The Data: Soft Machines' Proof

Ever since the initial announcement of the VISC architecture in 2014 there has been a element of it sounding too good to be true, and Soft Machines' 2016 announcements have come with yet more questions as well. Aside from questions requiring more information about the architecture and ISA, the big money questions relate to performance. We mentioned a couple of pages back that the original 28nm design made its way to silicon at 500 MHz and was shown as a proof of concept. At the 2014 conference, the platform was compared to both ARM and x86 and offered better scores on Denbench compared to both while also using less power. Now that Shasta is on the 16nm node, the big question is how the new design at a 2 GHz frequency compares, and if the increase in frequency has upset some of the IPC gains.

So it’s at this point that we have to show this graph before we can progress any further. This is a graph which has caused a lot of commotion among the analyst community, because it can be a very difficult graph to digest and work out what is going on. I’ll take you through it. But to start, ignore the vertical dashed lines.

This is a graph from Soft Machines attempting to show the efficiency of several CPUs cores: Cortex-A72, Apple’s Twister (A9X), Skylake, Shasta, Shasta+ (2017), and Tahoe (2018). It is a graph of the average power consumed per unit SPEC2006 score plotted against the SPEC2006 score, and each dot on a line shows the relative score of each CPU at a given frequency.

The reason why this graph has caused a lot of commotion is that it shows a lot of data based on a lot of assumptions displayed in a very odd way. The following points are worth mentioning

The Scores #1 This graph shows the geometric mean of SPEC2006int and SPEC2006fp, the integer and floating point parts of the SPEC2006 set of benchmark tools. Because different architectures focus on integer and floating point performance to differing degrees (more units focused on INT or FP), these results are typically given separately, with individual subtest scores. Practically no-one in the industry puts them together as a geometric mean, which has some analysts wondering if there are certain subtests where VISC scores particularly low.
The Scores #2 This graph shows only single threaded results, even though each CPU is listed as having two cores in the data but running a single thread. This puts the Soft Machines cores in the best light, as a single thread has access to all the ports on two cores as well as two re-order buffers and two cores' worth of L2 cache.
Conversion #1 All the results have been converted as if each CPU design has 1MB of last level cache per core. This means that designs such as the A9X and Skylake CPUs have been reduced, and scores have been adjusted by ambiguous ‘industry standard techniques’ according to SMI. A number of analysts say that this is not a fair conversion, as an A9X core or Skylake core with less cache would be arranged differently in silicon to take advantage of more space for other things or lower latencies.
Conversion #2 All the results have been converted to 16nm FinFET+ on TSMC, again by ‘industry standard techniques’. This is a hard one to grasp, because core designs are not simply ‘shrunk’ from one node to another. Similar to the cache situation, each process node can be optimized for metal layers and arrangement for latency and bandwidth optimizations. Each conversion, such as Intel’s 14nm to TSMC 16nm, or TSMC’s 28nm -> 16nm, would have to be thoroughly examined. Extrapolating from 28nm to 16nm would be an exasperating task to be accurate (and this level of extrapolation wouldn’t be acceptable even in a high school classroom as I pointed out).
Testing #1 Not all points on the graph come from direct data. Each line has had several points taken from data and the rest are interpolated given basic power formulas.
Testing #2 The platforms used are not all what they appear to be. So for example, the best Cortex-A72 16nm data point would be the Kirin 950 in the Huawei Mate 8, but instead a dual A72 was used from the Amazon Fire TV which as a 28nm MediaTek MT8173 running at 1.98 GHz. One could argue that A72 is new enough and only recently on 28nm that it isn’t fully optimized for the process yet and this is probably a low end version of that silicon. The Apple A9X numbers are actually taken from a 14nm A9 and the assumption was made that the dynamic power in a cold environment was similar to the A9X. The Skylake numbers were a mid-range Core i5-6200U in a Dell laptop, which could be prone to variable turbo modes or overheating, and that specific SKU is hardly the most power efficient model in Intel’s Skylake lineup.
Compilers In order to ‘normalize’ the data, each of the actual data points taken were as a result of SPEC2006 being compiled on GCC 4.9 (or Clang for Apple). Typically for SPEC we normally consider the peak numbers possible with the best compiler, and as pointed out by some analysts, Intel’s results on their compiler can get scores more than double that of GCC, which can put a negative bent on Intel’s numbers.
Simulations Almost all of SMI’s numbers come from internal RTL simulation of their IP designs. With the 28nm proof-of-concept chip, we were told that the difference between simulation and physical was around 5-10% on performance and power, but some chip designers have pointed out that performance on a simulated processor can be anywhere from 33-50% inaccurate from the peak theoretical performance when you actually put it into silicon.
Optimizations The data shown in this graph for the VISC processors is based on assumptions relating to process optimization. The way the design is to be sold means that licensees can work with the foundries to optimize the metal stack layers or other design characteristics to get better power or higher frequencies. I was told that this was put into the graph at an assumed value around 10%, and the data in the graph includes this.

Typically any one of these points in most contexts would be grounds to be apprehensive about the results. The fact that there are nine salient points here listed (and I may even have missed one or two) means that the data should be thrown out entirely.

Clarification on the Data from Soft Machines

Because we were one of the last media outlets to speak with Soft Machines, and I had already seen some discussion around these points, I posed the issues back to them, as well as a few questions of my own. Because of the response that had been presented, we managed to get a lot of details around the simulation and assumptions aspects. So to start, here’s the testing methodology that everyone was provided with:

For clarity on the VISC processors, simulations were done to be both signal accurate and cycle accurate, and data taken from 16nm design configurations. Both power and results were taken from these.

For the power on the other parts, the power consumption was taken at the wall. To remove system power from the equation, the system was run in 2C vs 1C modes and 1C vs idle modes at various frequencies to find the dynamic power. Each platform was tested in a cold environment to ensure the maximum temperature did not go above 55C. Each of the power numbers are estimates that have removed the production yield ‘guard’ (i.e. protection overestimate), which was about ~15%, giving benefits to the non-VISC core results.

For the power conversion to 16nm FinFET+ on TSMC:

  • The A72 28nm TSMC used TSMC’s numbers for scaling.
  • For the A9X numbers, the A9 numbers were taken in a cool environment to minimize Samsung 14nm leakage and the dynamic power for the A9X is assumed the same as the A9.
  • For Intel, using public data it was assumed that 16FF+ voltage is 10-15% higher and area is 30% larger, giving ~1.65x power on 16FF+.
  • Leakage scales with die area.

For performance testing:

  • Linaro GCC 4.9 with -O3 -mcpu=cortex-a57 -static -flto -ffast-math
  • If any test failed, results were taken from previous generations and expected percentage increases. For example, FORTRAN on A9 failed, so estimates were taken from the floating point numbers in C.
  • For the cache adjustments, for 3MB to 2MB was reduced 3.5-4%. Because Intel has L2+L3, this is reduced a further 5%. For A72 moving from 1MB L2 to 2MB L2 in total, score was increased by 4%.

For a full rundown, this was the slide provided to us:

As part of these assumptions, I did ask about the raw data collected and if that would ever be presented. I did mention that they really need to split up INT and FP results, and I was told that it may happen at a later date but not right now. What I was given though was the effect of the cache adjustment on Skylake.

The orange line and the blue line next to it represents the movement from a multi-cache hierarchy of L2+L3 to 1MB of L2 cache per core only. The blue line still has the assumption of moving from 14nm to 16FF+, and using the GCC compiler, but the orange line has that extra assumpton.

Personally, these assumptions make me uneasy. Conversions like this are typically only done as back-of-the-envelope types of calculations during the early stages of design, because they are very rough and do not take into account things like silicon floor-plan optimization that would occur if you chose a smaller/larger cache arrangement, or changing from 8-way to 4-way associativity in the caches and so on. Typically we see companies in similar positions to SMI provide the raw or semi-modified data, using one assumption at most, with a split between FP and INT results - e.g. taking all the results as-is with GCC. The reason why it all comes into one graph is for brief simplicity, which doesn't particularly endear any reader/investor who might decide to be heavily invested in this project.

We pointed out a lot of concerns with this data to Soft Machines, including the list above of assumptions and how some of them simply do not make sense and should be restricted to that those quick rough calculations, especially when presenting at a conference. They gave us the graph above showing the effect of cache changes on Skylake, but I have asked in the future for them to display the data in a less complicated way, using standard industry metrics (such as INT or FP). Ideally the graphs are also kept to two or three data sets without requiring a 9-point interpretation scheme to understand what is happening - we typically get a dozen or so graphs from Huawei, ARM, AMD or Intel when they are describing their latest architecture or microarchitecture designs. This allows more understanding of what is happening under the hood and can be used to validate the results - as it stands, it is difficult to validate anything due to the assumptions and conversions made.

Soft Machines, VISC and Roadmaps But 16FF+ Silicon Exists
POST A COMMENT

97 Comments

View All Comments

  • Bleakwise - Tuesday, March 14, 2017 - link

    "Floating point code"

    "Integer code"

    Do you have any idea what you're talking about?

    Bulldozer does "flaoting poitn code" faster than the fucking 1080Ti

    At least one one thread. Unless you're going to go wide it doesn't help.

    The point of this isn't to "go wide" it's to massively increase speculation ability.

    The 1080Ti has ZERO speculative ability, NONE. GPUs simply don't do branching, that's not what GPUs do, they rely on ACE units and SMX units and so on to balance thousands of cores.

    A CPU on the other hand has more speculative branches than cores.

    SIMD and SIMT that GPUs do are not "FPU code"

    FFS
    Reply
  • dcbronco - Friday, February 12, 2016 - link

    AMD helped finance this. They may already have a stake and I would bet some right to first refusal. They used their investment in HBM to get earlier access than NVIDIA, I doubt they would have invested without some sort of incentives for themselves. Reply
  • Bleakwise - Tuesday, March 14, 2017 - link

    Of course not. Reply
  • bcronce - Saturday, February 13, 2016 - link

    There is no such thing as a free lunch. They are trading something. Their benchmarks are for single thread performance, which the graphs showed a much greater efficiency and performance than Intel. Very impressive and I'm sure they'll be great for something.

    The problem is the platform sounds great for highly coupled cores and very wide single thread execution with few data dependencies. Could be great for computation.

    What I'm wondering is how their platform scales for IO workloads like web servers, file servers, or event video games. Suddenly a large part of the work is communicating with other devices and synchronizing many cores.

    One thing that has helped ARM for a long time is they were mostly single core and only recently multi-core. They didn't use to have a complex cache-coherency like x86. This dramatically reduced transistor counts, increase efficiency, and allowed for great decoupled core performance. But as soon as you wanted two cores to work together, it went to crap. Cache-coherency is hardware accelerated inter-core communication. Amdahl's law was not very forgiving to ARM's non-cache-coherency cores for anything except GPU like workloads.

    Based on the description, VISC sounds like it needs highly couples cores to maintain low latency and high bandwidth. This is probably why they also seem to have lower frequency. Keeping many parts far away from each-other in sync takes time. But lower frequency also means lower voltage, and power consumption scales with the square of the voltage and linear with frequency.

    I wonder how tightly coupled they can keep 4, 8, or 16 cores. Maybe they don't need the core counts for their target workloads or possibly they can stay competitive with a fraction the core counts by having better efficiency in power and IPC.

    In the end, I'm sure they'll at least find a niche market and I'm glad some new ideas are making it out there. I wouldn't be surprised if they can take over the dual or quad core market, forcing Intel to add more cores.
    Reply
  • Bleakwise - Tuesday, March 14, 2017 - link

    It's not a "free lunch"

    Obviously all of this crap is going to cost DIE space, it's not free.

    If all we cared about was raw processing power we'd just make 2046kb wide vector units and ignore branching and speculation all together.

    Bulldozer has better theortical performance than Haswell i5s. I'd rather have the extra out of order pipes, the SMT unit to use any unused pipes, better branch prediction and so, and in the real world this stuff wins the day.

    Not everyone can become a world class programmer and re-factor all their code so that it spreads across thousands of cores like it can on a GPU.

    Sometimes it's not even possible. Sometimes what you need is branch prediction, branch prediction lets you see the future, LITERALLY this is what the CPU does. Obviously the more branches you predict, the more cycles you're wasting on that thread, because the more speculations you get wrong.

    You also reduce the number of misses and increase cache hits.

    As for coupling 4 or 16 cores, they haven't even talked about going beyond 4 cores. Obvoiusly it doesn't scale into infinity, if you're getting 90% speculative accuracy you can only gain 10% more. Spending 30% of your transitor budget to bind up 8 or 16 cores when spending 10% of your budget on 4, for a 10% performance gain would be dumb.

    You'd be much better off going for more clock speed, or reducing latency, adding a victim cache, or l2 cache coherency, or beefing up the GPU, a better memory controller, or just beefing up your underlying branch predictor,
    Reply
  • Bleakwise - Tuesday, March 14, 2017 - link

    You'll never get perfect speculation anyway. Unless a language is developed that puts limits on the number of branches possible per X lines of code and keep the number of branches below the number the CPU can handle. You're going to ALWAYS have to deal with the risk of cache misses.

    Not sure there is even anything you could gain from 100% target prediction hit grantee beyond having no lost cycles on a miss. Getting there even through a core-binding fabric/bus like this across 16 cores would blow your transistor budget to the point that you could hardly afford a reasonable size cache in the first place.

    You'd be better off just reducing the number of stages in the pipelines or just adding more pipelines to each core instead of blowing your budget on this fabric.

    For example, binding together 100 in order CPUs to make a virtual 100 pipeline CPU would be ridiculously expensive and power hungry vs just having an 8 core superscaler CPU with 12 out of order pipelines in each CPU.
    Reply
  • tipoo - Friday, February 12, 2016 - link

    Question, since this is testing their core design in isolation and the rests of the package hasn't been built around it, is that accommodated for in the comparisons to other SoCs, which all have far more die area dedicated to non-core stuff than the cores? Reply
  • Flunk - Friday, February 12, 2016 - link

    If VISC is not an Acronym then don't capitalize it, idiots.

    The technology looks like it could be really good, I'm hoping we see some practical applications.
    Reply
  • smilingcrow - Friday, February 12, 2016 - link

    They can capitalize it for any reason they like; it's just a word so nothing to GYKIATO (Get your kickers .... over). Reply
  • andychow - Friday, February 12, 2016 - link

    It's an acronym, you can't trademark acronyms, so now they claim it's not an acronym. Legal bs 101. Reply

Log in

Don't have an account? Sign up now