Current vs Broadwell vs Skylake

[nanogirl21]

Can someone please explain
exactly what the differences are between the current chip that's in the 13 inch rMBP, Broadwell, and Skylake chips. I don't know much about computer terms and parts. Sorry for my ignorance on the subject.

 

[crsh1976]

Can someone please explain
exactly what the differences are between the current chip that's in the 13 inch rMBP, Broadwell, and Skylake chips. I don't know much about computer terms and parts. Sorry for my ignorance on the subject.

Intel follows a tick-tock scheme, where the tick is a die shrink and design refinements, and tock typically introduces a new architecture and adds major features.

Broadwell (tick) is a die shrink of Haswell (move from 22 to 14 nm), Intel is adding some minor features but it's mainly an improved design of the Haswell chips - slightly better, cooler, faster, with slightly better on-chip graphics (to simplify).

Skylake is a new architecture (tock), it adds major features such as support for DDR4 memory, Thunderbolt 3.0, better graphics, PCIe 4.0 slots, more bandwidth/faster memory lanes, etc.

Cannonlake is Skylake's tick after that, the die will go down to 10 nm, etc.

 

[dusk007]

Broadwell is basically 20% faster clock speed. means 20% faster CPU though not really because Turbo is almost identical so for burst it feels no faster.
And the GPU is the next generation which should in theory be 20% faster at least. It is called Intel HD 5500 instead of 4400 as current. In reality it seems at times no faster at all and other times it is hard to tell if it isn't just newer drivers (software improvements that both benefit from if you'd retest the old 4400 with new software) which help results.

Skylake will offer some bigger GPU improvements because graphics desperately need DDR4. Otherwise it is supposed to be mostly a big improvement on the chipset side but the cpu itself will probably not be much faster. I think the quad cores in the 15" will benefit much more than the 13" dual cores which already have extremely low power chipsets.
It is always possible that skylake will bring quad cores to the 13" rMBP but I expected this to happen already and it hasn't so who knows. If it happens that is significant upgrade but Apple doesn't seem interested because they could already fit a low power quad core into the 13" if they wanted to.
Well four cores is better than two obviously. It is just more power efficient to run more cores at lower clock speeds than fewer cores at higher clock speeds since power draw grows exponentially. So you get more total maximum performance in four cores at 30W than in two.

 

[bradleyjx]

Can someone please explain
exactly what the differences are between the current chip that's in the 13 inch rMBP, Broadwell, and Skylake chips. I don't know much about computer terms and parts. Sorry for my ignorance on the subject.

Like most industries, Intel works to improve it's computer chips, and releases upgrades in approximately yearly cycles. These upgrades come in two forms:

New Architecture: where much of the design of a chip is overhauled.

Die shrink: where the size of the components that make up a chip shrink. A normal die shrink comes close to doubling the amount of components that can fit in a given surface area.

During both of these, general improvements happen, integrating new technologies in hardware which the chips interface with; architecture changes tend to have the more drastic improvements, (as they are architected specifically to handle the new challenges those new technologies brought) but die shrinks can also have new features.

The current rMBP has a chip from Intel's "Haswell" line of processors. These chips were a new architecture release, but are also getting long in the tooth: while there's been a "refresh" of the line, the processors have been in rMBPs for close to 18 months at this point.

"Broadwell" processors, like the similarity of the name implies, are based on the same architecture as Haswell, and represents a die shrink. Intel has been having major issues with this die shrink, however, as these chips are getting so tiny that there are both technical challenges (building the chips with current technology) and physical constraints (quantum effects are beginning to cause major problems) with the size of connections in this generation. Some Broadwell chips have been released, but no Apple device currently uses Broadwell chips.

"Skylake" processors are on the next new architecture, at the same die size as Broadwell. The question that's been going around recently is in how Skylake will be released. Intel indicates that it will be releasing it on schedule, possibly skipping Broadwell in some markets, (such as high-end laptops) but there is some concern that they may hold off on the release as a means to make more money off of the Broadwell release.

 

[whitedragon101]

Skylake will offer some bigger GPU improvements because graphics desperately need DDR4.

How do graphics desperately need DDR4? Could you explain?

Thanks

 

[yjchua95]

How do graphics desperately need DDR4? Could you explain?

Thanks

Integrated graphics get their VRAM from the computer's RAM, and the bandwidth of the iGPU is henceforth limited by the RAM's bandwidth.

DDR4 brings faster RAM and much larger bandwidth, and hence will improve iGPU performance.

 

[thekev]

Integrated graphics get their VRAM from the computer's RAM, and the bandwidth of the iGPU is henceforth limited by the RAM's bandwidth.

DDR4 brings faster RAM and much larger bandwidth, and hence will improve iGPU performance.

That's true, but it remains to be seen whether it's the biggest bottleneck to intel's integrated graphics. Filling gpu memory is also expensive on discrete graphics, but that doesn't make it always the primary issue. Having plenty of bandwidth definitely makes it easier to program things, but it doesn't state whether current solutions could run faster if they had higher bandwidth. There's also the issue of how close memory clock comes to a multiple of cpu clock with a tightly integrated design.

For the OP, if there are appropriate Broadwell chips, they'll be out this year. There were rumors of skylake this year, but intel wouldn't release chips designed around Broadwell then follow that up with chips designed to fit the same sockets and power specifications in a different architecture a few months later. Basically what you see at the next refresh is what you can actually count on for the year. Apple sometimes does mid-generation updates, but they are typically just a minute bump 8-10 months in if there won't be a major update for a while. The best use of those is to get a good price on a refurbished or closeout unit, because the difference isn't really meaningful.

 

[yjchua95]

That's true, but it remains to be seen whether it's the biggest bottleneck to intel's integrated graphics. Filling gpu memory is also expensive on discrete graphics, but that doesn't make it always the primary issue. Having plenty of bandwidth definitely makes it easier to program things, but it doesn't state whether current solutions could run faster if they had higher bandwidth. There's also the issue of how close memory clock comes to a multiple of cpu clock with a tightly integrated design.

Nah, the limiting factor is now more on Intel's 128MB eDRAM Crystal Well cache. Perhaps they should bump up that cache.

But AnandTech agrees that even the eDRAM cache's bandwidth is somewhat limiting, in their Iris Pro 5200 review.

 

[thekev]

Nah, the limiting factor is now more on Intel's 128MB eDRAM Crystal Well cache. Perhaps they should bump up that cache.

But AnandTech agrees that even the eDRAM cache's bandwidth is somewhat limiting, in their Iris Pro 5200 review.

That's a little disappointing. What I was getting at earlier is that I don't expect an enormous benefit to be realized from day 1, because the issue of memory bandwidth influences overall design. In anything sent to the gpu, the manner in which it is sent becomes a huge issue, because it involves expensive operations. It is worth noting that the last major update was in 2013, so we should be near the end of a cycle.

 

[leman]

Broadwell is basically 20% faster clock speed. means 20% faster CPU though not really because Turbo is almost identical so for burst it feels no faster.

Don't forget that Broadwell reaches the Turbo much quicker — its capable of rapid micro-bursts. Given that the common user-machine interaction involves very short action-reaction scenarios, these CPUs will most likely make the computer 'feel' faster by providing immediate power exactly when its required. That's why Core M CPUs perform quite well in browser benchmark tests despite being slower than current Haswell ones.

 

[dusk007]

How do graphics desperately need DDR4? Could you explain?
Thanks

CPU need specific data as fast as possible but not all that much.
GPUs need lots of data constantly but it usually knows which data beforehand so it just need to arrive immediately but at a constant flow to keep all the processing units always busy. Integrated GPUs as in the dual cores share the memory access and while it is enough for the CPUs the GPUs don't have enough. It is hard to make the GPU faster by adding more processing power if you don't get enough data there to process.
For the last generations integrated GPUs had to few bandwidth and worked faster every time memory speed improved.

The GPU in the 15" MBP quad core has the same 40 EUs(processing units) as the one in the 13". But it has the extra eDRAM with about 2 times the bandwidth of the RAM which takes a lot of load from the main memory access by caching up to 128mb in this eDRAM.
That GPU performs about 70% faster which clearly shows that bandwidth is needed. DDR4 will give some and it will allow integrated GPUs without such an expensive eDRAM to reach new performance levels.

Nvidia's Maxwell (their newest generation) got a lot of its speed and efficiency from figuring out a way to make the whole chip less bandwidth hungry by using a bigger 2MB cache. Not having to access data over the memory interface saves a lot of power (the further data travels the more power you need for the work) and it offers speed because mainstream mobile GPUs have for many years now been stuck at 128bit GDDR5. So it was basically an exercise of eking as much performance as possible out of a bandwidth starved situation. Nvidia is currently best at it.
But Intel and AMD all face the same problem. DDR4 helps quite a bit but the real deal will be memory call HBM (high bandwidth memory) which is stacked DRAM that offers much higher bandwidth at much lower power consumption. But until HBM shows up for CPU SoC like Intels it may still take some time, maybe the one past skylake. DDR4 helps somewhat but only HBM will actually deliver enough to make the bottleneck go away.