Originally Posted by xlink
How much does FP performance matter? I remember hearing a while back that FP mattered more than integer for most things. Has the pendulum swung back the other way?
Floating point performance is still as critical as ever. Integer performance is key as main loops and a majority of variables are integers in most software. Tho once you get into gaming and heavily compute workloads than floating point performance is a big factor. This is why online games like Battlefield take a huge performance hit once online vs offline mode. For example player position is represented by a float, not just one but at least four (for the basics x, y, z, radius) of them for each and every player. On a 64 player map (64 * 4 = 256) every game client has to calculate 256 floating point numbers at least every second just to keep up with proper character positions and movements in the game. Now that's only the new character positions, the client has to store the existing character positions to know which way the character is moving etc. That brings you up to already 512 floats to be calculated nearly every second ingame. And that's only the start when it comes to heavy floating point numbers that are being calculated frequently. I am in the process of writing a benchmark tool for calculating CPU performance. I can tell you this much floating point performance on my A10-6800k is roughly 50% slower than integer performance. Completing 1 million integer loops at around 20 seconds and 1 million float loops at around 30 seconds. That's with me patching out x87 and strictly using SSE instructions as well (to avoid skewed results with Bulldozer). So yea FPU performance is still as critical as ever, and it will continue to be. There is no getting rid of floats in programming, its just something that simply cannot be done. Hence why AMD is pushing forward HSA and trying to move floating point calculations to the GPU where they are done much faster (faster than integer core performance). In laymans terms if games like Battlefield used the GPU to calculate all the floating point numbers ingame like character position. You would actually see a frame rate increase while online vs offline mode (that is if offline didn't as well). Due to the CPU not being bogged down by these slower operations. The fact that the CPU has to do them is why there is an easy 10-20 frame rate difference between offline and online in games like Battlefield. Hopefully in the future developers will use the iGPU for all floating point calculations, so the CPU can reap full integer performance while gaming. With hQ that will be entirely possible as the iGPU doesn't have to be queued by the CPU. It can be issued and complete instructions entirely on its own.
Originally Posted by NaroonGTX
Carrizo is supposed to be the Excavator APU scheduled for Q1 2015 release.
Carrizo is actually suppose to be Steamroller refresh. I would expect context switching and everything else to come along with it sometime 2014. Kaveri is already showing DDR4 support that others may have mistaken for GDDR5 support in the past. DDR4 runs a single channel per dimm so each module runs on its own dedicated channel to the controller. It would explain the following source outlining four separate channels within the boot code.
// MEMORY-SPECIFIC DATA STRUCTURES
#define MAX_CHANNELS_PER_SOCKET 4 ///< Max Channels per sockets
#define MAX_DIMMS_PER_CHANNEL 4 ///< Max DIMMs on a memory channel (independent of platform)
#define UMA_ATTRIBUTE_ON_DCT0 0x40000000ul ///< UMA resides on memory that belongs to DCT0
#define UMA_ATTRIBUTE_ON_DCT1 0x20000000ul ///< UMA resides on memory that belongs to DCT1
#define UMA_ATTRIBUTE_ON_DCT2 0x10000000ul ///< UMA resides on memory that belongs to DCT2
#define UMA_ATTRIBUTE_ON_DCT3 0x08000000ul ///< UMA resides on memory that belongs to DCT3
Basically what this says is Kaveri may have quad channel support built into its memory controller just with two channels disabled. AMD could of jumped the gun and added DDR4 support so when its refresh time all they have to do is tweak the cores and launch the new product (Carrizo). No reason for them to implement Excavator yet as Kaveri will breath new life once it gets DDR4 support. Especially for those interested in building APU based gaming rigs. This would allow AMD to launch a new product later in 2014 without having to rush Excavator. Sorta like how Richland was a space gap after Trinity. Except this time we would obviously see marginal improvements other than clock speed.