Trinity Improvements Include Updated Piledriver Cores and VLIW4 GPUs
Subject: Processors | May 4, 2012 - 06:11 PM | Tim Verry
Tagged: trinity, piledriver, llnao, fm2, APU, AMD A series
EXP Review has managed to get their hands on a set of AMD slides containing information on one of the company’s upcoming processor lines. The Llano successor, known as Trinity, is a new APU due out later this year that is said to bring increases in performance thanks to several architectural enhancements.
A Trinity APU die sitting next to a USB flash drive
Llano is AMD’s currently available Accelerated Processing Unit, or APU. The chips combine updated “Stars” mobile Phenom II CPU cores and Radeon 6000 series graphics cores into a single package. Further, the APUs contain a PCI-E 2.0 controller, integrated memory controller, and UVD3 hardware video decoding units. Some models also support AMD’s Turbo Core and Hybrid Graphics Technology which allow them to automatically boost CPU clockspeeds when lower GPU usage leaves TDP headroom, and to pair with a discrete Radeon HD 6450, 6570, or 6670 GPU in a Crossfire-like configuration. Built on a 32nm silicon on insulator (SOI) manufacturing process by GlobalFoundries, the APUs employ 1.45 billion transistors and have a die size of 228mm2 for the desktop versions. Desktop parts have TDPs of 65 watts or 100 watts depending on the particular chip and connect to the motherboards using the FM1 socket (which was a new socket for AMD, it has 905 contacts). There are both desktop and mobile Llano parts, though they are essentially the same chips. The mobile parts are scaled down desktop Llano chips that run at lower clockspeeds, top out DDR3 support at 1600MHz (versus DDR3 1866MHz on the desktop parts), have lower TDPs of either 35W or 45W, and use a slightly different socket (FS1).
In our review, and what many other users noted, is that Llano’s CPU performance really left something to be desired. Fortunately for AMD, the GPU portion of the chip delivered on performance and made the APU desirable for certain niches. The low power chips had a place in home theater PCs (HTPCs), cheap desktops, and even budget gaming rigs to an extent. Still, the CPU performance really held Llano back in terms of popularity and adoption among enthusiasts.
Llano APU in action during overclocking and gaming tests.
The upcoming Trinity processors bring quite a few enhancements to the table, foremost of which is a revamped CPU part that ditches the old Phenom II processor cores in favor of updated Piledriver architecture CPU modules. The move to the Piledriver x86 cores promises an increase in IPC, leakage reduction, CAC reduction, and increased clockspeeds according to the leaked slides, but the most important change is the increased performance per clock numbers. The Trinity APUs are set to replace the A8–or performance series of–Llano APUs with quad core Trinity processors that utilize two Piledriver modules that each share 2MB cache for 4MB of total L2 cache. In that respect, Trinity will be similar to Llano in that it does not employ any L3 cache that is shared between the CPU and GPU cores. Interestingly, that may mean that using higher clocked RAM can improve performance on Trinity just as it did with Llano. If true, that would make Trinity’s improved DDR3 support–up to DDR3 2133MHz– all the better. On the GPU side of things, Trinity moves to a “Northern Islands” VLIW4 architecture with up to 384 stream processing units. Although the GPU area is physically smaller, it is said to be more efficient than the GPU cores in Llano APUs. The new GPU core is DirectX 11 and OpenCL 1.1 compliant. Also, it includes an updated hardware tessellator engine and hardware encoding unit (AMD Accelerated Video Converter).
Trinity will continue to offer 65W and 100W TDPs as well as a 35W part. The TDPs are the same as those in Llano, but AMD has managed to lower the voltages needed to run Trinity out of the box. Also, AMD is claiming the new Trinity chips will sip power at idle–as low as 1.08 watts.
Trinity also ratchets up the automatic overclocking with Turbo Core 3 support which can boost the CPU clockspeed up to 19% or the GPU clockspeed up to 20% above stock clocks. Even better, the APU is able to allocate power to either the GPU or CPU depending on which area needs the boost and how much TDP headroom the chip has when doing certain tasks. For example, AMD shows that the A10-4600M APU can downclock the GPU from the default clockspeed of 685MHz to 496MHz, allowing the x86 Piledriver cores to achieve up to a 900MHz overclock at a clockspeed of 3.2GHz. Alternatively, when the GPU is needed, it can run at 685MHz while the CPU sits at 2.3GHz. They are likely not able to push the GPU much further as any more reductions in CPU speeds would need to be much bigger than any accompanying GPU increases. And at that point, the GPU would likely become bottlenecked and the system would be starved of too much CPU power anyway.
The Trinity APUs continue to be based on GlobalFoundries’ 32nm SOI manufacturing process, but this time the chips are slightly larger with a die size of 246mm^2. Although the APU is wholly larger than Llano, they actually have fewer transistors at 1.303 billion versus the 1.45 billion in Llano. Although that may seem like a step in the wrong direction, the new CPU modules and GPU cores are much more efficient than those in Llano so it should all balance out and Trinity should come out on top despite the lower transistor count. The Trinity APUs will also feature an improved instruction set that includes AVX, AVX1.1, FMA3, AES, and F16C which should help the CPU in certain tasks.
Overall, Trinity is looking like an improved part versus Llano, especially in the CPU department. Although AMD’s numbers should be taken with more than a grain of salt, they are claiming 26% better desktop system performance as a result of the CPU overhaul. Granted, Bulldozer was not a CPU powerhouse itself when compared to the competition, but it is–at least on paper–a good design. When paired with a relatively good GPU, as is the case of Trinity, the Piledriver [architecture based] (a refined version of the Bulldozer architecture with some under-the-hood tweaks) cores should at the least not hold the GPU back, and at best make the CPU processor performance good enough to make the Trinity APU all the more desirable to an even wider range of potential buyers. Pricing of the new APUs is still up in the air, but they are set to release later this month if a certain leak is to be believed.
I think that we can expect to see an all around better chip with Trinity, though pricing will be the ultimate factor in determining how popular it is. I suspect that Intel will still carry the CPU crown, but if the price is right, AMD can sell a lot of Trinity chips to builders that only need decent CPUs to support good integrated GPU cores in systems where the GPU is more important. I am anxiously awaiting reviews of the new Trinity chips and hoping that AMD continues to have successful chips with their line of APUs.