Report: 'Crystal Well' Haswell IGP to Have 64MB L4 Cache
The "Crystal Well" variant of Intel's upcoming Haswell lineup reportedly features a 64 MB L4 cache dedicated for graphics.
According to a report published by Fuzdilla, the "Crystal Well" variant of Intel's Haswell range of processors will feature a massive 64 MB cache dedicated for graphics. Though the presence of L4 caches on graphics cards is nothing new, it has yet to make an appearance on CPUs since it would produce a "huge chip" due to the high transistor count of GPU cache memory.
While Intel certainly has the ability to fabricate processors with die sizes large enough to accommodate a large L4 cache, the size of this cache merits a degree of skepticism. Fuzdilla has also received reports that suggest that L1, L2, L3 and L4 memory will instead be "shared between the CPU and GPU."
From what we know, Crystal Well technology has been reserved for Intel's high-end GTS-based processors, so it is likely to only make an appearance on the most expensive Core i7 models. Regardless of whether these reports are true, we can safely expect a large improvement in Intel's onboard graphics when Haswell is launched later this year.

L1 cache isn't even shared between cores, let alone between the CPU and GPU. I call BS.
2. However, it makes sense to put a 64MB cache in the mobile chips for ultrabooks where the option for dedicated graphics isn't possible due to reduced battery capacity, heat, and space on the motherboard.
64MB sounds expensive but it may be necessary to provide enough bandwidth to keep the GPU from starving the CPU. Building it into the die using 22 NM may make it cheap enough to provide enough bang for the price.
L1 cache isn't even shared between cores, let alone between the CPU and GPU. I call BS.
2. However, it makes sense to put a 64MB cache in the mobile chips for ultrabooks where the option for dedicated graphics isn't possible due to reduced battery capacity, heat, and space on the motherboard.
64MB sounds expensive but it may be necessary to provide enough bandwidth to keep the GPU from starving the CPU. Building it into the die using 22 NM may make it cheap enough to provide enough bang for the price.
Not if the L4 is on a separate die from the rest of the CPU.
With the 386, 486 and Pentiums, L2 cache used to be on the motherboard. With the Pentium Pro, Pentium 2 and early Pentium 3, the L2 cache used to reside on separate chips on the CPU package.
There is plenty of precedent for on-package/off-die cache.
Also, a 64MB SRAM is more than 3.2 billion transistors which would make this cache larger than the whole CPU die which does not make much sense cost-wise. Logically, this would indicate that the cache is DRAM-based to keep surface area and cost in check. Since DRAM and high-speed CMOS processes do not play well together, this would also point towards an off-die DRAM chip.
So my bet is custom on-package DRAM chip.
L1 cache isn't even shared between cores, let alone between the CPU and GPU. I call BS.
I'm skeptical about L1, L2 being shared, L3 and L4 might, though...but i think it's more likely that only L4 is shared, or that it's exclusive to graphics.
Yeah, but if this gives you GT650m performance, while consuming far less power, you could actually game on the go, without plugging it in. Not intended to compete with 660m and above, though, i would guess. Might be able to undercut the discrete mobile market if they're smart.
2. However, it makes sense to put a 64MB cache in the mobile chips for ultrabooks where the option for dedicated graphics isn't possible due to reduced battery capacity, heat, and space on the motherboard.
64MB sounds expensive but it may be necessary to provide enough bandwidth to keep the GPU from starving the CPU. Building it into the die using 22 NM may make it cheap enough to provide enough bang for the price.
1. GT3e is for notebooks and -R series BGA parts only.
2. Agreed, but this isn't for ultrabooks afaik.
3. That's exactly why they're doing it, because making it shared GDDR would induce latency issues for the CPU. Best of both worlds.
It's for notebooks and BGA desktop parts (read: AiOs) only.
Not if the L4 is on a separate die from the rest of the CPU.
With the 386, 486 and Pentiums, L2 cache used to be on the motherboard. With the Pentium Pro, Pentium 2 and early Pentium 3, the L2 cache used to reside on separate chips on the CPU package.
There is plenty of precedent for on-package/off-die cache.
Also, a 64MB SRAM is more than 3.2 billion transistors which would make this cache larger than the whole CPU die which does not make much sense cost-wise. Logically, this would indicate that the cache is DRAM-based to keep surface area and cost in check. Since DRAM and high-speed CMOS processes do not play well together, this would also point towards an off-die DRAM chip.
So my bet is custom on-package DRAM chip.
You're pretty much dead on, check this out:
http://www.anandtech.com/show/6892/haswell-gt3e-pictured-coming-to-desktops-rsku-notebooks