Kaveri and Counting Cores

With the move towards highly integrated SoCs we've seen a variety of approaches to core counts. Apple, Intel and Qualcomm still count CPU cores when advertising an SoC. For Apple and Qualcomm that's partially because neither company is particularly fond of disclosing the configuration of their GPUs. More recently, NVIDIA took the somewhat insane stance of counting GPU CUDA cores on its Tegra K1 SoC. Motorola on the other hand opted for the bizarre choice of aggregating CPU, GPU and off-die companion processors with the X8 platform in its Moto X smartphone. Eventually we will have to find a way to characterize these highly integrated SoCs, particularly when the majority of applications actually depend on/leverage both CPU and GPU cores.

AMD finds itself in a unique position with Kaveri where it has a truly unified CPU/GPU architecture and needs to establish a new nomenclature for use in the future. With 47% of the Kaveri die dedicated for GPU use, and an architecture that treats both CPU and GPU as equals, I can understand AMD's desire to talk about the number of total cores on the APU.

AMD settled on the term "Compute Core", which can refer to either an x86 (or maybe eventually ARM) CPU core or a GCN compute unit. The breakdown is as follows:

  • Each thread on a CPU is a Compute Core
  • Each Compute Unit on the IGP is a Compute Core
  • Total Compute Cores = CPU Compute Cores + IGP Compute Cores

This means that the high end SKU, the A10-7850K will have a total of 12 compute cores: four from the CPU (two Steamroller modules supporting four threads) and eight from the IGP (due to eight compute units from the R7 graphics).

There are some qualifications to be made on this front. Technically, AMD is correct – each compute unit in the IGP and each thread on the CPU can run separate code. The Hawaii GCN architecture can spawn as many kernels as compute units, whereas a couple of generations ago we were restricted to one compute kernel on the GPU at once (merely with blocks of work being split across the CUs). However, clearly these 12 compute units are not equivalent: a programmer will still have to write code for the CPU and GPU specifically in order to use all the processing power available.

Whenever AMD (or partners) are to promote the new APUs, AMD tells us clearly that two sets of numbers should be quoted in reference to the Compute Cores – the total, and the breakdown of CPU/GPU on the APU. Thus this would mean that the A10-7850K APU would be marketed at a “12 Compute Core” device, with “(4 CPU + 8 GPU)” following immediately after. I applaud AMD's decision to not obfuscate the internal configuration of its APUs. This approach seems to be the most sensible if it wants to tout the total processing power of the APU as well as tell those users who understand a bit more what the actual configuration of the SoC is. The biggest issue is how to address the users who automatically assume that more cores == better. The root of this problem is very similar to the old PR-rating debates of the Athlon XP. Explaining to end users the intracacies of CPU/GPU programming is really no different than explaining why IPC * frequency matters more than absolute frequency.

When a programmer obtains an APU, the OpenCL profiler should locate the eight compute units of the GPU and display that to the user for offloading compute purposes; at the same time it is up to the programmer to leverage the threads as appropriately as possible, even with AMD’s 3rd generation Bulldozer modules implementing a dual-INT + single-FP solution.

At launch, AMD will offer the following configurations:

  • A10-7850K: 12 Compute Cores (4 CPU + 8 GPU)
  • A10-7700K: 10 Compute Cores (4 CPU + 6 GPU)
  • A8-7600: 10 Compute Cores (4 CPU + 6 GPU)

The problem with distilling the processing power of the APU into a number of compute cores is primarily on the CPU side. AMD will be keeping the GPU at approximately the same clock (720 MHz for these three) but the CPU frequency will differ greatly, especially with the A8-7600 which has configurable TDP and when in 45W mode will operate 300-400 MHz less.

Steamroller Architecture Redux

We've already documented much of what Steamroller brings to the table based on AMD's Hot Chips presentation last year, but I'll offer a quick recap here.

Steamroller brings forward AMD's Bulldozer architecture, largely unchanged. We're still talking about a dual-core module featuring two independent integer execution cores with a single shared floating point execution core capable of executing instructions from two threads in parallel. A single module still appears as two cores/threads to the OS.

In Bulldozer and Piledriver, each integer core had its own independent scheduler but the two cores shared a single fetch and decode unit. Instructions would come in and decodeded operations would be fed to each integer pipe on alternating clock cycles. In Steamroller the decode hardware is duplicated in each module, so now each integer core gets its own decode unit. The two decode units are shared by the one FP unit.

L1 instruction caches increase in size from 64KB to 96KB per module, which AMD claims reduces misses by up to 30%. There's an updated branch predictor which is responsible for reducing mispredicted branches by up to 20%. Both integer and FP register files grow in size as well as an increase in the size of the scheduling window, a combination of which increases the number of dispatches per thread by up to 25%.

There are huge improvements on the store side. Steamroller can now issue up to 2 stores at the same time compared to 1 in Bulldozer/Piledriver. The load/store queue sizes go up by around 20% as well. It's remarkable just how much low hanging fruit there was in the Bulldozer design.

GCN in an APU

The integrated graphics solution on Trinity/Richland launched with a Cayman-derived VLIW4 architecture, which unfortunately came just shortly after desktop side of the equation finished moving from VLIW5/VLIW4 to GCN. Having a product stack with largely different GPU architectures doesn't help anyone, particularly on the developer front. Looking forward, sticking with GCN was the appropriate thing to do, as now Kaveri is using GCN, the same architecture found in AMD’s high end R9-290X GPU, based on the Hawaii platform.

This enabled AMD to add in all the features they currently have on Hawaii with little to no effort – the TrueAudio DSP, the upgraded Video Coding Engine and Unified Video Decoder are such examples. Whether or not AMD decides to develop an APU with more than 8 GCN CUs is another matter. This is a point we've brought up with AMD internally and one that I'm curious about - do any AnandTech readers have an interest in an even higher end APU with substantially more graphics horsepower? Memory bandwidth obviously becomes an issue, but the real question is how valuable an Xbox One/PS4-like APU would be to the community.

Kaveri: Aiming for 1080p30 and Compute Llano, Trinity and Kaveri Die: Compared
Comments Locked

380 Comments

View All Comments

  • geniekid - Tuesday, January 14, 2014 - link

    Would've been nice to see a discrete GPU thrown in the mix - especially with all that talk about Dual Graphics.
  • Ryan Smith - Tuesday, January 14, 2014 - link

    Dual graphics is not yet up and running (and it would require a different card than the 6750 Ian had on hand).
  • Nenad - Wednesday, January 15, 2014 - link

    I wonder if Dual Graphics can work with HSA, although I doubt due to cache coherence if nothing else.

    While on HSA, I must say that it looks very promising. I do not have experience with AMD specific GPU programming, or with OpenCL, but I do with CUDA (and some AMP) - and ability to avoid CPU/GPU copy would be great advantage in certain cases.

    Interesting thing is that AMD now have HW that support HSA, but does not yet have software tools (drivers, compilers...), while NVidia does not have HW, but does have software: in new CUDA, you can use unified memory, even if driver simulate copy for you (but that supposedly means when NVidia deliver HW, your unaltered app from last year will work and use advantage of HSA)

    Also, while HSA is great step ahead, I wonder if we will ever see one much more important thing if GPGPU is ever to became mainstream: PREEMPTIVE MULTITASKING. As it is now, still programer/app needs to spend time to figure out how to split work in small chunks for GPU, in order to not take too much time of GPU at once. It increase complexity of GPU code, and rely on good behavior of other GPU apps. Hopefully, next AMD 'unification' after HSA would be 'preemptive multitasking' ;p
  • tcube - Thursday, January 16, 2014 - link

    Preemtion, dynamic context switching is said to come with excavator core/ carizo apu. And they do have the toolset for hsa/hsail, just look it up on amd's site, bolt i think it's called it is a c library.

    Further more project sumatra will make java execute on the gpu. At first via a opencl wrapper then via hsa and in the end the jvm itself will do it for you via hsa. Oracle is prety commited to this.
  • kazriko - Thursday, January 30, 2014 - link

    I think where multiple GPU and Dual Graphics stuff will really shine is when we start getting more Mantle applications. With that, each GPU in the system can be controlled independently, and the developers could put GPGPU processes that work better with low latency to the CPU on the APU's built in GPU, and processes for graphics rendering that don't need as low of latency to the discrete graphics card.

    Preemptive would be interesting, but I'm not sure how game-changing it would be once you get into HSA's juggling of tasks back and forth between different processors. Right now, they do have multitasking they could do by having several queues going into the GPU, and you could have several tasks running from each queue across the different CUs on the chip. Not preemptive, but definitely multi-threaded.
  • MaRao - Thursday, January 16, 2014 - link

    Instead AMD should create new chipsets with dual AMU sockets. Two A8-7600 APUs can give tremendous CPU and GPU performance, yet maintaining 90-100W power usage.
  • PatHeist - Thursday, February 13, 2014 - link

    Making dual socket boards scale well is tremendously complex. You also need to increase things like the CPU cache by a lot. Not to mention that performance would tend to scale very badly with the additional CPU cores for things like gaming.
  • kzac - Monday, February 16, 2015 - link

    Having 2 or more APUs on a logic board would defeat the purpose of having an APU in the first place, which was to eliminate processing being handled by the logic board controller. With dual APU sockets, there would need to be some controller interjected to direct work to the APUs which could create a bottle neck in processing time (clock cycles). This is the very reason for the existence of multi core APUs and CPUs of today.

    Its my expectation that we will start to observe much more memory being added to the APU at some point, to increase throughput speeds. Essentially think of future APUs becoming a mini computer within, the only limitations currently to this issue are heat extraction and power consumption.
  • 5thaccount - Tuesday, January 21, 2014 - link

    I'm not so interested in dual graphics... I am really curious to see how it performs as a standard old-fashioned CPU. You could even bench it with an nVidia card. No one seems to be reviewing it as a processor. All reviews review it as an APU. Funny thing is, several people I work with use these, but they all have discrete graphics.
  • geniekid - Tuesday, January 14, 2014 - link

    Nvm. Too early!

Log in

Don't have an account? Sign up now