That link from Mar 2014 only mentions Z87 boards getting Haswell refresh not Broadwell update.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
The Intel Broadwell Core i7-5775C and i5-5765C Tested
- Thread starter HardOCP News
- Start date
That link from Mar 2014 only mentions Z87 boards getting Haswell refresh not Broadwell update.
Yep. I think it was the Anandtech article that said that 32 megabytes would be fine but they decided to double it, and then double that. It's just probably is one of those situations where it's so big because they needed the space for the interconnect so they filled it other than they just decided to be extremely generous.
defaultluser
[H]F Junkie
- Joined
- Jan 14, 2006
- Messages
- 14,398
Right, the reason you don't see much *consistent* improvement iwith the 128MB L4 cache is because:
They went with eDRAM for high density and low power compared to SRAM (which probably would, have limited them to 32MB in the same chip space). eDRAM, much like main memory has slow access times, but wide bandwidth. Because of their unique choice for L4 cache, it exhibits the following behaviors:
1. When the L4 cache hits, it's fast, but not massively faster than main memory like the L3 cache.
2. When L4 cache misses, due to that extra layer of cache now, the miss access time is significantly LONGER (it always takes time to check if data is in L4 cache). See here where worst-case access times rise by 10-15 clocks; ALL memory accesses take longer.
So if you don't get a high hitrate on the L4 cache, you don't see any benefit. High hitrate on large, slow cache like Crystalwell depends mostly on spacial locality: i.e. when we have a miss, the cache loads an entire line: that memory location plus X bytes in the same line into L4. This means simple contiguous data structures like textures (and large databases) benefit most from the eDRAM, while your average consumer application with more random data access patterns benefits less.
And then there's the fact that most applications people work with fit just fine into L3 cache. if you have consistent hits in L3, then you're not going to benefit from L4.
They went with eDRAM for high density and low power compared to SRAM (which probably would, have limited them to 32MB in the same chip space). eDRAM, much like main memory has slow access times, but wide bandwidth. Because of their unique choice for L4 cache, it exhibits the following behaviors:
1. When the L4 cache hits, it's fast, but not massively faster than main memory like the L3 cache.
2. When L4 cache misses, due to that extra layer of cache now, the miss access time is significantly LONGER (it always takes time to check if data is in L4 cache). See here where worst-case access times rise by 10-15 clocks; ALL memory accesses take longer.
So if you don't get a high hitrate on the L4 cache, you don't see any benefit. High hitrate on large, slow cache like Crystalwell depends mostly on spacial locality: i.e. when we have a miss, the cache loads an entire line: that memory location plus X bytes in the same line into L4. This means simple contiguous data structures like textures (and large databases) benefit most from the eDRAM, while your average consumer application with more random data access patterns benefits less.
And then there's the fact that most applications people work with fit just fine into L3 cache. if you have consistent hits in L3, then you're not going to benefit from L4.
Last edited:
Wiseguy2001
2[H]4U
- Joined
- Nov 28, 2001
- Messages
- 3,470
It's a shame the L4 cache only has 50GB/s bandwidth, which is very low for a GPU (~1/10 of AMD's HBM). Maybe things would have been more interesting if they halved the size and doubled the speeds.
Hopefully unified memory makes a bigger difference.
Hopefully unified memory makes a bigger difference.
Crystalwell's eDRAM has a bit more than 100GB/s bandwidth (~50GB/s bidirectional, a design choice favoring the type of work it's made for). Intel's integrated GPUs since Sandy Bridge have also used the higher bandwidth L3 cache (~180GB/s on Haswell) for graphics use. It's not "AMD's HBM". HBM a JEDEC standard and should be available to all foundry customers where it's available (Nvidia's Pascal uses it too).
Comparing it to multi-chip implementations of HBM for video cards is missing the point, both in what's feasible to put on a regular size microprocessor MCM and the segment integrated video targets. The JEDEC info I've seen has a single multi DRAM stack running at ~100GB/s bandwidth. Crystalwell has been shipping for 2 years vs we'll see about that stacked DRAM next year, maybe?
Wiseguy2001
2[H]4U
- Joined
- Nov 28, 2001
- Messages
- 3,470
No shit, I mentioned AMD because they're about to release a product with HBM; which is a similar release date to Broadwell.
The L4 cache isn't that fast, how much the performance benefit outweighs the overhead of another layer of cache is questionable. I was hoping for more.
The L4 cache isn't that fast, how much the performance benefit outweighs the overhead of another layer of cache is questionable. I was hoping for more.
Yeah, for a high end card, obviously not the same as an iGPU.No shit, I mentioned AMD because they're about to release a product with HBM; which is a similar release date to Broadwell.
The L4 cache isn't that fast, how much the performance benefit outweighs the overhead of another layer of cache is questionable. I was hoping for more.
Since there were apparently no performance regressions but only several benefits in the reviews of Broadwell I've read so far, it's not really a "questionable" choice*. If it's disappointing to you, then it is expectations that are out of line with reality. I could recommend one of a few good computer organization texts if you'd like to understand how a multi-level cached memory subsystem works.
The eDRAM is much faster and has higher bandwidth than main memory, and is larger than other caches, plus it offers large performance benefits to the GPU. It's just odd to be so negative towards it because it's obviously a benefit to the CPU.
* with a 95% L4 hit rate, the 5% of cache misses increasing by 10% in clock cycles is still a net benefit, by far
defaultluser
[H]F Junkie
- Joined
- Jan 14, 2006
- Messages
- 14,398
Yeah, not trying to downplay the benefits, just saying it's.a purpose built for large data sets
Unified memory will be slow.
This is the most exciting part to me. With DX12 we will all benefit from having a strong iGPU.
Michaelius
Supreme [H]ardness
- Joined
- Sep 8, 2003
- Messages
- 4,684
Yeah would be nice if iGPU could take a role of dedicated physX card but now with much wider support due to all those igpus put in cpus over last few years.
Why would it be slow? It's just mapping like what goes on for all other memory managed by the CPU, and is quite efficient in current processors.
Those are exactly all the channels a non-unified memory architecture would go through when you're talking about heterogeneous systems. Unified memory is at the very least a programming convenience, or possibly better when it's integrated with on-die LLC.
I know very well, its just that people are thinking its some users are thinking its much more than what it really is. Much like the person I quoted, using it in the same post as with L4 cache. Saying L4 cache was slow then talked about unified memory.
Complete review at computerbase.de
Wiseguy2001
2[H]4U
- Joined
- Nov 28, 2001
- Messages
- 3,470
What we're talking about here is not your typical CPU cache. Typically, cache is used to boost performance (and uses quite a bit of power). This isn't that; this 'cache' is used to boost power efficiently (making it so big means the DRAM controllers can go into sleep more), it offers very little performance* benefit over RAM (and adds another layer of cache to check).
Now, the reason I mention unified memory is because the memory on GPU's is sky-rocketing right now (GPU's have been stuck around 200GB/s for years, feels like a decade). This month, that's jumping to 512GB/s, nVidia's next gen release (with unified memory) it's doubling again to 1TB/s. Yeah, there is latency around PCIE, but it's still pretty fast (PCIE4.0 should be out), plus there's nvLink. If the [real] CPU cache can feed the CPUs then should should bring some decent benefit. I don't care how you cut it, 1TB/s of memory bandwidth on consumer hardware is going to be interesting!
*Performance is the combination of low latency and high bandwidth.
PS: Don't be a dick and try to put people down. We're all technology enthusiasts here (with different knowledge levels), let's play nicely.
The rest of your post wasn't really on the table but this last part is only in your head.
If you don't like people quoting you then don't post in forums.
ps. Don't take everything personal.
ccityinstaller
Supreme [H]ardness
- Joined
- Feb 23, 2007
- Messages
- 4,236
MrCaffeineX
[H]ard|Gawd
- Joined
- Aug 22, 2011
- Messages
- 1,604
Any idea when/if these things are actually going to be available at retail/e-tail? I haven't even seen anything up on Ebay yet...
These CPUs are available absolutely nowhere. I've checked dozens of sites and they all say that they will ship someday... but when? After the overall Broadwell delays due to their awful yield issues, this is not confidence-inspiring at all. I really want to build a totally fanless i7-5775C but if Intel delays any further they'll be shipping right ahead of the first Skylakes. However, it seems that the actual "enthusiast desktop" Skylakes won't be out until sometime in 2016 so I'm still interested in the i7-5775C... if I can get one!
So maybe the yield issues are not yet resolved? It's not like Intel to blow launch dates quite so badly.
It wouldn't surprise me with Iris and all that L4 cache. They took a very long time to delay broadwell, enough to not delay their next release.
I'm glad to hear that they are selling somewhere as that is a good sign that this may not be a vapor launch after all. I'm sure that if they actually are selling in Japan the forums will have someone who has actually managed to get their hands on a retail CPU.
The i7 5775C is really underwhelming as a new CPU but the reason I want it so badly is the Iris 6200. The goal is to have an entire system without a single moving part, and since I don't game, the 6200 is more than enough GPU for my needs. What I need is absolute silence... so come on, Intel, ship those LGA Broadwells here!
Well Skylake should be right around the corner, so ..
I'm not clear about this as it seems that I'm reading some reports that state that the LGA Skylakes won't be out first, as Intel will give priority to the mobile and NUC CPUs with the desktop silicon coming on later. I may be wrong about that but that's what I recall. Not only that, but I have DDR4 in my current 5960X system and in the type of applications I run I have to be honest that I really can't distinguish any difference to DDR3. I know the DDR4 advantages are not going to jump out and slap me across the face as they're primarily noticeable in benchmarks, but for day to day use at this point at least DDR4 seems like a waste of money. Furthermore, I think I'd like to wait at least 6 months for the LGA1151 platform and the new 100 series chipsets to go through a couple of revisions. I can't afford to be a guinea pig for Intel and since I have a very specific requirement for the Iris 6200 for a "no moving parts" system I will be happy to cough up the money for an i7 5775C... if they finally start shipping here at least!
I spoke to NCIX today and they said that they have absolutely no clue when the LGA Broadwells are arriving. They expected them three weeks ago but not a word from Intel. There has to be some reason for this late shipment and I'm surprised that the tech media isn't all over this. Is it possible that something happened between June 2 and now to wreck the launch altogether? Are there ever going to be any LGA Broadwells?
If it is true that the LGA Broadwells won't launch until late July then I just can't see LGA Skylakes showing up a month later. Neither Intel nor AMD (to my memory) have ever launched processor families that close. It makes no sense from a marketing standpoint unless they are trying to promote Broadwells as APUs and Skylakes as the next Haswells.
Maybe Intel is trying to decide if they'll use crappy TIM or solder?
My personal belief is that they have not overcome their 14nm yield issues and therefore both Broadwell and Skylake launches are sliding badly. But who knows what is going on behind the big blue facade?
Comixbooks
Fully [H]
- Joined
- Jun 7, 2008
- Messages
- 22,022
The lower power use at less clock speed is impressive to me anyway.
Can't wait to get my hands on one in December or sooner =)
http://www.maximumpc.com/intel-broadwell-dt-core-i7-5775c-review/
Can't wait to get my hands on one in December or sooner =)
http://www.maximumpc.com/intel-broadwell-dt-core-i7-5775c-review/
The thing about the date is that it's gone up a day each day it has been listed, it was the 23rd when I posted it here, yesterday it was the 24th and today it's the 25th, so it seems he has no idea when it will be available.
I spoke with a PC builder yesterday who is advertising i7 5775C systems as "shipping" but he admits he doesn't have anything and has no date for delivery. Just "someday."
Just a suggestion: Kyle could use his industry cachet to look into this and turn it into one helluva article that would get really wide exposure: "Why is Intel lying about Broadwell?"
Just a suggestion: Kyle could use his industry cachet to look into this and turn it into one helluva article that would get really wide exposure: "Why is Intel lying about Broadwell?"
Michaelius
Supreme [H]ardness
- Joined
- Sep 8, 2003
- Messages
- 4,684
Hmm our Polish shops show i5 OEM as available around 35 EUR more than i5 4690k.
Is that the 5675C? Is it actually shipping?