Intel doesn't care about the low-end 50cent chips, don't you know.
Intel's IoT is more than $2B per year, not does have 95% MSS and is growing. In the last few years it has had higher operating margin than their software or NAND group.
Intel doesn't care about the low-end 50cent chips, don't you know.Using Atom in IoT chips makes little sense except for the relatively small higher-end IoT market. For the mass market, you want very small and efficient chips, something x86 isn't good at. Cortex-M cores are what you want
I still stand to my opinion that Atom is completely useless except for low-cost machines with Windows legacy support.
Intel are just doing standard embedded stuff (as many have done for dozens of years) and claim it's IoT. You gotta love buzzwordsIntel's IoT isn't focused on consumer IoT
Intel are just doing standard embedded stuff (as many have done for dozens of years) and claim it's IoT. You gotta love buzzwords
But There Shall Be Fifty Billion IoTs.It's just marketing
The three decoders explain a lot of the improved performance.Intel Pentium N4200 (Apollo Lake/Goldmont) AIDA64 CPUID dump
Intel Pentium N4200 (Apollo Lake/Goldmont) Instruction latency dump
Airmont vs Goldmont:
Caches:
- L1D: same characteristic (24 KB, 6 way, 3 clk latency), but independent 16B load + 16B store ports instead of 16B shared load/store port (according to REP MOVS* throughput)
- L1I: same characteristic (32 KB, 8 way)
- L2: same characteristic (1 MB / 2 core, 16 way, ??? clk latency)
- 4K DTLB doubled (256 -> 512)
ISA extensions:
- SHA
- CLFLUSHOPT
- RDSEED
- PT
- MPE
- SMAP
- FSGSBASE
- no-VEX required ISA extensions e.g. BMI, AVX, F16C
Integer:
- Triple decode, triple issue. Goldmont can sustain 3 basic ALU instruction (MOV, ADD, SUB, CMP, TEST, AND, OR, XOR, NEG, NOT, NOP) per clock
- MOV ellimination works for 32 and 64 bit GPR
- ~2.5x faster 64b IDIV (107 -> 44)
- 2 operand write (I)MUL (7 -> 6)
- REP MOVS 16 -> 32 B/clk peak
- CRC32 6|6 -> 3|1
vector fp:
- Out-of-Order(*), fully pipelined, SP-DP L|T symmetrical
- NOV ellimination works for MOVA|UPS|D
- ADDPS 3|1 -> 3|1
- MULPS 5|2 -> 4|1
- ADDPD 4|2 -> 3|1
- MULPD 7|4 -> 4|1
- shuffle, pack 1|1 -> 1|0.5
- CVT*, EXTRACT*, INSERT* fully pipelined
vector int:
- NOV ellimination works for MOVDQA|U
- shuffle, pack, shift, PMOV* 1|1 -> 1|0.5
- PSHUFB: 5|5 -> 1|1
- Quadword add/sub 4|4-> 2|1
- PMUL* 5|2 -> 4|1
- AESENC/DEC 9|5 -> 6|2
HW SHA instrucions:
SHA1RNDS4 5|2
SHA1NEXTE 3|1
SHA1MSG1 3|1
SHA1MSG2 3|1
SHA256RNDS2 8|4
SHA256MSG1 3|1
SHA256MSG2 3|1
x87:
- FXCH doesn't block
- some uCode faster
ROB size:
- 48 -> ???
Intel removed the 6W TDP Apollo Lake parts from ark.intel.com database (Pentium N4200, Celeron N3450, N3350)
AFAIK Intel Denverton server (Atom C3xxx) and probably Knights Mill and Knights Hill will be based on this core
(*)Current form of the instruction latency dump doesn't implicate the OoO property, but I've seen a written statement about this from an Intel engineer
I hope later I can provide the computing evidence of OoO FPU, L2 latency and ROB size
QuadCore Intel Pentium N4200 (Apollo Lake) 605C9 CPUID, InstLat dump added (3x ALU, OoO fp FPU,HW SHA included)
Ok, I revive the thread, but this time about Intel SOFIA.
Recently I got an Intel SOFIA Based tablet and decided to do a Benchmark with Geekbench 4 and here is the results compared to the competition.
https://browser.primatelabs.com/v4/cpu/609232
MEDIACOM M-MPI7A3G (Tablet) - Intel SOFIA 1.05 GHZ Quad Core
Single Thread: 371
Multi Thread: 828
Antutu 6.2.1 - 18879 - And even it didn't run the Space 3D bechmark
--------------------
https://browser.primatelabs.com/v4/cpu/82703
Lenovo A 2010 - ARM A53 1.0 Ghz - Quad Core
Single Thread: 457
Multi Thread: 1221
--------------------
https://browser.primatelabs.com/v4/cpu/609236
Moto G 1st Gen - ARM A7 1.2 Ghz - Quad Core
Single Thread: 433
Multi Thread: 1093
-------------------
https://browser.primatelabs.com/v4/cpu/609226
CUSTOM OCed Xperia L - Krait 1.7 Ghz - Dual Core
Single Thread: 590
Multi Thread: 914
Antutu 6.2.1 - 24315
------------------
https://browser.primatelabs.com/v4/cpu/609204
Xiaomi Redmi Note 2 - ARM A53 2.0 Ghz - Octo Core.
Single Thread: 798
Multi Thread: 3139
Antutu 6.2.1 - 45463
-------------------
In few words, SOFIA sucks a LOT CPU wise.... is even worse than the worst phone which is the Lenovo A2010 manages to outperform it!
Even ARM A7 defeats SoFIA... what the hell Intel tought on going that way instead of continuing with Broxton?Wow yeah, SOFIA is kind of terrible. No value prop to this SoC compared to a cheap ARM A53 based SoC and probably more expensive to make. What a terrible showing.
By the way, some new scores for Apollo Lake @ Geekbench 4:
- Celeron N3350 @ 1.10 GHz (2C/2T):
SC: 1556
MC: 2722
https://browser.primatelabs.com/v4/cpu/548293
~50% faster multi-core score than current Celeron N3350 @ 1.10 GHz. Good news for future Chromebooks.
- Pentium N4200 @ 1.10 GHz (4C/4T):
SC: 1567
MC: 4433
https://browser.primatelabs.com/v4/cpu/567708
Some reviewers already have the notebooks so I hope we get something better than GB soon.
The range comprises two microATX models featuring the Celeron J3355 and J3455, two mini-ITX featuring the very same SoCs, as well as a final mini-ITX model rocking the Pentium J4205. We're hearing that Apollo Lake offers some serious improvement over Braswell and that the J4205 is on par with 1st gen Core M (with HEVC 10-bit decoding on top of it).
✓ J3355M
✓ Dual-Core Celeron J3355 (up to 2.5 GHz)
✓ 2 DIMM slots (Supports DDR3/DDR3L 1866 memory)
✓ 1 PCIe 2.0 x16 (x1 mode)
✓ 2 PCIe 2.0 x1
✓ 2 x USB 3.0 / 2 x USB 2.0
✓ 2 x SATA 3
✓ D-Sub, HDMI, DVI-D
✓ J3455M
✓ Quad-Core Celeron J3455 (up to 2.3 GHz)
✓ 2 DIMM slots (Supports DDR3/DDR3L 1866 memory)
✓ 1 PCIe 2.0 x16 (x1 mode)
✓ 2 PCIe 2.0 x1
✓ 2 x USB 3.0 / 2 x USB 2.0
✓ 2 x SATA 3
✓ D-Sub, HDMI, DVI-D
✓ J3355B-ITX
✓ Dual-Core Celeron J3355 (up to 2.5 GHz)
✓ 2 SO-DIMM slots (Supports DDR3/DDR3L 1866 memory)
✓ 1 PCIe 2.0 x16 (x2 mode)
✓ 3 x USB 3.0 / 1 x USB 2.0
✓ 2 x SATA 3
✓ D-Sub, HDMI
✓ J3455-ITX
✓ Quad-Core Celeron J3455 (up to 2.3 GHz)
✓ 2 SO-DIMM slots (Supports DDR3/DDR3L 1866 memory)
✓ 2 PCIe 2.0 x1 / 1 M.2 (Key E)
✓ 2 x USB 3.0 / 2 x USB 2.0
✓ 4 x SATA 3
✓ D-Sub, HDMI, DVI-D
✓ J4205-ITX
✓ Quad-Core Pentium J4205 (up to 2.6 GHz)
✓ 2 SO-DIMM slots (Supports DDR3/DDR3L 1866 memory)
✓ 2 PCIe 2.0 x1 / 1 M.2 (Key E)
✓ 2 x USB 3.0 / 2 x USB 2.0
✓ 4 x SATA 3
✓ D-Sub, HDMI, DVI-D
That's why they should had continued with Broxton....Intel Celeron J3455 @ 1.50 GHz @ Geekbench 3 Windows x86
Single-Core Score
1427
Multi-Core Score
4827
https://browser.primatelabs.com/geekbench3/8050780
That's a 1.5 GHz (up to 2.3 GHz Turbo) 4C/4T Apollo Lake for desktops.
That's why they should had continued with Broxton....
Agreed, but considering that Intel could scale it well down... they could reach ARM A72 levels at 14 nm.Keep in mind they are only getting that score by upping the TDP beyond what you could reasonably put in a smartphone. Or a tablet really.