5nm / 5LPE: What do we know?

Starting off with the biggest new change of this generation, both the Snapdragon 888 and the Exynos 2100 are manufactured on Samsung’s new 5nm process node, which is the biggest unknown in today’s comparison.

What’s important to remember is that although Samsung calls this node 5nm, its design and characteristics are more similar to that of their 7nm node. Key new characteristics of the new node here are the reintroduction of single diffusion breaks (SDB) on an EUV process node, as well as slight changes in the cell libraries of the process.

Advertised PPA Improvements of New Process Technologies
Data announced by companies during conference calls, press briefings and in press releases
  7LPP
vs 10LPE
6LPP
vs 7LPP
5LPE
vs 7LPP
3GAE
vs 7LPP
Power 50% lower 20% 50%
Performance 20% ? 10% 35%
Area Reduction 40% ~9% <20% 40%

Per Samsung’s own numbers, the foundry claims that 5LPE is either 20% lower power than 7LPP, or 10% more performance. These are actually quite important figures to put into context, particularly when we’re comparing designs which are manufactured on TSMC’s process nodes.

In least year’s review of the Galaxy S20 series and the Exynos 990 and Snapdragon 865 SoCs, an important data-point that put things into context was Qualcomm’s Snapdragon 765 SoC which was also manufactured on Samsung’s 7LPP node, and featuring Cortex-A76 cores. In that comparison we saw that The Exynos 990 and the Snapdragon 765’s A76 cores behaves very similarly in terms of power consumption, however they fell behind TSMC-based Cortex-A76 cores by anywhere from 20% to 30%.

In that context, Samsung’s 5LPE process node improving power by 20% would mean they’d only be catching up with TSMC’s 7nm nodes.

Cortex A55 @ 1.8GHz - SPEC2006 456.hmmer Power

An interesting comparison to make – and probably one of the rare ones we’re actually able to achieve today, is the comparison between the Cortex-A55 cores inside of both the Snapdragon 865 and the new Snapdragon 888. Both SoCs feature the same IP cores, clock them at the same 1.8GHz frequency, and both feature the same amount of L2 cache, with their only real difference being their process nodes.

Using SPEC’s 456.hmmer – because it’s a workload that primarily resides in the lower cache hierarchies and thus, we avoid any impact of the possibly different memory subsystem, we can see that both SoCs’ power consumption indeed is almost identical, with performance also being identical with a score of 6.84 versus 6.81 in favour of the new Snapdragon 888.

So at least at first glance, our theory that Samsung’s 5LPE merely just catches up with the power consumption and power efficiency of TSMC’s N7/N7P nodes seems to be valid – at least at these frequencies.

Further interesting data is the voltage curves of the CPUs on the Exynos 2100. I’ve extracted the frequency voltages tables of both my devices, a regular S21 and an S21 Ultra, with the above curves being the better binned chip inside of the smaller S21.

Generationally, Samsung seems to have been able to greatly reduce voltages this generation. On the Cortex-A55 cores, the cores now only require 800mV at 2GHz whilst the Exynos 990 last year in our review unit they required over 1050mV. Similarly, although the comparison isn’t apples-to-apples, the Cortex-A78 cores at 2.5GHz only require 862mV, while the Cortex-A76 cores of the previous generation required also 1050mV.

What’s also very interesting to see is the voltage curves of the Cortex-X1 cores versus the Cortex-A78 cores: they’re both nigh identical to each other, which actually lines up with Arm’s claims that the new X1 cores have the same frequency capabilities as the A78 cores, only being larger and increasing their power consumption linearly in relation to their frequency.

Samsung’s frequency tables indicate that they had been testing the A55 up to 2.6GHz, and the X1 and A78 cores up to 3.2GHz – however voltages here are quite higher and it’s also likely SLSI wouldn’t have been able to achieve similar chip yields.

Unfortunately, I wasn’t able to extract data from my Snapdragon 888 S21 Ultra, so I can’t tell exactly where it falls in terms of voltages compared to the Exynos 2100. One thing I can confirm as being quite different between the two SoCs is that Samsung does actually give the Exynos 2100’s Cortex-X1 core its own dedicated voltage rail and PMIC regulator, while the Snapdragon 888 shares the same voltage rail across the X1 and A78 cores. In theory, that could mean that in more mixed-thread workloads, the Exynos has the opportunity to be more power efficiency than the Snapdragon 888.

Generally, the one thing I want people to take away here is that although Samsung calls this their 5nm node, it’s quite certain that it will not perform the same as TSMC’s 5nm node. Usually we don’t care about density all too much, however performance and power efficiency are critical aspects that effect the silicon and the end-products’ experiences.

The Snapdragon 888 & Exynos 2100 Memory Subsystem & Latency: Quite Different
POST A COMMENT

122 Comments

View All Comments

  • eastcoast_pete - Monday, February 8, 2021 - link

    Andrei, also special thanks for the power draw comparison of the A55 Little Cores in the (TSMC N7) 865 vs the (Samsung 5 nm) 888! That one graph tells us everything we need to know about what Samsung's current " 5 nm" is really comparable to. I really wonder if QC's decision to chose Samsung's fabbing was more based on availability (or absence thereof for TSMC's 5 nm) or on price? Reply
  • DanD85 - Monday, February 8, 2021 - link

    Well, seems like Apple hogging most of TSMC 5nm node leaves other with no other choice but going with the lesser foundry. Reply
  • heraldo25 - Monday, February 8, 2021 - link

    For such a thorough review it is shocking to see that software versions (build number) used during tests are not stated.
    It is absolutely essential that the review contains software versions, so that other can try to replicate results, and for the reviewing site, to have references during re-tests.
    Reply
  • name99 - Monday, February 8, 2021 - link

    The milc win is certainly from the data prefetcher. In simulation milc also benefits massively from runahead execution, ie same principle (bring in data earlier).

    Has anyone identified a paper or patent that indicates what ARM are doing? A table driven approach (markov prefetcher) still seems impractical, and ARM don't go in for blunt solutions that just throw area at the problem. They might be doing something like scanning lines as they enter L2 for what look like plausible addresses, and prefetching based on those, which would cover a large range of pointer-based use cases, and seems like the sort of smart low area solution they tend to favor.
    Reply
  • trivik12 - Monday, February 8, 2021 - link

    Hope Qualcomm moves next gen flagship SOC to TSMC again. Cannot be at so much disadvantage. Of course Samsung 3nm could narrow the gap, but that is more for 2023 flagships.
    Disappointing to see Exynos disappoint again. How is Exynos1080 as a mid range chipset?
    Reply
  • geoxile - Monday, February 8, 2021 - link

    Their 3nm is expected to be on par with TSMC N5. The expect gains over 7nm are only 30% higher performance, 35% die area reduction, and 40-50% power reduction. Considering 5LPE is still behind N7P it's not much and will be barely be on par with N5 in density let alone efficiency. Reply
  • jeremyshaw - Monday, February 8, 2021 - link

    In other words, Samsung strangled then killed SARC for their failures, only to find the failures were with SSI itself. Reply
  • geoxile - Monday, February 8, 2021 - link

    You must be kidding... The Exynos 2100 is at least somewhat close to the Snapdragon 888 in CPU performance. Mali continues to be a problem, and remains so even for the Kirin 9000 on TSMC N5. Mongoose was an abomination that belonged maybe in 2015. Samsung Semiconductor is less competent than TSMC but SARC's mongoose team was a joke. Reply
  • EthiaW - Monday, February 8, 2021 - link

    All those attempts to spend transistors niggardly and boost performance by high frequency have failed miserably.
    Single transistor performance seems to be decaying from node to node now. Flat & Not more enough transistor count=performance regression.
    Reply
  • eastcoast_pete - Monday, February 8, 2021 - link

    Andrei, when you're testing the actual phone, could you check the battery life with the 5G modem on and off, respectively? 5G modems are supposedly quite power hungry also, and, if it's possible to turn 5G off (but leaving 4G LTE on), it would be interesting to see just how much power 5G really consumes. Reply

Log in

Don't have an account? Sign up now