August 29, 2013, 1:44 PM — Advanced Micro Devices (AMD) has gained and then given up a lot of ground to Intel in recent years because it can't keep up with Intel in the engineering race, and if the rumors are true, it's going to go slow in the future as well, at the risk of even more market share.
That's if the rumors are true. They come via DigiTimes, a Taiwanese publication that's got a hit and miss track record.
For example, DigiTimes claims Kabini, an APU is aimed at the low-power/subnotebook/netbook/ultra-thin form factor markets, would enter the market in the first quarter of 2014. But AMD told me it's already shipping Kabini.
Kabini is based on the Jaguar architecture and a SoC design of Jaguar, featuring a 8-core CPU, GPU and other integrated electronics will be used in both the Xbox One and PlayStation 4, and those are shipping this fall.
The article also discussed Kaveri, a high-end processor with four cores and a Graphics Core Next (GCN) design that will be seen for the first time in Kaveri. This GPU can access cache data from the system memory and also reference the data from CPU's cache, making for the most tightly integrated CPU/GPU union yet. It also supports the Heterogeneous System Architecture and comes with an ARM Cortex-A5 CPU as well all integrated into the chip.
DigiTimes said it was due for late 2013 but now will come out in March, 2014. However, AMD says Kaveri was always on track for a 2013 release. So this means Kaveri's successor, Carrizo, won't emerge towards the middle 2015.
This CPU will feature a whole new core based on the "Excavator" architecture, which replaces "Steamroller," the current CPU generation and the one on which Kaveri is based. Here's what is known about Excavator:
-- It introduces the "Clustered Integer Core" micro-architecture, original developed by DEC in 1996 with the Alpha 21264 RISC microprocessor. For the unfamiliar, former AMD CEO Dirk Meyer was previously a chip designer at DEC and developed the Alpha in the early 1990s. He then moved to AMD where he built a new x86 core later known as Athlon that had a lot of common technology with Alpha, including sharing the same socket. So there is a history here.
The Clustered Integer Core uses something called Clustered Multi-Thread (CMT), where instead of cores you have "modules." Each module has two out-of-order x86 processing cores sharing the early pipeline stages, an FPU and the L2 cache. This means dedicated hardware for two threads instead of one at a time and hopefully an increase in speed, especially in some multi-threaded integer cases.
-- Two symmetrical 128-bit FMAC (fused multiply–add capability) floating-point pipelines per module that can be unified into one large 256-bit-wide unit for AVX instructions and two symmetrical 128-bit pipelines for x87/MMX/SSE instructions.
-- All "modules" present share the L3 cache as well as an Advanced Dual-Channel Memory Sub-System (IMC - Integrated Memory Controller).
-- DDR4 supported.
The best rumor is that Carrizo will be fully compatible with the motherboards launched for Kaveri, which will use the FM2+ socket and A88X chipsets. It will also support the HSA platform. One thing you have to give AMD credit for is it does a much better job of preserving socket compatibility over multiple processor generations than Intel has ever done.
So what's the takeaway? AMD is taking it slow again and possibly ceding more ground to Intel. As it is now, AMD's best can't compete with Haswell and are barely competitive with Ivy Bridge, as benchmarks from Tom's Hardware Guide show.
AMD is well ahead of Intel in the CPU/GPU process. They've been doing it longer and have a vastly superior GPU to work with in the form of AMD's technology. But it's all for naught if you are routinely late and fall short on performance.
I have a briefing with them planned and one thing I want an answer on is the effort behind the HSA push. HSA is a system architecture that allows accelerators like the GPU to operate at the processing level as the system's CPU. This means those processors don't need separate memory or scheduling. It all sounds nice but I wonder what the payoff is, because AMD needs one. It's got a great SoC business and stands to make