AMD has spent the previous couple of years difficult Intel around the desktop, server, and cellular markets, however the hole between the 2 corporations is arguably greatest in server. At the moment, AMD ships as much as 64 cores in one socket, the place Intel has most effective stepped as much as transport 40 cores this week with the release of Ice Lake SP. Earlier Intel Cascade Lake CPUs crowned out at 28 cores. A brand new leak suggests Intel’s next-generation CPU platform, codenamed Sapphire Rapids, will after all search to scale back one of the vital gaps between itself and AMD’s Epyc.
As all the time, take this leak along with your day-to-day dose of salt. This slide comes from VideoCardz and it builds on some information we’ve prior to now noticed.
Sapphire Rapids, when it launches, will (supposedly) specify every other TDP building up — as much as 350W, this time. AMD’s present “Milan” CPUs peak out at 280W, identical to Rome. Reminiscence strengthen strikes to DDR5, as anticipated, and the slide claims Sapphire Lake gives 1TB of bandwidth on 64GB of HBM2E. We knew Sapphire Lake used to be going to provide HBM2E as an possibility, however 64GB of on-die reminiscence with 1TB/s of bandwidth is massive. It’d be in point of fact attention-grabbing to peer how gadget efficiency scaling adjustments with this configuration in comparison with fashions with out HBM2.
A top-end Sapphire Rapids, if those rumors are correct, would supply a small pool of ultra-high bandwidth reminiscence, subsidized by means of a a long way greater pool of lower-bandwidth reminiscence. An eight-channel DDR5 gadget the use of DDR5-4800 would supply 307.2GB/s of reminiscence bandwidth to as much as 4TB of RAM (assuming Intel keeps current Ice Lake SP limits).
Sapphire Rapids is claimed to characteristic as much as 80 PCIe 5.0 lanes on some SKUs, with others restricted to simply 64 lanes. It’s a four-tile design. This meshes with what we’ve discovered about Intel’s plan for tiles, which might be more or less analogous to AMD’s chiplets, however with other methods for I/O, bundle routing, and interconnects.
As for when those chips can be in-market, that’s just a little laborious to learn presently. Intel has made noise about transport Sapphire Rapids in 2021, however we’ve additionally heard that the chip wasn’t prone to release prior to 2022. Previously, there was once a big distinction between TSMC and Intel when it got here to the query of “quantity manufacturing.” That distinction is shrinking.
Intel would use the time period only some months prior to a chip went on sale, whilst TSMC may announce quantity manufacturing so long as a 12 months prior to chips become to be had to customers. Intel claimed to be in quantity manufacturing for Ice Lake SP in January 2021 and introduced in April, however reports from Dell recommend servers with the CPU received’t be to be had till Would possibly, and that that is “in sync with Intel’s timelines.” A January quantity announcement with Would possibly availability is a four-month extend. That’s longer than is standard for Intel.
As of this writing, we’re guessing Sapphire Rapids will pattern in 2021 however no longer release till 2022. It’ll compete in-market towards a mix of Milan and Genoa portions. Genoa is anticipated to be constructed on 5nm and to make use of AMD’s Zen 4 structure. There are rumors of an extra core depend building up, as much as 96 cores, however that can or will not be true.
With Zen 3, AMD fascinated about making improvements to Infinity Cloth efficiency and clock speeds, but it surely wound up spending considerably extra energy on “uncore” actions than Rome did. The corporate may select to concentrate on making improvements to IF and CPU potency with Zen 4 and dangle core counts equivalent, or it should decide to make the most of 5nm’s density enhancements and push core counts as soon as once more. 96 cores with out HBM2 and 12 reminiscence channels tackling 56 cores with HBM2 and 8 reminiscence channels? Sounds attention-grabbing to us.
This slide additionally mentions third-generation Optane, aka Crow Go, and claims bandwidth might be progressed by means of as much as 2.6x in combined learn/write eventualities. Not one of the information referring to Optane has been excellent in recent times, to the purpose that we’re gazing to peer if Crow Go even involves marketplace. Assuming that it does, then again, it looks as if the reminiscence same old will after all get an actual efficiency kick. No phrase on whether or not Crow Go helps PCIe 4.0 or PCIe 5.0, however Intel is obviously pushing to get Xeon again on a aggressive footing. Ice Lake SP is a cast effort for Chipzilla, but it surely doesn’t fully shut the space with AMD. Sapphire Rapids offers Intel every other shot at doing so.