Bristol Ridge is faster than my current system.
If Summit Ridge was a regression from that level (which is doubtfull), then i'd just not "plop" in an upgrade.

 

[Exophase]

The L2 would have to be really awful for this to be true.

Excavator doubled the L1 dcache size, that could go a long way in undoing the damage of the halved L2 cache.

 

[Azuma Hazuki]

Not as much as making it write-back instead of write-through would :/

Although I'm no expert in CPU design, the thing that stood out to me when Bulldozer first hit is how slow and insufficient the caches were. I really think the poor computation cores are starving most of the time

 

[deasd]

http://www.bitsandchips.it/52-english-news/6815-speculations-about-zen-after-our-april-s-fool

The news of yesterday was an April's Fool, and at the present moment is not a secret anymore. The AIDA64 screens were fakes, but there is a kernel of truth.

First of all, A0 samples of 8 cores Zen exist, and it seems that AMD has already delivered them to its partners for the preliminary tests. The base frequency seems to be very high (3 GHz, Turbo still not enable with these ES) for a high end 8 cores CPU, also produced on a new node like the 14nm LPP FinFET of Samsung/GloFo (The first ES of Bulldozer worked at 2,8 GHz). It is a promising start.
About Zen, I can tell you some my speculations (NO OFFICIAL INFO) and some news from USA (from the same guy who told me about Zen, 3 days earlier the Official Presentation made by AMD).
If 14nm LPP will be good, Zen base frequency (8 cores version, TDP 95W) will be 3.7-3.8 GHz. Turbo 4.1-4.2 GHz. If 14nm LPP will be very good (or if AMD will commercialize a limited edition, like the Athlon Slot A 1GHz during the good old times), Zen will have a base frequency of 4 GHz (Turbo 4.2-4.3 GHz). Also, the overclockers will have a lot of fun, because of the soldered IHS!
Zen seems to be a High Performance and High Frequency uArch, like Skylake and Kaby Lake, but with some little tweak. The design team lead by Keller, as you know, has chosen to simplify some features in order to limit the Power Consumption and improve the Frequency, due to the low availability of appropriate software at the present moment (e.g. 2 x 128 Bit FMA instead of one big 256 Bit FMA SIMD). FPU units, however, will be very strong. In theory, Zen will be a monster in video games tasks (It's just an example). We can expect an IPC at Broadwell level. And if some companies want some features, AMD can do a Custom Zen CPU/APU (It has a modular design).
Last, but not the least, the part of the news about Intel was true. It seems that Intel will commercialize a 10 cores Broadwell-E due to its own internal Zen simulations (Intel has some data we don't know, yet, and has skilled engineers that know what they do). Zen will have 8 cores, while Broadwell-E 10 cores. So, Intel will have the “King of the Hill”, but Zen seems to be very good in the mid range of the market (AKA, gaming and enthusiast configurations).

 

[jhu]

Sometimes cache speed matters more than size

It's how you use it that matters.