mobile soc 2014
TRANSCRIPT
Antonio Luca Alfeo
Maurizio Palmieri
ARM CORTEX
1. Snapdragon 810 (big.LITTLE)
2. Mediatek MT6752
3. Nvidia Tegra K1
4. HiSilicon Hi3798M V100
5. Mediatek MT8173 (big.LITTLE)
INTEL
1. Clover trail+
2. Bay trail
3. Moorefield
4. Cherry trail
High reliability (less than 10 fault per year).
Long time battery life (at least 1 day).
High computing performance (more than 2 GB Ram, 4
IPC,1.7Ghz ).
The Z35XX (Moorefield) family of Intel SoC offers fast and
smart performance and long battery life for Android
devices.
It is LTE optimized: supports Intel's XMM 7260 LTE silicon
which is CAT 6 capable.
Moorefield is based on the Silvermont CPU:
1. 64-bit microprocessor.
2. Out-of-order execution.
3. Back-end decodes and issues 2 instructions per cycle.
4. Dispatches 5 operations per cycle.
5. Consuming under 1W/core.
2 predictors:
1. BTB that provides low latency predictions over
instruction fetching.
2. Decode predictor controls the speculative
instructions issuing into the back-end and can
override earlier predictions.
A quad-core system with frequency from 1.8 up to 2.3
GHz(thanks to Intel SpeedStep technology)
24KB D-cache(6 ways) and 32KB I-cache(8 ways) for L1 caches.
2*1MB L2 cache(16 way) each shared by a couple of cores.
Up to 4GB of Main memory(2*LLDDR3) with a max bandwidth
of 12,8 Gbps.
A 533 MHz PowerVR G6430 GPU
The L2 is non-inclusive and non-exclusive relative to the
L1D and is a writeback design with full ECC protection
and pseudo-LRU replacement.
System agent allows reordering of memory requests
optimize for performance and quality of service.
Silvermont integrates Intel’s Core IDI, a lightweight point to
point interface.
This interface has independent read and write channels, and
it features high bandwidth and support for out-of-order
transactions.
In Silvermont, the IDI and CPU frequencies are largely
decoupled - enabling good bandwidth out of the cores even
at low frequency levels.
50% higher IPC for Silvermont
versus the previous generation.
Out-of-order scheduling provides 30% and perhaps 5-
10% for better branch
prediction.
The 22nm process technology
reduces Vmin by 100mV, and
increases clock frequencies by
20-30%.
Online intensive gaming (MMORPG for 4 hours)
Complex GPU processing (at least 60 fps)
High quality streaming video and audio (more than FHD
video)
High resolution screen
Tagliare l’immagine
Both chips are made on a 20nm process, big.LITTLEcortex A57 + A53, clocked at 2.5 GHz+.
Both have their LPDDR 4 memory clock speed at 1.6GHz, with peak memory bandwidth reaching up to 25.6GB/s.
Optimization procedure are different.
GPU: Tegra X1's GPU 256 unit at clock speed reaches a whirring 1GHz, while the Adreno 430 288 GPU cores purrs at 600MHz.
More (from 1 to 4, not
symmetric neither the
same) core clusters
switchable by
cpufreq driver to
balance
power/performances
ratio.
Very scalable
solution, 40
microsecond switch,
possible all cpu on.
Cache coherency tecnique:
1. Software: flush and invalidate.
2. Hardware: more efficient and simply done by AMBA 4 ACE
bus interface through L1 shared cache and L2 different (for
purpose) but visible cache.
A snoop filter reduce the snooping traffic. When replacement
of one of the entries is required, the snoop filter selects for
replacement the entry representing the cache line or lines
owned by the fewest nodes. This can be improved by further
locality based algorithm.
3.5x the performance of Cortex-A15 processor.
Frequencies of 2.5 GHz in a 16nm FF+ process and
scalable to higher frequencies.
75% reduction in energy consumption on heavy tasks.
Combining with a Cortex-A53 CPU in big.LITTLE
configurations, saves an additional 40-60% energy.
Smooth performances on multiple light app (e.g.
checking social network, email, news reader…)
Low cost (<250 €)
Notable design
Raw computational power leads to thermal issue. When
not needed must be avoided. If not the measures taken
to react to this issue could be heavy (e.g. thermal
throttling):
Mobile markets are growing and changing year after year,
low-cost is becoming the most important target"
INTEL FAMILY
INTEL ATOM Z3580 [~55$ ]
Asus Zenfone 2 = 350 $
Nokia N1 (tablet) = 250 $
Dell Venue 8 7000 (tablet) = 400 $
BIG.LITTLE FAMILY
SNAPDRAGON 810 [~65$ ]
LG Flex 2 = 700 $
HTC One M9 = 650 $
SAMSUNG EXYNOS 7420 [~30$]
Samsung Galaxy s6 = 750 $
A53 (derived A7) FAMILY
SNAPDRAGON 410 [~30$ ]
Samsung Galaxy A5 = 300 $
HTC desire 620 = 250 $
MEDIATEK MT6735 [~40$ ]
Sony Xperia E4g = 130 $
Lenovo A5000 = 150 $
Some low-cost SoC promise the same (or better
performance) at lower costs: e.g. snapdragon 410 cost
25% more than Mtk comparteur providing 30% less
computational power.
We have seen that different market targets lead to
different meaning of word ENOUGH: enough battery life,
enough performance, enough graphic quality.
To achieve these goals trade-off and different strategies
must be taken into account. Some are better.
http://www.anandtech.com/show/6936/intels-silvermont-architecture-revealed-getting-serious-about-mobile/5
http://www.pcper.com/reviews/Mobile/Dell-Venue-8-7000-Review-Intel-and-Android-Perfected/Battery-Life
http://www.realworldtech.com/silvermont/
http://download.intel.com/newsroom/kits/atom/comms/pdfs/Intel_Atom_Z35XX_FactSheet.pdf
http://www.arm.com/files/pdf/CacheCoherencyWhitepaper_6June2011.pdf
http://www.embedded.com/electronics-news/4419448/Benchmarking-ARM-s-big-little-architecture
http://mobiletoday.it/3227/android/arm-cortex-a72-e-gpu-mali-t-880-tecnologia-per-il-2016.html