As part of its 2021 CES activities, Intel gave an update on the Ice Lake Xeon generation. This statement is in-line with what we heard in the Q3/Q4 timeframe that the final revision of silicon was going to be a new spin that would push out Ice Lake a bit further. In early 2020 Intel was communicating a late 2020 launch for Ice Lake with some SKUs arriving in early 2021. With the latest spin, production of the final silicon was pushed back so we now have production in Q1 2021.
Intel’s Statement on Ice Lake Production
Here is the excerpt from Intel’s press release:
As part of its disclosures, Intel announced the recent production of its 3rd Gen Intel® Xeon Scalable processors (code-named “Ice Lake”) with volume ramp taking place during the first quarter of 2021. Intel’s 10nm Xeon Scalable processors feature architectural and platform innovations that boost performance, security and operational efficiency within data centers.
The 3rd Gen Intel Xeon Scalable processors help deliver the computing power customers need to thrive in this new, complex world thanks to significant increases in core count, performance, integrated artificial intelligence (AI) and security.
The latest Intel Xeon Scalable processors will enable cloud, enterprise and networking customers to deploy innovative systems and services, such as next-generation cloud services, that strengthen data privacy, AI video analytics, and microservices at the edge. The new Intel Xeon Scalable processors will be formally announced in the coming months. (Source: Intel PR excerpt)
For many wondering what this means, we are going to quickly give some context here.
Intel Ice Lake Xeon Status
When AMD EPYC 7003 “Milan” did not launch, as we expected in Q4 2020, it may have seemed like Intel has the upper hand again with 10nm manufacturing. Since Milan is largely based on the Zen 3 architecture, AMD has been in production of Milan since at least September 2020 and delivering parts to large customers. Milan’s public launch we were hearing for some time was pushed to January, now our best guess is still in Q1, but likely in the last third of the month. Again, we know that Milan is in production with “retail marked” parts, but we are awaiting the formal launch.
Ice Lake Xeons bring a large number of features to the Intel Xeon platform. We get more cores, PCIe Gen4 support with 64 lanes per socket, 8-channel DDR4-3200 support. Intel is also bringing trusted computing with SGX (differentiated by tier) and a host of other tweaks along with the new core architecture. 10nm for Intel is a big deal. In Q3 2018, we expected 2020 to be the year of Ice Lake and 2019 to be the year of Cooper Lake.
By Supercomputing in November 2019, it was fairly clear that we would not get Cooper Lake in 2019. STH covered the news that Cooper Lake would not be coming to the Whitley platform shared with Ice Lake.
A few quarters later, we saw a similar story. Ice lake would be coming in 2020, with 2021 being the Sapphire Rapids part. Sapphire Rapids uses a new platform, so the timing between Ice Lake and Sapphire Rapids is important.
With today’s announcement, it seems as though Intel is confirming that it is in production with the retail stepping of 10nm Ice Lake Xeons. That is a big deal since it puts us on track for a Q2 2021 launch assuming some of the recent production to launch timelines are the correct guidance. This is still a big change for Intel’s partners from early 2020 when it was expected that there would be some retail parts for server launches in 2020 with the availability of other die variants available in early Q1 2021. Clearly, starting production does not mean available. While Intel has been delivering previous steppings to partners to the point that it is fair to say partners and large customers had them in 2021, from what we understand those were not final stepping parts.
This may seem like a nuance, but it is extremely important. AMD has been shipping production Zen 3 parts since 2020 both on the desktop (Ryzen) and in the server (EPYC) segments. The delta between the CCDs in desktop and server parts for AMD is nowhere near Ice Lake desktop/ mobile and server parts. AMD was shipping its next-gen in Q4 2020, even if the EPYC 7003 “Milan” launch we have pegged for the back half of Q1 2021.
For Intel, saying Ice Lake Xeons were shipping in 2020 is a big deal if the retail stepping is in production now, ramping for a Q2 2021 launch. The reason it is important is that by Intel moving its key milestones to shipping low volume silicon or pre-retail shipping to customers, it is getting closer to the way Arm is messaging Neoverse. When Arm announces a core, there is some lag time between when that core is ready and when partners take the core and turn it into production silicon with memory controllers, I/O, and simply going through the fabrication steps. We know Neoverse N1 cores have had a number of wins such as Ampere Altra for the Oracle cloud and the AWS Graviton2.
While AMD has been shipping retail stepping Zen 3 since Q4 2020, even with Milan’s public launch delayed, Intel is not at the point yet where all of its partners are doing qualifications/ customer demos are happening on final stepping silicon.
As a result, Intel saying Ice Lake was a 2020 product, is becoming more akin to how Arm announces cores before retail stepping parts are available.
There is a lot of nuance here. One can “launch” a product which is a paper launch announcement. Products can be launched well after all of the large customers have parts. Of note, shipping to cloud customers early is important so infrastructure can be deployed at the public launch. That same practice also weakens the channels not selling to hyper-scale since it creates a systemic incentive for end customers to go cloud for the latest technology (and cloud providers pay less for chips and so are effectively being subsidized by enterprise customers as well.) Now that those same cloud customers are building their own chips, this early launch strategy is trading gains today for future share. In years past, this was not as big of a concern, but as these schedules get pushed, it is a practical tool that can be used to maintain share today.
The concern, based on current roadmaps is what a Q2 2021 Ice Lake Xeon launch means for non-cloud customers. Whitley is a “one-and-done” platform, and we typically see 6-9 months for entire server portfolios to get filled out after a new platform. If Sapphire Rapids is launched in 2021-early Q1 2022, it effectively will get launched as the normal platform transition to a new processor generation would be happening. That creates a strange dynamic where Intel and its partners are saying “here is our new Ice Lake server you should buy” but then also are saying “our new Sapphire Rapids is already shipping to customers and is a huge upgrade over Ice Lake.” This is the premise of our The 2021 Intel Ice Pickle How 2021 Will be Crunch Time, piece and video.
While the financial markets may have a different interest in 10nm ramps, our interest is when the chips can be purchased outside of NDA realms. As such, we are going to say that Intel did not launch Ice Lake Xeons in 2020. AMD is an interesting case, but we are going to count the EPYC 7003 “Milan” as launched at its public launch as well. Going forward, at STH, we are going to define a “launch” as:
The time when the average customer can configure and buy a system with a given processor, from a Tier 1/2 OEM, without any special access restrictions, and with a reasonable delivery timeframe.
Cooper Lake was a bit interesting in 2020 given there was a public launch with later availability.
For “shipping” we are going to use Pre-Production Shipping and Production Shipping. To describe that phase of launch and cover cases. This will be based on retail steppings. If a stepping is not viable for retail customer use, it is not the generally saleable offering. Instead, it is something that is more akin to a custom part/ line. A previous stepping chip sold to AWS that AWS is expected to work around bugs, we are going to treat as its own distinct custom chip line. Likewise, a genal customer is not expected to buy and use the ES Ice Lake Xeon chip we saw at Intel Architecture Day 2018. Clearly, there was a gap between when that chip was shown and when it is available in the market.
By those definitions, AMD EPYC “Milan” would be in a Production Shipping status but not yet launched. Arm Neoverse N1 products such as the Ampere Altra/ Graviton are now Launched since a prerequisite of this is going through Production Shipping, but Neoverse V1 is not.
This may seem futile to try to standardize on terminology, but the current state is untenable for customers. At STH, I need our team to be able to communicate to our audience basic milestones such as when they can expect to purchase and install generations of systems so we need to standardize how we decribe platform status across vendors.
If you were hoping for Ice Lake Xeons in 2021, the wait will be a bit longer, but hopefully, this is a milestone towards Launch.