Apple A9x chip discussion

[dahorns]

Just doing some quick math on expected GPU numbers:

Manhattan
A9X = 37 (A8X score) * 2 = 74 fps
HD 515 = 54.1 (HD 530) / 1.15 (boost over HD 515) = 47

T-Rex
A9X = 70 * 2 = 140 fps
HD 515 = 127 / 1.15 = 110 fps

Ice storm Unlimited Overall
A9X = 21,000 x 2 = 42,000
HD 515 = 49,000 (HD 5300) * 1.4 = 68,600 OR
HD 515 = 84,000 (HD 530) / 1.15 = 73,000

Ice Storm Unlimited Graphics
A9X = 31,000 x 2 = 62,000
HD 515 = 59,000 x 1.4 = 82,600

I don't have a break down for the HD 530 Ice Storm Graphics score so I can't do the calculation in reverse. It is interesting that applying a 40% over 5300 or dividing 530 by 1.15 for the total Ice Storm score come up with very similar numbers.

 

[Fjodor2001]

How did you arrive at the 1.15 factor?

 

[Sweepr]

Just doing some quick math on expected GPU numbers:

Manhattan
A9X = 37 (A8X score) * 2 = 74 fps
HD 515 = 54.1 (HD 530) / 1.15 (boost over HD 515) = 47

T-Rex
A9X = 70 * 2 = 140 fps
HD 515 = 127 / 1.15 = 110 fps

Ice storm Unlimited Overall
A9X = 21,000 x 2 = 42,000
HD 515 = 49,000 (HD 5300) * 1.4 = 68,600 OR
HD 515 = 84,000 (HD 530) / 1.15 = 73,000

Ice Storm Unlimited Graphics
A9X = 31,000 x 2 = 62,000
HD 515 = 59,000 x 1.4 = 82,600

I don't have a break down for the HD 530 Ice Storm Graphics score so I can't do the calculation in reverse. It is interesting that applying a 40% over 5300 or dividing 530 by 1.15 for the total Ice Storm score come up with very similar numbers.

HD Graphics 530 numbers from ArsTechnica:

68.1 FPS Manhattan Offscreen and 139.7 FPS T-Rex Offscreen.

 

[dahorns]

How did you arrive at the 1.15 factor?

HD 530 has 15% higher boost than HD 515.

 

[dahorns]

HD Graphics 530 numbers from ArsTechnica:

68.1 FPS Manhattan Offscreen and 139.7 FPS T-Rex Offscreen.

Ah, I pulled it from GFX website. May have been a preliminary bench. Applying the same formula above you get:

68.1 / 1.15 = 59.2 FPS
139.7 / 1.15 = 121.47 FPS

Obviously not exact, but it'll be interesting how close the results are given the marketing from both.

 

[Edrick]

thats your confirmation bias right there because you don't know that. Second AVX2 did not double fp throughput , AVX2 doesn't even have FP instructions but int SIMD instructions, so dont let facts get in the way.........

What increased AVX throughtput from IB forward has been FMA and actually being able to load store 512/256bit a cycle.

AVX2 introduced FMA, not AVX1. And since FMA is part of AVX2, my statement stands as true. But thank you for the attempt.

 

[Thala]

For 176.gcc on integrate.i compiled with gcc 4.8.3 for x86-64 and gcc linaro 4.9 2014.09:
x86-64. 7250M instructions
AArch64 6690M instructions

Interesting isn't it? The ISA is the single most important interface between compiler and machine. If it is lacking register, the only chance to tell the machine you need an operand later is to push it to stack. Similar with 2 operand instructions, the only chance to tell the machine that you need an operand later is to push or move it. These push/pop/move instruction are executed like any other instruction and take up valuable CPU resources. Its not only fetch, decode and ALU resources but also the cache footprint is higher.

In conclusion, yes, the ISA is very important in how efficient a core's uArch can possibly be.

 

[name99]

Yeah, let's wait until this thing is thoroughly benchmarked before we jump to any conclusions purely based on marketing-speak.

(a) Which of Swift, Cyclone, and Typhoon have NOT matched Apple's performance claims? Every one of those CPUs was aggressively benchmarked by AnandTech (and others) and matched what Apple said.

(b) When Apple introduced the iPad Air 2 they called it 40% faster. (Same language as yesterday's keynote). Since that involved a transition from 2 to 3 cores, this suggests that Apple is referring to SINGLE-THREADED performance.
This is quite feasible --- it requires the CPU to run at say 2.3GHz, with 12% or so misc IPC improvements, or 2.4GHz with fewer IPC improvements.

Intel has Core-M's that turbo to 3.1GHz (though probably in a higher power/thermal envelope), so we're in the same ballpark, we're not requiring Apple to do anything outside the laws of physics.
Obviously this peak speed cannot be sustained indefinitely --- EVERY mobile core sprints for as long as it can then throttles down --- but if it can sustain that speed for say 60 seconds (as long as an A8 can sustain its peak speed) that's good enough for most "snappiness" purposes.

As for the indefinite duration speed (what you get after a long period of heating) for Intel that's around 1.6GHz for core-M. I expect Apple would be in the same ballpark.

 

[itsmydamnation]

AVX2 introduced FMA, not AVX1. And since FMA is part of AVX2, my statement stands as true. But thank you for the attempt.

FMA doesn't come anywhere near doubling throughtput like you are claiming, It gives a few percent on average. It also isn't apart of AVX2 it just came out at the same intel uarch. Thats why FMA has its own CPU flag and its instructions aren't listed in AVX2 instruction set list.

But thank you for your attempt, see agner for how wrong you are:
http://www.agner.org/optimize/instruction_tables.pdf

AVX2

Integer vector instructions are available in 256-bit versions. Furthermore, the
following instructions are added in AVX2: ANDN, BEXTR, BLSI, BLSMSK,
BLSR, BZHI, INVPCID, LZCNT, MULX, PEXT, PDEP, RORX, SARX, SHLX,
SHRX, TZCNT, VBROADCASTI128, VBROADCASTSS, VBROADCASTSD,
VEXTRACTI128, VGATHERDPD, VGATHERQPD, VGATHERDPS,
VGATHERQPS, VPGATHERDD, VPGATHERQD, VPGATHERDQ,
VPGATHERQQ, VINSERTI128, VPERM2I128, VPERMD, VPERMPD,
VPERMPS, VPERMQ, VPMASKMOVD, VPMASKMOVQ, VPSLLVD, VPSLLVQ,
VPSRAVD, VPSRLVD, VPSRLVQ

FMA3

(FMA): Fused multiply and add instructions: VFMADDxxxPD, VFMADDxxxPS,
VFMADDxxxSD, VFMADDxxxSS, VFMADDSUBxxxPD, VFMADDSUBxxxPS,
VFMSUBADDxxxPD, VFMSUBADDxxxPS, VFMSUBxxxPD, VFMSUBxxxPS,
VFMSUBxxxSD, VFMSUBxxxSS, VFNMADDxxxPD, VFNMADDxxPS,
VFNMADDxxxSD, VFNMADDxxxSS, VFNMSUBxxxPD, VFNMSUBxxxPS,
VFNMSUBxxxSD, VFNMSUBxxxSS.

 

[IntelUser2000]

Apple introduced the iPad Air 2 they called it 40% faster. (Same language as yesterday's keynote). Since that involved a transition from 2 to 3 cores, this suggests that Apple is referring to SINGLE-THREADED performance.

I don't think so. Anandtech benchmarks showed 20% ST and 60% MT. It's probably combined.

 

[IntelUser2000]

Ah, I pulled it from GFX website. May have been a preliminary bench. Applying the same formula above you get:

68.1 / 1.15 = 59.2 FPS
139.7 / 1.15 = 121.47 FPS

Obviously not exact, but it'll be interesting how close the results are given the marketing from both.

HD 5300 scores:

Manhattan:
-Yoga 3 Pro: 23
-Macbook 2015: 21.2
-HP Elitebook 1020: 25.7

T-Rex:
-Yoga 3 Pro: 45
-Macbook 2015: 54
-HP Elitebook 1020: 43.6

Now add 40% on top of that per Intel(best case) and you get your scores which are far lower than "extrapolated".

- Apple releases new CPU with more cores and new instructions that only show up in certain benchmarks. Apple fanboys response: "Apple is great. We get 80% performance increase almost every year. Who cares if it is only on specific benchmarks or marketing slides. Who cares if developers have to update their code to utilize it. Intel sux!"

Take a look again at claims and benchmarks of the previous products and you'll see they correlate pretty well.

 

[jhu]

That iPad Pro seems drool-worthy. I'd take that over a Surface Pro any day...

Running iOS? I wouldn't. Maybe if it was running MacOS instead. Can run full desktop programs in the Surface Pro (for me the relevent ones are Blender, Guild Wars2, and Java). At the very least, if/when Microsoft stops supporting the Surface Pro, other OSes can still be installed. Same can't be said for the iPad Pro unless Apple allows a way to unlock the bootloader (which actually won't happen).

 

[StrangerGuy]

There is something I don't really get: If Apple loves PowerVR so much in for their GPUs why don't Apple just outright buy them.

 

[Space69]

AVX2 doesn't even have FP instructions but int SIMD instructions, so dont let facts get in the way..........

This is incorrect. AVX2 has FP instructions.

 

[asendra]

There is something I don't really get: If Apple loves PowerVR so much in for their GPUs why don't Apple just outright buy them.

I mean, why bother? They already control a big part of PowerVR, and I'm pretty sure they dictate heavily the direction of their GPU design. Plus, they even get special custom design/configuration from them, or at least permission to modify them freely.

The only thing I see they could gain is exclusivity, but they don't compete directly against the rest of OEMs because they are the only OEM for their platforms.

 

[itsmydamnation]

This is incorrect. AVX2 has FP instructions.

anything outside of gather?

even intel themselve describe AVX2 as an int SIMD instructions set:

https://software.intel.com/en-us/blogs/2011/06/13/haswell-new-instruction-descriptions-now-available

AVX2 - Integer data types expanded to 256-bit SIMD. AVX2’s integer support is particularly useful for processing visual data commonly encountered in consumer imaging and video processing workloads. With Haswell, we have both Intel® Advanced Vector Extensions (Intel® AVX) for floating point, and AVX2 for integer data types.

But at this point we are so far removed from the original point which was all intel did to double their fp performance was remove there own throughput bottlenecks and it had nothing to do with AVX2.

 

[Nothingness]

This is incorrect. AVX2 has FP instructions.

AVX2 has FP instructions but AVX already had most of them. The only additional FP instruction AVX2 brought was FMA (fused multiply-accumulate).

This Intel page shows the impact of FMA on LINPACK (basically a best-case for FP improvements brought by AVX2).

 

[Space69]

There is something I don't really get: If Apple loves PowerVR so much in for their GPUs why don't Apple just outright buy them.

Recently both Apple and Intel was key shareholders of Imagination Technologies, but Intel decided to sell its share this year.

 

[Space69]

anything outside of gather?

Yes - permute (and I would classify FMA as part of AVX2). Gather was a big addition, but rather slow in some circumstances.

 

[Sweepr]

HD 5300 scores:

Manhattan:
-Yoga 3 Pro: 23
-Macbook 2015: 21.2
-HP Elitebook 1020: 25.7

T-Rex:
-Yoga 3 Pro: 45
-Macbook 2015: 54
-HP Elitebook 1020: 43.6

Now add 40% on top of that per Intel(best case) and you get your scores which are far lower than "extrapolated".

Straight from Notebookcheck:

Broadwell-Y @ Macbook 12'' (using dual-channel RAM)
T-Rex Offscreen 66.2 FPS
Manhattan Offscreen 34.1 FPS

Intel has been quoting 3DMark 11 for their graphics comparisons for quite some time now (that's where the 41% number comes from), so who knows how they actually compare @ GFXBench.

 

[itsmydamnation]

Yes - permute (and I would classify FMA as part of AVX2). Gather was a big addition, but rather slow in some circumstances.

But its not it is its Own instruction set, amd had fma3 on piledriver without avx2.

 

[Eug]

Some comments from Geekbench itself about Geekbench scores here:

http://forums.anandtech.com/showthread.php?t=2404965&page=3

This is a thread from last year describing performance on A8X. Look for jfpoole's posts.

PS. I like the posts from one guy saying A8X can't be tri-core.

There was also a post correctly predicting that A9X would be the one to look at, in terms of its speed against Core M.

Also lots of posts about SHA on ARM skewing results.

 

[dahorns]

HD 5300 scores:

Manhattan:
-Yoga 3 Pro: 23
-Macbook 2015: 21.2
-HP Elitebook 1020: 25.7

T-Rex:
-Yoga 3 Pro: 45
-Macbook 2015: 54
-HP Elitebook 1020: 43.6

Now add 40% on top of that per Intel(best case) and you get your scores which are far lower than "extrapolated".

1) I didn't apply the 40% to the GFX benches because I think they are broken on Gen 8 and prior architecture. Instead I based my numbers on the known scores of the HD 530. I think that is a pretty reasonable way to approach it. Actually, since you wouldn't expect the 15% frequency advantage of the HD 530 over the HD 515 to result in a straight 15% performance increase, my numbers may be low.

Also, as Sweeper mentioned, the 40% was related to 3dmark (Skydiver I think). There was a huge jump in GFX scores from the HD 4600 to the HD 530.

2) Where did you get those numbers? They are quite a bit lower than other scores I've seen.

3) Why the attitude? My assumptions were provided in my post. I also explicitly stated that the estimates may have no connection to actual product performance. It is just something to discuss until we have actual numbers for both the A9X and the Skylake Core M.

 

[Edrick]

But its not it is its Own instruction set, amd had fma3 on piledriver without avx2.

You clearly have your own opinion which seems to differ from most others.

But regardless of FMA being part of AVX2 or not, my original point still stands. Please stop getting caught up in semantics.

 

[name99]

There is something I don't really get: If Apple loves PowerVR so much in for their GPUs why don't Apple just outright buy them.

I think this is Apple's backup plan.
But right now, the existing situation is basically perfect for them.
They are PowerVR's ONLY large customer. Which means they can simultaneously work very closely with PowerVR to design a GPU that EXACTLY meets their needs
AND
they can sell downsized/crippled versions of that same chip to the current PowerVR customers (companies like Allwinner, Rockchip, and Mediatek) which allows some offset against the initial design costs.

In some ways (with respect to costs anyway) it's an even better situation than the CPU situation --- Apple gets the advantages of a custom-design while being able to offload some of the costs in a way they can't with their CPUs (unless, of course, there is some secret program they have selling their CPUs to another unnamed company...)