Intel has already disclosed that it will have a next generation Atom core, code named Tremont, which is to appear in products such as the Foveros-based hybrid Lakefield, as well as Snow Ridge designed for 5G deployments. In advance of the launch of the core and the product, it is customary for some documentation and tools to be updated to prepare for it; in this case, one of those updates has disclosed that the Tremont core would contain an L3 cache – a first for one of Intel’s Atom designs. is an Intel website which hosts all of its open source projects. One of those projects is perfmon, a simple performance monitoring tool that can be used by developers to direct where code may be bottlenecked by either throughput, memory latency, memory bandwidth, TLBs, port allocation, or cache hits/misses. In this case, the profiles for Snow Ridge have been uploaded to the platform, and one of the counters provided includes provisions for L3 cache monitoring. This provision is directly listed under the Tremont heading.

Enabling an L3 cache on Atom does two potential things to Intel’s design: it adds power, but also adds performance. By having an L3, it means that data in the L3 is quicker to access than it would be in memory, however there is an idle power hit by having L3 present. Intel can mitigate this by enabling parts of the L3 to be powered on as needed, but there is always a tradeoff. There can also be a hit to die area, so it will be interesting to see how Intel has changed the microarchitecture of it’s Atom design. There is also no indication if the Tremont L3 cache is an inclusive cache, or a non-inclusive cache, or if it can be pre-fetched into, or if it is shared between cores or done on a per-core basis.

Intel’s Atom roadmap, as disclosed last year at Architecture day, shows that the company is planning several more generations of Atom core, although beyond Tremont we get Gracemont in 2021, and beyond that is ‘increased ST Perf, Frequency, Features’ listed around 2023. In that time, Intel expects to launch Sunny Cove, Willow Cove, and Golden Cove on the Core side.


The first public device with Tremont inside is expected to be the Core/Atom hybrid Lakefield processor, which uses Intel’s new Foveros stacking technology. We know that this design will have one Sunny Cove core and pair it with four Tremont cores. Intel expects chip production of Lakefield for consumer use by the end of the year.

Related Reading

Source: InstLatX64,



View All Comments

  • IntelUser2000 - Wednesday, July 17, 2019 - link

    You don't just put DRAM on-die and call it done. It doesn't work that way. You have to make a CPU-process specific DRAM and that's why they call it eDRAM. eDRAM also has lot lower density compared to regular DRAM. It's still better than SRAM in density, but not like order of magnitude(and in case its not clear, 1 order of magnitude = 10x) or greater better. Intel's eDRAM for example had a density advantage of just 3x over their SRAM.

    Yes you have a point about the power compared to system memory but that power is in addition to being quite a bit slower. You are talking slower in both latency and bandwidth. The density increase isn't enough to displace DRAM anyways.

    And reducing power in complex systems require tackling it in multiple areas. Just because its in ok in DRAM doesn't make sense to increase it on the CPU.

    Don't assume things without doing more research.
  • mode_13h - Wednesday, July 17, 2019 - link

    Thanks for the explanation.

    > Don't assume things without doing more research.

    Well, I wasn't sure so I asked. You didn't have to answer, but I'm glad you did. Thanks.
  • Santoval - Tuesday, July 16, 2019 - link

    L3 cache is always (or practically always) SRAM based. L4 cache is usually DRAM based. SRAM is much more expensive die area wise, but it's also quite faster and more power efficient than DRAM. DRAM is not fast enough for an L3 cache.

    Perhaps in the future (STT-)MRAM will instead be used for L4 cache, or maybe even L3 cache. MRAM is almost as fast as SRAM but it's much denser (thus much more cache can fit in each mm^2), it's more power efficient and it's also non volatile. I have no idea if CPU cache non volatility can become a useful feature, but I imagine it might.
  • mode_13h - Wednesday, July 17, 2019 - link

    Persistent L3 cache would make it more efficient for a CPU to sleep and periodically wake up to do some small amount of work.

    However, for security reasons, I'm thinking persistence wouldn't be utilized across reboots, etc.
  • HStewart - Monday, July 15, 2019 - link

    I very curious about the efforts that Intel is doing here, I believe there is more happening here because there is a major change it chip design here - it sounds to me that the Lakefield has a fast core for computer and 4 Tremont cores for backgound tasks - which sounds idea for portable always on computer.

    But this Snow Ridge is a server based product and I believe it more than just Xeon D. There might be reason for Xeon Phi discontinuation and Snow Ridge could be the reason. With larger cache and more cpus in device, this could be excellent low power server. I would not doubt 16 or more cores is likely in very small space.

    Toms Hardware has interest article about Intel plans past Foveros and part it discusses CO-EMiB which is designed for datacenter - which sounds like it combine 4 Foveros together which each having 8 chiplets - to me that sound like at least 32 core system that is only if one core per chiplet which I would assume it could have more

    Do me Snow Ridges sound like next generation of C3000's series.
  • HStewart - Monday, July 15, 2019 - link

    Supremicro has C3xxx based system aim at cloud computing
  • IntelUser2000 - Wednesday, July 17, 2019 - link

    Snow Ridge is specifically set to act as a compute node for 5G base stations. It's not C3000 successor which is general purpose.

    They'll likely have a Tremont-based successor to C3000.
  • abufrejoval - Monday, July 15, 2019 - link

    So that's the reason you can score Gemini Lake Atoms or J5005 all of a sudden again?

    They used to be near impossible to obtain!

    I got one last week; turned out very much more responsive than the J1900 and N3700 I already had, at least on Windows 10: Really nothing to complain about even on a 4K screen, can't say that an Nvidia 2080ti on an 5GHz i7 or a 4GHz Xeon 18 core is dramatically faster on Firefox or Office.

    I had DDR4 SO-DIMMs lying around 1x 32GB and 2x 16GB that seemed sad to waste (and were priceless a year ago), so that's why I clicked on 'order' when a full Mini-ITX ASRock MoBo could be had for much less than the price of that RAM at the time (turns out, it's quite reasonable these days).

    Actually, after the really impressive initial tests on Windows I got greedy and ordered another two, to create an oVirt Gluster for functional testing but at zero acoustics and very little power.

    Alas, the on-board RealTek has issues with more complicated things on Linux, some of which can be healed by cold-booting (warm boots seem to loose the network on CentOS...)

    It's also not generally as snappy on a Linux desktop as it is on Windows, but that's now how I plan on using them.

    The common theme on all Atoms since the J1900 (Braswell, I think): Those 8GB RAM limitations are 'market segmentation lies', whatever RAM you can physically fit into those SO-DIMM slots, the Atoms will address it all. With DDR3 16GB were never any issue, with DDR4 32GB work just fine, too.

    Few will actually feel tempted to load a $100 system with 32GB of RAM, but I thought it worth mentioning, that you could.

    I only wish they'd make these boards with some higher quality 5Gbit NBase-T onboard NICs, too for say a $20 premium, because that would make them a nicely rounded fanless/noiseless low power home appliance.
  • Jorgp2 - Monday, July 15, 2019 - link

    Isn't the J5005 NuC discontinued? Reply
  • abufrejoval - Tuesday, July 16, 2019 - link

    Not using a NUC, because those have fans. These are ASRock Mini-ITX boards (€116 with VAT) in a chassis just big enough to house it with up to 2 SATA SSDs using a massive heat-sink.

    100% silent and typically 6-8 Watts at the wall plug on idle.

    The chassis (€40 with VAT) comes with a 65 Watt 12V power brick and an internal 12V -> ATX power conversion unit that isn't optimal at that load point, but who knows what will be in there two years down the road? I prefer modularity.

    There are silent conversion kits for NUCs, but those are only worth paying for on high-end variants with i7 or similar.

    Some of those configurations (NUC7i7BNH, for example) would be rather nice to have as Mini-ITX, alas Intel won't have it!

    I have an i7-7700T (35 Watt) running as pfSense appliance with a Noctua NH-L9i fan in a chassis like that and you have to put your ear right next to it to hear anything. Under constant high loads (deep inspection of 400MBit traffic) it becomes noticeable but never a bother. Normal loads and even shorter spikes (several seconds) get completely buffered inside the high mass of the Noctua cooler.

    Whereas even with 15 Watt NUCs I found them far too noisy on high loads and far too nervous with the fans: Almost an audible CPU graph... yucks!

    If NUCs had Noctuas, I might go for them, especially with Thunderbolt/USB4 replacing PCIe slots. So far they are design over function.

Log in

Don't have an account? Sign up now