News AMD reveals Zen CPU details

It's great to hear it will hopefully offer much better performance(even if it will still be behind Intel's best) but I'm hoping the APU also see's a huge boost. Could make more a perfect HTPC or decent gaming rig for my nephew.

 

I'm sure this is a dumb question, but why are L2 and L3 caches so small and why do they never seem to get bigger?

Is there a reason CPUs don't have 10 MB caches?

 

Why should they get larger? It's not supposed to be a source of on-chip RAM. The smaller they are, the better latency read/write speeds become. In many cases, L2 has double or even triple the latency of L1, and can be more than 3x slower in read/write speed. The caches are meant to have just barely enough data to hold a set of instructions for the CPU to process and nothing more. RAM itself is technically a form of cache, but as you may know, the speed and latency of RAM is significantly slower than even L3 cache. Cache is useless if the registers are sitting there twirling their thumbs waiting for the cache to fill up with more instructions. So, the engineers likely design the caches to always fill up faster than the rest of the CPU can process.

For the record, RAM is no different. There is such thing as "too much" RAM (depending on your workload anyway). In many dual-channel DDR3 setups, you start to lose performance once you go beyond 16GB. I like to think of it like trying to grab a book in a large empty room. It isn't really all that hard to find it, but it still takes forever just to walk over there and get it.


Also for the record, there are CPUs out there with 10MB caches. They're just stupidly expensive. Interestingly, cache is one of the most expensive components of a CPU. An extra MB or 2 can have an equivalent cost to an IGP.

 

Disclaimer: Not a fan boy either way

The last AMD I owned was an X2 3800+ on the 939 platform. I'm the type of consumer that will buy what gives me the best perf / cash ratio.

So im genuinely excited for AMD as hopefully they have an alternative product for me and / or will force Intel to lower their prices. Same goes for the AMD/NVIDIA thingy.

I've got home upgrades which I will be holding off on for now in light of this.

 

That 128MB isn't processor cache - it's on-die RAM for the IGP.

 

No... the 5775C doesn't disagree with what I wrote. First of all, as you said, it's an L4 cache - that's pretty removed. Secondly, if you haven't noticed, each layer of cache is always multitudes larger than the previous. Seeing as L4, as you pointed out, is pretty rare, it isn't surprising that it is so large. That being said, it was made for specific workloads. I can assure you that it does not offer a significant performance enhancement to all (maybe even most) workloads; even L3 caches sometimes offer a negligible performance difference, but L3s became more popular thanks to IGPs. Lastly, if 128MB L4 caches were as useful as you seem to imply, they'd be more common. Again, I'm not denying that there's a performance benefit, but for the average application, the performance benefit is way too minimal considering the expense of making it. The 5775C is not a cheap product when you consider it's a quad core.

The point of my post was to say that caches are intentionally small for performance (and cost) reasons. If you had two alike CPUs, where one had a 256KB L1 cache and the other with a 2MB cache, the one with the bigger cache will run slower in just about every situation. If larger caches had no performance detriment, then what's the point of having so many layers? Why not just use one giant 64MB cache for each core? I'm not making this up - run memtest86+ and you will see with your own eyes that the caches are significantly faster when you compare the lower levels to the higher ones.

 

No, it CAN heavily increase performance. Big difference there. It doesn't always increase performance. I have no doubt that in tests involving the IGP, the 5775C will always win, but in strictly CPU or discrete GPU tests, it will NOT win every time. Here's some proof of that:
http://techreport.com/review/28751/intel-core-i7-6700k-skylake-processor-reviewed/4
http://techreport.com/review/28751/intel-core-i7-6700k-skylake-processor-reviewed/10
Look at benchmarks involving games with a discrete GPU and you'll find the 5775C isn't really a whole lot faster, but there isn't really any significant performance difference between the other i7s either.

Uh... because it is overpriced? Compare it to the i7-5820K, which:
* Is only about $10-30 more expensive
* Has 2 additional physical cores (and 4 total additional threads)
* Has 15MB of L3 cache (versus the 6MB of the 5775C) and L3 is significantly more useful than L4.
* Has 28 PCIe lanes (versus the 16 of the 5775C)
* Doesn't prioritize an IGP that nobody wants to use anyway
That isn't just mildly better, that's a whole other league for a very minimal price hike. And as shown in the links I provided, the 6700K is a substantial performer considering how much cheaper it is.

 

Don't forget that Skylake benchmarks at launch were hit by the FCLK issue lowering apparent performance. As Per Anandtech:

There's also overclocking to consider: excellent Broadwell-C examples can hit 4.2GHz, 4.3GHz if you're lucky in the silicon lottery and have excellent cooling. Skylake chips seem to comfortably hit 4.6GHz plus.

 

You're really grasping at straws here. Looking at only the bar graphs for games, the 5775C wins in 6 tests, loses in 4, and ties in 2. Of the tests it won, the greatest FPS difference was 6% faster, and in another test it had 7ms shorter frame time. That isn't an overwhelming win, especially when you consider the price. Meanwhile, you seem to intentionally ignore the significant losses in non-gaming tests.

Yes, the cache very well likely is the reason why this CPU can compete (in gaming tasks) with a newer one that is several hundred MHz faster, but the simple fact of the matter is, it is overall slower than the less-expensive 6700K and a lot less capable than the similarly priced 5820K.

 

I do think it's a bit of a shame that there aren't more Intel CPUs with the Iris Pro GPU, as it's pretty tasty.

 

The problem is (to my knowledge) Intel won't put it in i5 models or lower - that's where people would need it most. Most desktop i7 users don't need the IGP at all. If you're a gamer or graphic designer, it's too underpowered, relative to how good the CPU is. If you're just an office user, Iris Pro is really overkill, unless you have 3+ monitors. But if you had an i3 with an Iris Pro as capable as an i7's, that would make a fantastic energy efficient and cost effective rig to play modern games at medium settings.

 

I thought Zen was released this month? any news?

 

Thanks Gareth, I'm holding off building a new system waiting for this, will it be worth it? To give an idea, I am looking at 6700K, GTX1070 16GB Ram as the base, can Zen provide the performance.

 

Just about nobody knows. There are a couple leaked benchmarks here and there but we don't even know the frequency of these new CPUs, let alone their names. Unless you talk to one of the AMD engineers directly, nobody can tell you how good it is.
But, what I can tell you is if you already own a 6700K, then Zen will absolutely not be worth the upgrade. Even if there will be a Zen part that is better than Skylake in terms of performance-per-watt or instructions-per-clock, it won't be enough of a difference.

Right now, Excavator on DDR3 is roughly the same performance as Sandy Bridge. Zen, according to AMD, is supposed to be 40% faster than Excavator clock-per-clock. This is assuming a cherry-picked result of their best-case scenario. So realistically it's probably more like 20% faster on average. I did the math elsewhere (and I can't remember exactly how I put it together), but that puts it very slightly behind Skylake.


However, if you don't currently own a Haswell or newer CPU, Zen could very well be worth the wait.

 

I don't have a 6700k, this was my line of thinking for my major system overhaul - currently sandy bridge i5K @ 4.8ghz, 16GB 1866mhz DDR3 and 4gb 770GTX

But reassuring to know that with AMD, there's potentially a 20% increase over what I have and/or price drops on current Intel line-up, well worth waiting for, as Gareth says, I can't really lose by being a little more patient.

 

Depends on usage, Gaming your wasting money. Sandy at 4.8ghz in games is not far behind skylake at stock. Yes you can overclock but the gains vary.

1070 would likely double or tripple fps vs the 770.

Excavtor benches close in 1-2 benches but is a lifetime behind once Overclocked.

That's why clock for clock are never good comparisons. If your chip only does 3ghz but your rivals does 5ghz the gap could still be massive

 

Thanks Guys,

Upgrading from my current rig might also reduce power usage.

I recall reading somewhere that a CPU upgrade might also mean a PSU upgrade to be compatible (currently a 5yr old corsair 850w modular)?

The 1070, it's still expensive, but there's no alternative at that performance point, so I'll bite the bullet (I set myself a target buy price of £350 but not sure it will reach that).

I realise in many FPS games that 770-1070 will be a huge increase in performance - However, I do use my PC for other things so an upgrade to the core for me feels well overdue, I've never had a CPU/GPU/MOBO for as long as I have had this one.