Page 1:AMD’s Gambit
Page 2:Llano Products And Platforms
Page 3:The West-Side CPU
Page 4:The East-Side GPU
Page 5:CPU + GPU = APU: East Meets West
Page 6:The Driver Interface And AMD Steady Video
Page 7:Test Setup And Benchmarks
Page 8:Graphics Benchmarks: Synthetic
Page 9:Benchmark Results: Crysis 2
Page 10:Benchmark Results: StarCraft 2
Page 11:Benchmark Results: Metro 2033
Page 12:Benchmark Results: F1 2010
Page 13:Benchmark Results: Just Cause 2
Page 14:Benchmark Results: H.A.W.X. 2
Page 15:Game Benchmarks: Anti-Aliasing
Page 16:CPU Benchmarks: Synthetic
Page 17:CPU Benchmarks: Productivity
Page 18:CPU Benchmarks: Content Creation
Page 19:CPU Benchmarks: Media Encoding
Page 20:Accelerated Video Encoding: APP Vs. Quick Sync
Page 21:HD Video Quality Benchmark: HQV 2.0
Page 22:Power Usage And Battery Life
Page 23:Conclusion: Llano Brings A Lot Of Potential To The Table
The West-Side CPU
The CPU portion of Llano’s die is based on the Stars architecture we know from the original Phenom, living on in today’s Athlon II and Phenom II processors. There are a few notable differences and improvements, though. The most obvious is AMD’s transition from 45 nm manufacturing to a 32 nm node. So, right out of the gate, we expect certain power- and thermal-oriented advantages compared to the same piece of logic manufactured at 45 nm. We don’t have that here, of course. Adding the Fusion-oriented components just about doubles Llano’s transistor count compared to Deneb, adding up to 1.45 billion.
Bear in mind that Sandy Bridge-based CPUs already employ a mature 32 nm process, and Intel is on the verge of making a shift to 22 nm with the Ivy Bridge die shrink later this year. AMD knows that Intel has a manufacturing technology advantage and is looking to compensate in other ways.
AMD claims that Llano’s processing cores demonstrate an average 6% improvement in successfully executed instructions per clock (IPC) versus previous Stars-based offerings. In some cases there will be zero improvement and in others the new core might show as much as a 15% increase. There are two main reasons for this on-average performance improvement: more L2 cache and a better hardware prefetcher.
As you likely noticed, the 6 MB shared L3 CPU cache is gone entirely, a loss AMD hopes will be mitigated by a doubling of the exclusive L2 cache from 2 MB to 4 MB, totalling 1 MB of dedicated cache per core. This might sound like a dire compromise at first, but keep in mind that although a large L3 on the back of the northbridge allows sharing of another high-speed repository between cores, it also increases latency and presents a problem for power management. Increasing the L2 might sacrifice a little overall scalability, but it helps augment multithreading operations and allows more granular control of the chip’s power management. Since this was one of the main goals of Llano’s design team, the change made sense. And although the L2 and L3 cache configurations are changed, each core has the same 64 KB L1 instruction and data cache (128 KB total L1) as the 45 nm Phenom II and Athlon II families.
AMD also put a lot of work into its hardware prefetcher. Traditionally, a hardware stride prefetcher looks at specific memory instructions over time, and if the address loaded differs by a constant stride, the prefetcher may decide that it’s beneficial to put that instruction into the cache. The problem with this approach is that many programs can make it difficult to extrapolate patterns because of spurious instructions that distract the algorithm. AMD enhanced the prefetcher intelligence with Instruction Pointer (IP)-based prefetching. IP understands the instructions accessing the memory and uses that intelligence to find patterns, improving its ability to find strides.
Aside from this, the buffer sizes are larger. The reorder buffer is about 20% larger and the load/store buffers are doubled in size. We’re also told that the hardware multiplier is better, but AMD won’t disclose exactly what it improved.
When all is said and done, the 6% average IPC increase over the Phenom II represents a modest gain, if it’s quantifiable at all in testing. It seems that improving CPU performance was not a major target for the design team. And again, this makes well-enough sense to us with the first Bulldozer-based CPU (Zambezi) in the wings.
Power management takes precedence, which is smart since Llano needs to share its thermal ceiling between the CPU and GPU. While the aging Stars architecture might be sufficient to show what the GPU side can do today, this APU’s successor replaces Stars with Bulldozer-derived silicon. That part may be what manages to challenge Intel on the processor side. AMD’s Rick Bergman showed off a Trinity-based APU at this year’s Computex and said the first silicon was showing up at the lab. Hopefully the company isn’t late with that part.
- AMD’s Gambit
- Llano Products And Platforms
- The West-Side CPU
- The East-Side GPU
- CPU + GPU = APU: East Meets West
- The Driver Interface And AMD Steady Video
- Test Setup And Benchmarks
- Graphics Benchmarks: Synthetic
- Benchmark Results: Crysis 2
- Benchmark Results: StarCraft 2
- Benchmark Results: Metro 2033
- Benchmark Results: F1 2010
- Benchmark Results: Just Cause 2
- Benchmark Results: H.A.W.X. 2
- Game Benchmarks: Anti-Aliasing
- CPU Benchmarks: Synthetic
- CPU Benchmarks: Productivity
- CPU Benchmarks: Content Creation
- CPU Benchmarks: Media Encoding
- Accelerated Video Encoding: APP Vs. Quick Sync
- HD Video Quality Benchmark: HQV 2.0
- Power Usage And Battery Life
- Conclusion: Llano Brings A Lot Of Potential To The Table