Intel's 128MB L4 Cache May Be Coming To Broadwell and Other Future CPUs 110
MojoKid writes "When Intel debuted Haswell this year, it launched its first mobile processor with a massive 128MB L4 cache. Dubbed "Crystal Well," this on-package (not on-die) pool of memory wasn't just a graphics frame buffer, but a giant pool of RAM for the entire core to utilize. The performance impact from doing so is significant, though the Haswell processors that utilize the L4 cache don't appear to account for very much of Intel's total CPU volume. Right now, the L4 cache pool is only available on mobile parts, but that could change next year. Apparently Broadwell-K will change that. The 14nm desktop chips aren't due until the tail end of next year but we should see a desktop refresh in the spring with a second-generation Haswell part. Still, it's a sign that Intel intends to integrate the large L4 as standard on a wider range of parts. Using EDRAM instead of SRAM allows Intel's architecture to dedicate just one transistor per cell instead of the 6T configurations commonly used for L1 or L2 cache. That means the memory isn't quite as fast but it saves an enormous amount of die space. At 1.6GHz, L4 latencies are 50-60ns which is significantly higher than the L3 but just half the speed of main memory."
So in the real world? (Score:1)
I have a Retina MacBook Pro with this Crystal Well processor. What advantages does it really bring?
Unsure of any real world benchmarks compared to standard Haswell processors.
Re:So in the real world? (Score:5, Informative)
The only benchmarks I have found is from SiSoftware. http://www.sisoftware.co.uk/?d=qa&f=mem_hsw [sisoftware.co.uk]
But how is this going to effect Firefox, Photoshop, or video conversion?
Does it have an effect on battery life?
Re:So in the real world? (Score:4, Interesting)
Re:So in the real world? (Score:5, Informative)
Photoshop? Considering that the adobe rgb or other color spaces combined with the file sizes of some of the larger images coming out of cameras, your gains in latency would really depend on Photoshop and the OS being able to handle the L4 cache and keep the right part of the image in the cache. Video editing, with file sizes into the gigabyte range would probably see no gains at all. Video conversion, with a program that keeps a reasonably sized buffer, should see a good performance gain; but it would require code that knows the L4 is available or the OS to predict that L4 is a good place to put a 10-50-100MB buffer. The real gain will be in common things: playing a video, browsing the web (seen how much memory a bit of javascript or the JRE can eat up lately? Or Silverlight/Flash?) and email clients (cache all your email in L4 for faster searching).
As for battery life, I have no idea. It might use more power, since DRAM requires constant power to refresh data where SRAM is pretty stable; but the lower leakage of using a single transistor instead of 6 might prove to be a benefit. It would take a good bit of time and some pretty good test code to figure the difference, I suspect.
Re: (Score:3)
I'm not an expert by a long shot, but I'm pretty sure that modern day applications don't go anywhere near that low a level and instead leave memory management up to the system.
Re:So in the real world? (Score:4, Informative)
Yes and no. Applications can't typically "put things into the cache", but algorithms can (and often are, when it comes to image processing) tuned to suit a particular cache size. Processing the image in an appropriate order, breaking the image into cache-sized chunks, and so on can all be effective strategies which pay off big-time in terms of performance.
Re: (Score:3)
Even if the whole files take up more than the cache, the filters and algorithms running on them may need to access only a part of the image/video (e.g. a access a frame of the video multiple times). The benefit of caching is highly dependent on the algorithms used
Re: (Score:3)
Re:So in the real world? (Score:5, Informative)
Given that even intel's marketing optimists don't say much about CPU performance (and also given that it's a mobile-only feature, you can't even buy an non-BGA part expensive enough to have it, which would be unusual if it actually improved CPU performance enough to get enthusiasts worked up; but is downright sensible if the target market is laptops sufficiently size/power constrained not to have discrete GPUs; but where pure shared memory was dragging GPU performance down.)
Re: (Score:2)
Re: (Score:2)
Interesting that the gigabit Ethernet controller on the latest Apple Mac's have 512MiB of DDR3. Any idea what this is for?
Re: (Score:3)
Re: (Score:2)
Actually, that amount on a NIC would be a great boon in keeping all network processing on the NIC instead of having to CPU/system memory-offload, especially when you turn on the bells and whistles like jumbo frames, etc. I can also see it helping out quite a bit when processing HD video packets when streaming video where it's pretty important to get them processed as quickly and efficiently as possible before passing them off to the main system. These packets tend to have a decent amount of overhead, etc an
Re: (Score:2)
The only sign, unless you delve into the part numbering alphabet soup, that you even have it, is a change in the designation of the graphics "Iris Pro 5200" vs "Iris Pro 5100", and it's only available in the highest-price laptop parts. I have no reason to suspect that it'll hurt performance on the CP
Re:So in the real world? (Score:5, Interesting)
Cache performance impact is very heavily dependant upon application characteristics. Specifically, active memory.
Best case, when you're working with an active set that's larger than L3 but under L4 - around 100MB or so - and you're accessing it on a repeating pattern, and the compiler hasn't found any tweaks to help, and you're not multitasking, and the OS isn't swapping you out every slice, and the stars are aligned in your favor... the theoretical maximum performance gain can be up to 2x. It's very rare you'll find a program that benefits that much, though. Closest I can think of is image processing.
So in the real world, anywhere from 'no benefit' to 'double the speed' depending on application.
Re: (Score:3)
Closest I can think of is image processing.
What you've quoted sounds more like a case for random accesses. Trees, graphs (!), and other complicated data structures, I'd guess. I believe that image processing can take care of itself most of the time by simple prefetching.
Re:So in the real world? (Score:4, Informative)
I have a Retina MacBook Pro with this Crystal Well processor. What advantages does it really bring?
Unsure of any real world benchmarks compared to standard Haswell processors.
I've written papers on the effect however I am unable to share them here. The bottom line is the application should be exposed to reduced minor page faulting and, if all goes well, improved context switching, all dependent on the way the CPU scheduler is configured - of course.
IMHO an L4 cache will alleviate the cache miss penalty when the CPU Scheduler looks for data in L1-3 however any increase in the penalty due to a cache miss will be highly dependent on the application and the way the CPU scheduler is configured.
The idea is to try and keep the L1-3 Cache as hot as possible, really it's because as programmers, many of us still have a long way to go to writing code that scales to parallel processing well (in the 21st century!!!) plus there is a lot of code out there already.
For Linux and Apple based systems (I can examine the code of these CPU Schedulers - just not the Microsoft as it is proprietary) this should mean that the amount of time the CPU spends on application tasks, as opposed to O.S tasks is increased, essentially boiling down to reduced application latency and improved "responsiveness". I don't mean to use such wishy-washy terms however at this level cpu instructions are carried out in the nano-femto second range and the duration imposed by a cache miss penalty and a context switch will also be dependent on the ram installed - which is another factor in the duration of a minor page fault.
Assuming that the schedulers, in a "fair and balanced" configuration I expect the following. For code that scales to parallelism you should see improvements because a task will exist on multiple cores well and not incur penalties for hogging CPU resulting in the L1-3 caches staying hot with application data longer (ideally, with threads running on multiple cores). For code that doesn't I expect it to hog a core, get pushed back to ram by the scheduler and be exposed to all of the performance penalties that come as a result.
Personally I have always thought it's a contest between Cycles and Cache - not a direct effect on battery life or power consumption however if the CPU is spending more time on application than OS then you are closer what the original Amdahl's law [wikipedia.org] sought to show - if your application allows it.
Re:first post (Score:4, Informative)
Seems other users have a bigger cache than yours...
Re: (Score:2)
I knew I should have waited for 129mb cache
Re: (Score:2)
0.128 bits does tend to be smaller than most caches.
If you screw with Broadwell (Score:2)
Just say no, David.
"half the speed of main memory"? (Score:3, Insightful)
"At 1.6GHz, L4 latencies are 50-60ns which is significantly higher than the L3 but just half the speed of main memory."
WTF? The correct would be, I think, half the latency of main memory...
Re: (Score:1)
Re: (Score:2)
It's also pretty poor, for a cache. That's why it's an L4 cache, rather than replacing the L3 or L2.
Half the speed of main memory? Why Bother? (Score:3, Informative)
At 1.6GHz, L4 latencies are 50-60ns which is significantly higher than the L3 but just half the speed of main memory.
Hmmm. L4 cache runs at half the speed of main memory? That doesn't seem right Why bother reading these summaries? The people posting them certainly don't
Re: (Score:1)
Umm, they are clearly using latency as their measure of speed, so yes, 'half the speed of main memory" does seem right. Sure it's not worded as well as it could be, but you should be able to understand and not bitch about it.
Re: (Score:1)
Re: (Score:2)
Oh please! It should be twice the speed if it has half the latency, not half the speed. Speed and latency are related, but not interchangeable synonyms!
If a cache has "half the speed" of your uncached memory, you need to disable that cache ASAP!
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
I suspect they meant half the latency
Re: (Score:2)
DDR3 is significantly faster than that.
I can't find any actual measurements with a quick web search, but I seem to remember that core to DDR3 latency is in the order of 80-100ns.
Re: (Score:3)
DDR3 (dram in general) can burst quickly but full random accesses are very slow in comparison.
-Matt
Re: (Score:2)
Why bother reading these summaries?
It's a puzzle. Is the summary wrong because of stupidity, or is it crafted that way for click bait?
Re: (Score:1)
Actually it is the other way round: The slower the CPU, the faster the memory interface in comparison and the less need for caches. What Intel does here makes no sense unless they are covering up an architectural problem. Memory clock is _not_ tied to CPU clock in a sane architecture. Otherwise you would need to buy memory by CPU clock. You do not need to to that for Intel or AMD.
Also, desperation engineering-wise has not necessarily any connection to business-desperation. And remember that Intel messed up
Why only 128 MB? (Score:1)
Broadwell represents a miniaturization step from 22 to 14 nm structures. Why do they keep the capacity of the Crystalwell L4 cache at 128 MB? They could put twice that memory onto a die with the same area as the 22 nm Crystalwell version. Is the Crystalwell die for the Haswell CPUs so large and expensive that they have to reduce its size?
It's not 14nm, that's why (Score:2)
Re:Why only 128 MB? (Score:5, Informative)
Broadwell represents a miniaturization step from 22 to 14 nm structures. Why do they keep the capacity of the Crystalwell L4 cache at 128 MB? They could put twice that memory onto a die with the same area as the 22 nm Crystalwell version. Is the Crystalwell die for the Haswell CPUs so large and expensive that they have to reduce its size?
From Anandtech's article on Crystalwell [anandtech.com]:
There's only a single size of eDRAM offered this generation: 128MB. Since it's a cache and not a buffer (and a giant one at that), Intel found that hit rate rarely dropped below 95%. It turns out that for current workloads, Intel didn't see much benefit beyond a 32MB eDRAM however it wanted the design to be future proof. Intel doubled the size to deal with any increases in game complexity, and doubled it again just to be sure. I believe the exact wording Intel's Tom Piazza used during his explanation of why 128MB was "go big or go home". It's very rare that we see Intel be so liberal with die area, which makes me think this 128MB design is going to stick around for a while.
I get the impression that the plan might be to keep the eDRAM on a n-1 process going forward. When Intel moves to 14nm with Broadwell, it's entirely possible that Crystalwell will remain at 22nm. Doing so would help Intel put older fabs to use, especially if there's no need for a near term increase in eDRAM size. I asked about the potential to integrate eDRAM on-die, but was told that it's far too early for that discussion. Given the size of the 128MB eDRAM on 22nm (~84mm^2), I can understand why. Intel did float an interesting idea by me though. In the future it could integrate 16 - 32MB of eDRAM on-die for specific use cases (e.g. storing the frame buffer).
Win95? (Score:4, Interesting)
With this 128MB cache, shouldn't this CPU be able to run an OS like Win95 of an older Linux without additional memory?
Re: (Score:3)
Re: (Score:3)
Re:Win95? (Score:4, Interesting)
You're right that motherboards won't post without memory sticks, but I don't see a good technical reason about why that should be. UEFI could be written so that it posts by using only the resources of the processor and its cache, if it detects no usable memory. I mean, never mind 128MB of L4. Even the 6MB of L3 that modern processors have is larger than the entire system memory of our parents' first computers. It should be more than enough to run something as simple as UEFI.
It would also be rather useful. Instead of issuing you beeps as it fails to boot, a motherboard with a correctly written UEFI implementation could post without working ram and run diagnostics on exactly which systems are working and which are not, and what exactly is going wrong. I really think this would increase everyone's system-building confidence and give the manufacturers who make it happen a leg up in the market.
Classic game consoles as a comparison point (Score:3)
Even the 6MB of L3 that modern processors have is larger than the entire system memory of our parents' first computers.
A 6 MB L3 cache is bigger than the RAM in the PlayStation, Nintendo 64, or Nintendo DS. A 128 MB L4 cache would surpass the RAM in a PlayStation 2 and an original Xbox combined. You don't need a lot of DDR [wikipedia.org] to play DDR [wikipedia.org], even if you live in the former DDR [wikipedia.org].
Re: (Score:2)
I'm an Apple user and have only ever really looked in to UFI on a Mac, UFI has its own memory space, current Macs have 512MiB of DDR3 separate to the main memory. Also have 512MiB of DDR3 for networking too.
Re: (Score:2)
Re: (Score:2)
Re: (Score:1)
Maybe they will get it right next time with the L5 cache.
L4 used to be main RAM (Score:2)
Re: (Score:2)
But the answer is no. The OS just wasn't designed to use that function of the processor in such way. Maybe if you wrote a VM that emulated RAM on the L4 cache, but the only purpose of this approach would be satisfy the curiosity.
Windows XP on 128 MB of RAM (Score:3)
Re: (Score:2)
Re: (Score:2)
Yeah*, but what's the point?
*Assuming the OS doesn't freak out - which will definitely happen. Let's just say there's no technical barrier to overcome.
Re: (Score:1)
Depends on what you're doing with it. I have a laptop (Pentium 3 / 128Mb RAM) with FreeBSD 10 on it. It works well but the application options running in X are limited unless you want to go into swap. A huge portion of what people consider regular computer usuage is "browse the internet". Good luck doing that these days with 128Mb RAM.
Re:Ours goes to 11 (Score:5, Insightful)
Let's see, the tiny amount of L1/2/3 cache currently is dictated by the energy budget of the CPU. Looking at the energy budget of the 4900MQ and the 4960HQ chips, you can take some wild arse guessing to get that the 2 megs of L3 cache sacrificed got back enough to power the 128 megs of L4. Then consider that there is only 64K (yes, kilobytes) of L1 or 256K L2 per core on the Haswell chips, and at 3.9GHz desktop chips you are looking at 84 watts of power dissipated . . . you can start to work out how much of that is due to leakage current from the 6 transistor L1/2/3 cache design.
Let's face it, SRAM isn't tiny, it leaks amps like a sieve at the tiny process size that everything is done at now days, and it's main advantage is that it doesn't take a controller to access and it's bloody fast and the bandwidth can be pretty sizable. A gig of SRAM on die would, I suspect, heat a small room; that much DRAM per core would slow the cores down due to the inherent latency of accessing DRAM.
So, sure, DRAM chips may be cheap, but putting them on the CPU die would be horrid. And SRAM still isn't cheap; either in die space, energy budget, or dollars!
Re: (Score:2)
Obviously, what we need is some XRAM.
1T-SRAM (Score:2)
Let's face it, SRAM isn't tiny, it leaks amps like a sieve at the tiny process size that everything is done at now days, and it's main advantage is that it doesn't take a controller to access and it's bloody fast and the bandwidth can be pretty sizable.
Then perhaps MoSys had the right idea: make a bunch of small, independent DRAM blocks and a front-end controller with as much SRAM as one block to hold cached results while waiting for the corresponding DRAM row to refresh.
128 MB L4 cache (Score:2)
This is making me feel old as I recall how happy I was to have once maxed a board with 32 MB of RAM, a previous one with 8 MB, another with 4 MB and so on. I love that about technology, it pretty much always gets better until DRM and politics get into the mix...
Re:128 MB L4 cache (Score:4, Interesting)
you can revisit those those nostalgic 8MB and 4MB days again with the latest AMD chips [wikipedia.org] as L2 cache. :)
just use a modifide version coreboot [coreboot.org] to bypass those silly POST tests and load to the CPU cache directly with Windows 3.11 :)
Re: (Score:3)
I got the same feeling when I got my first Android phone. 576MB of RAM...in a phone. I've recently upgraded and my new device has 3GB of RAM. It feels like only recently that I hit that amount in a desktop computer, now I have it in a device that fits in my pocket - never mind the quad-core CPU or 64GB of internal storage.
10 years ago, that would have been a reasonably powerful desktop machine.
Re: (Score:3)
Re: (Score:2)
It would still be a "reasonably powerful desktop machine" if your use case is still the same as 10 years ago.
However, contrary to what many geeks think, people don't just browse, do email, watch youtube etc.. A fair amount of non-geeks do CAD, image/video editing, 3D graphics, create music etc with their desktop machines, and routinely have workloads that would bring that 10 year old computer into thrashing hell...
In fact, I think the whole "oh, ordinary people just need enough power to browse, email etc" i
Re: (Score:3)
It would still be a "reasonably powerful desktop machine" if your use case is still the same as 10 years ago.
Yeah, I was a hard- and software developer 10 years ago, my use case didn't shift too much. OTOH, I happen to /do/ CAD on this nearly 10 year old computer (single-core and all that, can you believe it?), so my information is first-hand.
[meaningless windows-centric gibberish excusing bloatware]
Whatever.
Re: (Score:2)
Re: (Score:3)
This is making me feel old
If your youthful recollections are about memory measured in MEGAbytes, then you are not old. Back in the 1970s, I worked on a controller board with a Z80, 256 bytes of ROM and ZERO RAM. All the state information had to be kept in registers (but, fortunately, a Z80 has two register banks). No RAM means no stack, so to call a subroutine, you had to save the return address in a register, so subroutines couldn't nest. As I recall, it just had to monitor a voltage and dial a phone number if it dropped too lo
not on die (Score:5, Informative)
128MB L4 cache. [...] on-package (not on-die) pool of memory
what this means is the memory is not on the same piece of silicon as the CPU, just stuffed in the same chip package. this means they have to be connected by a lot of tiny wires instead of being integrated directly. the downside to this is that there is bandwidth between the L4 memory and the CPU is very limited and it uses more power. like AMD's first APUs where just two ICs on the same chip, i dont not think this will result in a drastic performance improvement but i'm unsure of the power savings. If AMD gets wise, they will beat Intel to the punch but then again. though if AMD is really smart, they would put out ARMv8 chips not just for servers(/desktops?) but for smartphones/tablets and laptops.
Re:not on die (Score:5, Informative)
what this means is the memory is not on the same piece of silicon as the CPU, just stuffed in the same chip package.
Which allows the designers to count on carefully controlled impedances, timings, seriously optimized bus widths and state machines, and all the other goodies that come with access to internal structures not otherwise available.
Such a resource could, if used properly, be a significant contributor to performance competitiveness.
Re: (Score:1)
Re: (Score:2)
Re: (Score:3)
Round trip time for an old school EDO DIMM is not the same thing as the burst cycle time of a synchronous dram. It take about the same time to get going, but the data bursts faster and wider.
Oh noes (Score:1)
All my algorithm development so far assumes small local caches.
Now I can start over again.
Aaaahhh!!!
Trade off makes sense.. (Score:1)
Don't you mean... (Score:2)
At 1.6GHz, L4 latencies are 50-60ns which is significantly higher than the L3 but just half the speed of main memory.
Don't you mean "but less than half the latency of main memory?"
Cool (Score:2)
I do look forward to this.
Re: (Score:2)
All these are cheap and worthless improvements! We need faster CPUs - 8GHz for single-core this year and 16GHz next year!
eDRAM may be many things, but cheap isnt one of them.
Eh, it's been done. (Score:2, Informative)
POWER8, anyone? With actual SMT instead of flakey HT, and lots more threads, and so on, and so forth.
Too bad they're unobtanium and if not cost too much. But otherwise... anything intel does has basically been done better before. Except process. That is the only thing they really lead with. The rest isn't half as interesting as most of the world makes it out to be.
Pretty obvious evolution (Score:2)
With Intel's 14nm so close, and 10nm production in another year or so, they need to use all that chip area for something that doesn't necessarily generate a ton of heat. RAM is the perfect thing. Not only is the power consumption relatively disconnected from the size and density of the cache, but not having to go off-chip for a majority of memory operations means that the external dynamic ram can probably go into power savings mode for most of its life, reducing the overall power consumption of the device
This is insane (Score:2)
No one ever needs more than 640KB. :P