The Ice Lake Benchmark Preview: Inside Intel's 10nm
by Dr. Ian Cutress on August 1, 2019 9:00 AM EST- Posted in
- CPUs
- Intel
- GPUs
- 10nm
- Core
- Ice Lake
- Cannon Lake
- Sunny Cove
- 10th Gen Core
Security Updates, Improved Instruction Performance and AVX-512 Updates
With every new microarchitecture update, there are goals on several fronts: add new instructions, decrease the latency of current instructions, increase the throughput of current instructions, and remove bugs. The big headline addition for Sunny Cove and Ice Lake is AVX-512, which hasn’t yet appeared on a mainstream widely distributed consumer processor – technically we saw it in Cannon Lake, but that was a limited run CPU. Nonetheless, a lot of what went into Cannon Lake also shows up in the Sunny Cove design. To complicate matters, AVX-512 comes in plenty of different flavors. But on top of that, Intel also made a significant number of improvements to a number of instructions throughout the design.
Big thanks to InstLatX64 for his help in analyzing the benchmark results.
Security
On security, almost all the documented hardware security fixes are in place with Sunny Cove. Through the CPUID results, we can determine that SSBD is enabled, as is IA32_ARCH_CAPABILITIES, L1D_FLUSH, STIBP, IBPB/IBRS and MD_CLEAR.
This aligns with Intel’s list of Sunny Cove security improvements:
Sunny Cove Security | |||
AnandTech | Description | Name | Solution |
BCB | Bound Check Bypass | Spectre V1 | Software |
BTI | Branch Target Injection | Spectre V2 | Hardware+OS |
RDCL | Rogue Data Cache Load | V3 | Hardware |
RSSR | Rogue System Register Read | V3a | Hardware |
SSB | Speculative Store Bypass | V4 | Hardware+OS |
L1TF | Level 1 Terminal Fault | Foreshadow | Hardware |
MFBDS | uArch Fill Buffer Data Sampling | RIDL | Hardware |
MSBDS | uArch Store Buffer Data Sampling | Fallout | Hardware |
MLPDS | uArch Load Port Data Sampling | - | Hardware |
MDSUM | uArch Data Sampling Uncachable Memory | - | Hardware |
Aside from Spectre V1, which has no suitable hardware solution, almost all of the rest have been solved through hardware/firmware (Intel won’t distinguish which, but to a certain extent it doesn’t matter for new hardware). This is a step in the right direction, but of course it may have a knock-on effect, plus for anything that gets performance improvements being moved from firmware to hardware will be rolled into any advertised IPC increase.
Also on the security side is SGX, or Intel’s Software Guard Instructions. Sunny Cove now becomes Intel’s first public processor to enable both AVX-512 and SGX in the same design. Technically the first chip with both SGX and AVX-512 should have been Skylake-X, however that feature was ultimately disabled due to failing some test validation cases. But it now comes together for Sunny Cove in Ice Lake-U, which is also a consumer processor.
Instruction Improvements and AVX-512
As mentioned, Sunny Cove pulls a number of key improvements from the Cannon Lake design, despite the Cannon Lake chip having the same cache configuration as Skylake. One of the key points here is the 64-bit division throughput, which goes from a 97-cycle latency to an 18-cycle latency, blowing past AMD’s 45-cycle latency. As an ex-researcher with no idea about instruction latency or compiler options, working on high-precision math code, this speedup would have been critical.
- IDIV -> 97-cycle to 18-cycle
For the general purpose registers, we see a lot of changes, and most of them quite sizable.
Sunny Cove GPR Changes | |||
AnandTech | Instruction | Skylake | Sunny Cove |
Complex LEA | Complex Load Effective Address | 3 cycle latency 1 per cycle |
1 cycle latency 2 per cycle |
SHL/SHR | Shift Left/Right | 2 cycle latency 0.5 per cycle |
1 cycle latency 1 per cycle |
ROL/ROR | Rotate Left/Right | 2 cycle latency 0.5 per cycle |
1 cycle latency 1 per cycle |
SHLD/SHRD | Double Precision Shift Left/Right | 4 cycle latency 0.5 per cycle |
4 cycle latency 1 per cycle |
4*MOV | Four repated string MOVS | Limited instructions | 104 bits/clock All MOVS* Instructions |
In the past we’ve seen x87 instructions being regressed, made slower, as they become obsolete. For whatever reason, Sunny Cove decreases the FMUL latency from 5 cycles to 4 cycles.
The SIMD units also go through some changes:
Sunny Cove SIMD | |||
AnandTech | Instruction | Skylake | Sunny Cove |
SIMD Packing | SIMD Packing now slower | 1 cycle latency 1 per cycle |
3 cycle latency 1 per cycle |
AES* | AES Crypto Instructions (for 128-bit / 256-bit) |
4 cycle latency 2 per cycle |
3 cycle latency 2 per cycle |
CLMUL | Carry-Less Multiplication | 7 cycle latency 1 per cycle |
6 cycle latency 1 per cycle |
PHADD/PHSUB | Packed Horizontal Add/Subtract and Saturate |
3 cycle latency 0.5 per cycle |
2 cycle latency 1 per cycle |
VPMOV* xmm | Vector Packed Move | 2 cycle latency 0.5 per cycle |
2 cycle latency 1 per cycle |
VPMOV* ymm | Vector Packed Move | 4 cycle latency 0.5 per cycle |
2 cycle latency 1 per cycle |
VPMOVZX/SX* xmm | Vector Packed Move | 1 cycle latency 1 per cycle |
1 cycle latency 2 per cycle |
POPCNT | Microcode 50% faster than SW (under L1-D size) | ||
REP STOS* | Repeated Store String | 62 bits/cycle | 54 bits/cycle |
VPCONFLICT | Still Microcode Only |
We’ve already gone through all of the new AVX-512 instructions in our Sunny Cove microarchitecture disclosure. These include the following families:
- AVX-512_VNNI (Vector Neural Network Instructions)
- AVX-512_VBMI (Vector Byte Manipulation Instructions)
- AVX-512_VBMI2 (second level VBMI)
- AVX-512_ BITALG (bit algorithms)
- AVX-512_IFMA (Integer Fused Multiply Add)
- AVX-512_VAES (Vector AES)
- AVX-512_VPCLMULQDQ (Carry-Less Multiplacation of Long Quad Words)
- AVX-512+GFNI (Galois Field New Instructions)
- SHA (not AVX-512, but still new)
- GNA (Gaussian Neural Accelerator)
(Intel also has the GMM (Gaussian Mixture Model) inside the core since Skylake, but I’ve yet to see any information on this outside a single line in the coding manual.)
For all these new AVX-512 instructions, it’s worth noting that they can be run in 128-bit, 256-bit, or 512-bit mode, depending on the data types passed to it. Each of these can have corresponding latencies and throughputs, which often get worse when going for the 512-bit mode, but overall assuming you can fill the register with a 512-bit data type, then the overall raw processing will be faster, even with the frequency differential. This doesn’t take into account any additional overhead for entering the 512-bit power state, it should be noted.
Most of these new instructions are relatively fast, with most of them only 1-3 cycles of latency. We observed the following:
Sunny Cove Vector Instructions | |||||
AnandTech | Instruction | XMM | YMM | ZMM | |
VNNI | Latency | Vector Neural Network Instructions | 5-cycle | 5-cycle | 5-cycle |
Throughput | 2/cycle | 2/cycle | 1/cycle | ||
VPOPCNT* | Latency | Return the number of bits set to 1 | 3-cycle | 3-cycle | 3-cycle |
Throughput | 1/cycle | 1/cycle | 1/cycle | ||
VPCOMPRESS* | Latency | Store Packed Data | 3-cycle | 3-cycle | 3-cycle |
Throughput | 0.5/cycle | 0.5/cycle | 0.5/cycle | ||
VPEXPAND* | Latency | Load Packed Data | 5-cycle | 5-cycle | 5-cycle |
Throughput | 0.5/cycle | 0.5/cycle | 0.5/cycle | ||
VPSHLD* | Latency | Vector Shift | 1-cycle | 1-cycle | 1-cycle |
Throughput | 2/cycle | 2/cycle | 1/cycle | ||
VAES* | Latency | Vector AES Instructions | 3-cycle | 3-cycle | 3-cycle |
Throughput | 2/cycle | 2/cycle | 1/cycle | ||
VPCLMUL | Latency | Vector Carry-Less Multiply | 6-cycle | 8-cycle | 8-cycle |
Throughput | 1/cycle | 0.5/cycle | 0.5/cycle | ||
GFNI | Latency | Galois Field New Instructions | 3-cycle | 3-cycle | 3-cycle |
Throughput | 2/cycle | 2/cycle | 1/cycle |
For all of the common AVX2 instructions, xmm/ymm latencies and throughputs are identical to Skylake, however zmm is often a few cycles slower for DIV/SQRT variants.
Other Noticeable Observations
From our testing, we were also able to prove some of the other parts of the core, such as the added store ports and shuffle units.
Our data shows that the second store port is not identical to the first, which explains the imbalance when it comes to writes: rather than supporting 2x64-bit with loads, it only supports either 1x64-bit write, or 1x32-bit write, or 2x16-bit writes. This means we mainly see speed ups with GPR/XMM data, and the result is only a small improvement for 512-bit SCATTER instructions. Otherwise, it seems not to work with any 256-bit or 512-bit operand (you can however use it with 64-bit AVX-512 mask registers). This is going to cause a slight headache for anyone currently limited by SCATTER stores.
The new shuffle unit is only 256-bit wide. It will handle a number of integer instructions (UNPCK, PSLLDQ, SHUF*, MOVSHDUP, but not PALIGNR or PACK), but only a couple of floating point instructions (SHUFPD, SHUFPS).
261 Comments
View All Comments
Phynaz - Saturday, August 3, 2019 - link
You guessed wrong, but considering your education level that should be expectedKorguz - Saturday, August 3, 2019 - link
i did ?? sure seems like you only want quad cores still... my guess, yours is maybe grade 10.AntonErtl - Friday, August 2, 2019 - link
I am guessing that SPEC CPU2017 Rate-1 means one copy of the benchmark, so it also gives us single-thread performance (and that explains why the power limit differences do not play a bigger role). Right?One thing that I find remarkable about the recent AMD and Intel launches is that Intel follows the pattern (already seen with Broadwell) of starting a new process with a low clock, and therefore with mobile products and then servers. If they follow their earlier pattern, we will see higher-clocked stuff on 10nm in the future.
By contrast, AMD (and the smartphone SoC vendors) manage to get the same or better clock rates with the newer process as with the older ones.
From what I hear, Intel is very much into hand-tuning their CPUs (that's why backporting Ice Lake to 14nm would be very expensive in terms of both labor and time-to-market), while AMD and the smartphone SoC vendors rely on more automatic tools. What is interesting is that AMD managed to get within a few % of clock rate of a highly tuned Intel chip (with several years of tuning both the process and the layout of the chip) with this more automatic approach, while having a similar IPC.
Intel has been working on Ice Lake and 10nm for a while, and still suffer from a clock rate deficit compared to Whiskey Lake (but in a way, 10nm is just becoming usefully online now, so maybe we should start the clock now). The fact that they have increased the TDP of the Y SKUs and decreased the base clock of the 15W SKUs (from 1.9GHz for the 8665U to 1.3GHz for the 1065G7) indicates that they still have problems being competitive wrt power consumption under load with Whiskey Lake. It will be interesting to see the final products (and compare them to Whiskey Lake and Picasso products with similar power delivery and cooling).
Phynaz - Friday, August 2, 2019 - link
AMD pushes their overvolted chips to the very limit. Intel doesn’t have to.Korguz - Friday, August 2, 2019 - link
" Intel doesn’t have to. " um.. i guess you havent seen the articles about how much intels cpus really use then. 95 watts, up to 200 watts in realityPhynaz - Saturday, August 3, 2019 - link
Wow, you’re really, really dumb. Do you get state aid?Korguz - Saturday, August 3, 2019 - link
oh how so ?? AT even posted a write up on this, obviously, you havent read it. so you are the one that needs aid, as you are the dumb one here....Alexvrb - Monday, August 5, 2019 - link
Reading isn't Phynaz' strong suit. However when you state a fact he can't counter, he immediately switches to low brow personal attacks - those are really more his speed.mobutu - Friday, August 2, 2019 - link
It doesn't mattter (for me), this time I'll buy AMD.HStewart - Friday, August 2, 2019 - link
Then why comment if it doesn't matter to you.