Note: Notice very carefully, the tiny but incredibly self-satisfied tag in the corner labeled ‘Rumor’. The sources behind this leak are given and are probably authentic, not to mention very difficult to fake; however this news has not been confirmed either by official or unofficial channels.
The motherload of leaks concerning Carrizo APU has just occurred. The test results of an Engineering Sample of the chip have leaked online and we now know everything from hardware details to assorted benchmarks, including GFX Bench and Sisoft Sandra. Do keep in mind however, that this is an engineering sample so, alot of things can change between here and the consumer version of Carrizo.
Massive leak reveals AMD ‘Gardenia’ Carrizo specs: 2MB L2 Cache, 4 Cores clocked at 2.6Ghz with 512 Stream Processors
I received the original leak (SiSoft Benchmark Database) courtesy of CHW.net, and from there on I did some rather successful (if I say so myself) Googling to net me the entirety of this leak. The first thing that you will notice is the interesting nomenclature employed by AMD before Carrizo, namely Gardenia. These suffixes usually denote the target market of a specific product in AMD’s lineup so Gardenia most likely denotes either an embedded or a workstation application. The second thing I noticed, and thanked the gods for, were two identifying IDs that I could use to track the APU in other sites as well. This was the device ID ‘2M1801C1Y4381_26/18/08/04_9874’ and the codename ‘Carrizo 9874’.
While the former yielded no results and led me back to the starting place the latter hit the jackpot and led me straight to GFXBench results of the chip in question. Lets start with the GPU side first. The new iteration of Carrizo is supposed to feature next generation GCN cores, although it is not known which specific architecture AMD chooses alongside this. A viable option would be the Pirates Island Architecture. The benchmarks show Gardenia Carrizo acing the T-Rex tests (OpenGL ES 2.0) but having some serious difficulty in Manhattan tests (OpenGL ES 3.0).
The graphics portion of the Carrizo APU dubbed “Carrizo 9874” will feature 512 stream processors which appears to be the sweet spot between CPU and GPU ratio on an APU. This is the same number that was featured in Kaveir. The SPs will be split into 8 CUs and clocked at 626Mhz. They will be coupled with DDR3 Memory Clocked at 1.6GHz effective and a bus width of 128 bits. The APU has a fill rate of 8.5GTexel/s. As far as benchmarks go, the APU’s performance in GFXBench’s Manhattan benchmark is not very good at around 9.8/13.5 frames per second. However, in T-Rex, both onscreen and offscreen, it scores pretty impressive numbers (46.6/71.7 frames per second). Once again, I would remind you guys that specifications such as memory configurations can change.
Interestingly, the numbers over at the compute side are pretty skewed. The clock speed that was used is not the maximum listed 2.6Ghz boost, but 600Mhz less. TDP mentioned is approximately 40W but this should increase to ~50W once you enter full load on CPU and GPU ( 2.6Ghz and all SPs firing). The general arithmetic of the core is 24.2 GFlops, which is pretty decent for an APU, I remember my old Core 2 Duo used to score around 10 GFlops. The only thing that really bugs me is the memory support of DDR3, which was also revealed in a leak published by us just a few weeks ago.
Slow memory bottlenecks the crap out of an efficient architecture such as an APU, and I sincerely hope that AMD decides to ditch the slow memory and proceed to a higher end configuration such as HBM or DDR4 at the very least. Oh and incase anyone is wondering, the fact that the benchmark shows ‘Normal Desktop’ means nothing. Mobile Chips are usually tested at desktops early in their cycle. In all probability and in accordance with all the leaks so far, the first APU we are going to be seeing is Carrizo-L; the mobile version of the chip. At any rate, we will know more about the Carrizo APU as we get ever closer to its leaked launch date (December).
— Usman Pirzada (@usmanpirzada) October 31, 2014