Follow us on Twitter
Follow us on FacebookWhy Ivy Bridge Is Still Quad-Core
@ 2011/12/06A couple days ago, we published our Ivy Bridge Desktop Lineup Overview in which we mentioned that Ivy Bridge will remain a quad-core solution. There are dozens of forum posts with people asking why there's no hex-core Ivy Bridge, so now seems like a good time to address the question. Fundamentally, Ivy Bridge is a die shrink of Sandy Bridge (a "tick" in Intel's world), and that usually means either the core count or frequency is increased due to the lower power consumption of the smaller process node. Thus, instead of hex-core, we get a chip that looks much the same as a year-old Sandy Bridge, only with improved efficiency and some other moderate tweaks to the design. Let's go through some of the elements that influence the design of a new processor, and when we're done we will have hopefully clarified why Ivy Bridge remains a quad-core solution.
Marketing
If we look at the situation from the marketing standpoint first, having a hex-core Ivy Bridge die would more or less kill the just released Sandy Bridge E. Sure, IVB is about five months away, but I doubt Intel wants to relive the Sandy Bridge vs. Nehalem (i7-9xx) situation--even Bloomfield vs. Lynnfield was quite bad. If Intel created a hex-core IVB die, they would have to also substantially cut the prices of SNB-E. The current cheapest hex-core SNB-E is $555, while IVB hex-core would most likely be priced at $300~$400 since it's aimed at the mainstream; otherwise very few SNB-E systems would be sold. Even then, most consumers would opt for the IVB platform due to cheaper motherboard costs and lower TDP. PCIe 3.0 should also make 16 lanes fine for dual-GPU setups, reducing the market for SNB-E even more.
Differentiating the lineup by keeping Ivy Bridge quad-core allows some market for SNB-E among enthusiast consumers. Ivy Bridge E isn't coming before H2 2012 anyway so SNB-E must please the high-end until IVB-E hits. In the end, we still recommend SNB-E primarily for servers and workstations where the extra memory channels, PCIe lanes, and dual-socket support are more important, but the lack of hex-core IVB parts at least gives the platform a bit more of an advantage.
Evolution from traditional CPU to SoC
There are more than just marketing reasons, though. If we look at the following die shots, we can see that CPUs are becoming increasing similar to SoCs.
These four (well, techically three because Kentsfield consists of two dual-core Conroe dies) chips are the only "real" quad-core CPUs from Intel. There are quad-core Gulftown Xeons, and there will soon be quad-core SNB-E CPUs, but they all have more cores on the actual die; some of them have just been disabled. Comparing the die shots, we notice that our definition of CPU has changed a lot in only five years or so. Kentsfield is a traditional CPU, consisting of processing cores and L2 cache. In 2008, Nehalem moved the memory controller onto the CPU die. In 2009, Lynnfield brought on-die PCIe controller, which allowed Intel to get rid of the Northbridge-Southbridge combination and replace it with their Platform Controller Hub. A year and a half later, Westmere (e.g. Arrandale and Clarkdale) brought us on-package graphics--note that it was on-package, not on-die as the GPU was on a separate die. It wasn't until Sandy Bridge that we got on-die graphics. The SNB graphics occupy roughly 25% of the total die area, or the space of three cores if you prefer to look at it that way, and IVB's graphics (a "tock" on the GPU side, as opposed to a "tick") will occupy even more space.
While we don't have a close-up die shot of Ivy Bridge (yet), we do know its approximate die size and the layout should be similar to the Sandy Bridge die as well. Anand estimated the die size to be around 162mm^2 for what appears to be the quad-core die (dual-core SNB with GT2 is 149mm^2, and even with the more complex IGP we wouldn't expect dual-core IVB to be larger). That's a 25% reduction in the die size when compared with quad-core SNB die (216mm^2). A 22nm quad-core SNB die would measure in at 102mm^2 with perfect scaling and assuming all the logic/architecture is the same; however, scaling is never perfect and we know there are a few new additions to IVB, so 162mm^2 for IVB die sounds right. Transistor wise, IVB counts in at around 1.4 billion, a 20.7% increase over quad-core SNB.
To the point, today's CPUs have much more than just CPU cores in them. We could easily have had a hex-core 32nm SNB die at the same die size if the graphics and memory controller were not on-die .We've actually got a pretty good reference point with SNB and Gulftown; accouting for the larger L3 cache and extra QPI link, Gulftown checks in at 240mm^2, though TDP is higher than SNB thanks to the extra cores. The same applies to Ivy Bridge. If Intel took away the graphics, or even kept the same die size as SNB, a hex-core would be more or less given. Instead, Intel has chosen to boost the graphics and decrease the die size.
Subjectively, this is not a bad decision. Intel needs to increase graphics performance, and will do just that in IVB. Intel's IGP solutions account for over 50% of the PC marketshare, yet the graphics are their Achilles' Heel. All modern laptops have integrated graphics (though many still opt to go discrete-only or use switchable graphics), and having more CPU cores isn't that useful if your system will be severely handicapped by a weak GPU. We've also shown in numerous articles how hex-core scaling over quad-core is largely unnecessary on desktop workloads (more on this below). Increasing the graphics' EU count and complexity while also adding CPU cores would have led to a larger than ideal die, not to mention the increased complexity and cost. Remember, Moore's Law was more an observation of the ideal size/complexity relationship of microprocessors rather than pure transistor count, and smaller die sizes generally improve yields in addition to being less expensive.
Marketing
If we look at the situation from the marketing standpoint first, having a hex-core Ivy Bridge die would more or less kill the just released Sandy Bridge E. Sure, IVB is about five months away, but I doubt Intel wants to relive the Sandy Bridge vs. Nehalem (i7-9xx) situation--even Bloomfield vs. Lynnfield was quite bad. If Intel created a hex-core IVB die, they would have to also substantially cut the prices of SNB-E. The current cheapest hex-core SNB-E is $555, while IVB hex-core would most likely be priced at $300~$400 since it's aimed at the mainstream; otherwise very few SNB-E systems would be sold. Even then, most consumers would opt for the IVB platform due to cheaper motherboard costs and lower TDP. PCIe 3.0 should also make 16 lanes fine for dual-GPU setups, reducing the market for SNB-E even more.
Differentiating the lineup by keeping Ivy Bridge quad-core allows some market for SNB-E among enthusiast consumers. Ivy Bridge E isn't coming before H2 2012 anyway so SNB-E must please the high-end until IVB-E hits. In the end, we still recommend SNB-E primarily for servers and workstations where the extra memory channels, PCIe lanes, and dual-socket support are more important, but the lack of hex-core IVB parts at least gives the platform a bit more of an advantage.
Evolution from traditional CPU to SoC
There are more than just marketing reasons, though. If we look at the following die shots, we can see that CPUs are becoming increasing similar to SoCs.
These four (well, techically three because Kentsfield consists of two dual-core Conroe dies) chips are the only "real" quad-core CPUs from Intel. There are quad-core Gulftown Xeons, and there will soon be quad-core SNB-E CPUs, but they all have more cores on the actual die; some of them have just been disabled. Comparing the die shots, we notice that our definition of CPU has changed a lot in only five years or so. Kentsfield is a traditional CPU, consisting of processing cores and L2 cache. In 2008, Nehalem moved the memory controller onto the CPU die. In 2009, Lynnfield brought on-die PCIe controller, which allowed Intel to get rid of the Northbridge-Southbridge combination and replace it with their Platform Controller Hub. A year and a half later, Westmere (e.g. Arrandale and Clarkdale) brought us on-package graphics--note that it was on-package, not on-die as the GPU was on a separate die. It wasn't until Sandy Bridge that we got on-die graphics. The SNB graphics occupy roughly 25% of the total die area, or the space of three cores if you prefer to look at it that way, and IVB's graphics (a "tock" on the GPU side, as opposed to a "tick") will occupy even more space.
While we don't have a close-up die shot of Ivy Bridge (yet), we do know its approximate die size and the layout should be similar to the Sandy Bridge die as well. Anand estimated the die size to be around 162mm^2 for what appears to be the quad-core die (dual-core SNB with GT2 is 149mm^2, and even with the more complex IGP we wouldn't expect dual-core IVB to be larger). That's a 25% reduction in the die size when compared with quad-core SNB die (216mm^2). A 22nm quad-core SNB die would measure in at 102mm^2 with perfect scaling and assuming all the logic/architecture is the same; however, scaling is never perfect and we know there are a few new additions to IVB, so 162mm^2 for IVB die sounds right. Transistor wise, IVB counts in at around 1.4 billion, a 20.7% increase over quad-core SNB.
To the point, today's CPUs have much more than just CPU cores in them. We could easily have had a hex-core 32nm SNB die at the same die size if the graphics and memory controller were not on-die .We've actually got a pretty good reference point with SNB and Gulftown; accouting for the larger L3 cache and extra QPI link, Gulftown checks in at 240mm^2, though TDP is higher than SNB thanks to the extra cores. The same applies to Ivy Bridge. If Intel took away the graphics, or even kept the same die size as SNB, a hex-core would be more or less given. Instead, Intel has chosen to boost the graphics and decrease the die size.
Subjectively, this is not a bad decision. Intel needs to increase graphics performance, and will do just that in IVB. Intel's IGP solutions account for over 50% of the PC marketshare, yet the graphics are their Achilles' Heel. All modern laptops have integrated graphics (though many still opt to go discrete-only or use switchable graphics), and having more CPU cores isn't that useful if your system will be severely handicapped by a weak GPU. We've also shown in numerous articles how hex-core scaling over quad-core is largely unnecessary on desktop workloads (more on this below). Increasing the graphics' EU count and complexity while also adding CPU cores would have led to a larger than ideal die, not to mention the increased complexity and cost. Remember, Moore's Law was more an observation of the ideal size/complexity relationship of microprocessors rather than pure transistor count, and smaller die sizes generally improve yields in addition to being less expensive.
