Intel Ice Lake-SP: Xeon CPUs with 40 cores are faster and cheaper

With Intel Ice Lake-SP, Intel is now offering Xeon CPUs for dual-socket servers with up to 40 cores, a new eight-channel memory interface and 64 fourth-generation PCIe lanes per CPU, which thanks to AMD Epyc are also significantly cheaper fail. The overview of the editorial team at the market launch shows models and innovations.

Intel Xeon will be much faster – and significantly cheaper

40 cores in full expansion with all options for 8,099 US dollars before taxes (RRP): This undoubtedly marks a turning point in Intel’s price strategy for server CPUs, which recently consistently exceeded 10,000 US dollars in the official price list for the top models despite growing competitive pressure. Dollars remained. But the pressure from AMD now also seems to have become too great in the server area.

For comparison: the flagship of Cascade Lake-SP with 28 cores and support for 4.5 TByte RAM costs 17,906 US dollars when it was introduced two years ago, last year it went down to 13,012 US dollars with Cooper Lake-SP. And now, despite more cores and new features, it’s another $5,000 less. AMD’s Epyc 7763 with 64 cores is still below that at officially $7,890.

24 cores for $ 1,273, 36 cores from $ 3,017

Not only in the absolute high-end area, although the prices have also fallen significantly there, also and especially just below that it becomes clear that Intel is accepting the price challenge posed by AMD.

New Xeons with 36 cores for just 3,027 US dollars in the form of the Xeon Platinum 8351N are sure to be a positive surprise for buyers, but 24-core models for just 1,273 US dollars such as the Xeon 5318Y also belong in this category. AMD’s new Milan processors cost at least as much here. In almost all areas, their prices are now met at least at a very similar level.

Model range at a glance

The 40-core has long been a secret, Intel announced in advance in a press workshop. Pessimists say that Intel cannot pull enough fully functional chips from the wafers for this, because it is undoubtedly a large caliber. Only a single SKU will offer 40 cores today, which underpins this skeptical assessment.

Intel has launched various SKUs as the Xeon 8300 series, they are aimed at very different areas of application. In addition, there are custom solutions and so-called off-roadmap processors, which are not put on the big bell, but can also quickly identify dozens of other solutions in the end. The official list is therefore primarily intended for regular OEMs such as HPE, Dell or Lenovo, who from today can equip their servers with it and ultimately be purchased by anyone. The TDP range extends from 270 watts down to 105 watts, whereby the smallest CPU in the form of the Xeon Silver 4309Y can also be configured as an 85 watt solution.

The basic designation is almost identical to that of the already existing Xeon 8300 H / HL, which stands for Cooper Lake-SP in 14 nm with its alignment for four- and eight-socket systems. The new 10 nm Ice Lake-SP models will only be specified for one and two sockets, but will run on the same LGA 4189 socket paired with the same chipset – the Whitley platform.

At the end there is a bevy of additional abbreviations that are intended to specialize models for many areas. But this approach is confusing, because some CPUs offer the features of three abbreviations or even more, Intel in turn only picks out the most important from the company’s point of view for the name. For example, S trumps Y in the name, but in this case does not mean that Y, the Performance Profile 2.0 from Speed ​​Select Technology, is not integrated in the S-SKUs – it is integrated even if Y is not on it.

But it can get worse. There are four variants of the Xeon Platinum 8352 model, one of which has 36 cores and three only 32 cores. At this point, at the latest, the confusion is at its maximum. It is therefore essential to consult the exact technical documentation for the chosen solution before a possible purchase.

Intel Xeon 8300 aka Ice Lake-SP

Architecture overview

Intel Ice Lake-SP uses the same technical basis as Ice Lake-U for notebooks. For this purpose, new processor cores with the code name Sunny Cove were launched many years ago, which were actually ready for series production at least in 2018, but had to wait until mid-2019 before they were allowed to debut in the notebook.

Replacement after four years of Skylake

At the time, Intel called the approach behind the new CPU architecture with a view to the server segment “Deeper, Wider, Smarter”. With Intel Ice Lake-SP, more execution units, larger caches, but above all the modification in the front end, which is particularly important for the high-performance segment, is now being used in series production. The bottom line is that Intel is now even promising a 20 percent increase in the IPC compared to its predecessor, and up to 19 percent in notebooks and desktops. Coupled with the more cores, 50 percent more performance is made available compared to the previous platform.

In addition to the new CPU cores, it is the additional features that allow Intel Ice Lake-SP to take a step forward. First and foremost is the new eight-channel memory controller for DDR4-3200, followed by the PCIe 4.0 interface with 64 lanes per CPU, in the typical dual-socket system there are consequently 128 usable lanes – with Intel, communication does not take place via PCIe. AVX-512 is still an essential unique selling point of the Intel platform, from now on all CPUs will always offer two FMA units.

The featured in the fall of new security features are TotalMemory Encrytion and for the first time in the upper class SGX on the program. According to the manufacturer, this is the most researched, updated and “battle-tested “security solution. In the past, this often included serious security gaps, which Intel of course does not mention today, but which are likely to have played their part in the further development. The feature that was previously only used in the small Xeon E is now offered in all solutions. Under the motto “Encrypt everything”, user data should be safe in every situation, explains Intel.

Intel has not done anything with the chipset. It is still the old Lewisburg as the C620A series, the three models C621A, C627A and C629A are primarily used for Ice Lake. These are still connected to the CPU via four DMI 3.0 lanes. Communication between the CPUs is still controlled via three UPI links, but Intel has given this a small upgrade from 10.4 to 11.2 GT / s. However, as usual, this does not apply to the very smallest CPUs, which are sometimes only operated with two UPI links at 10.4 GT / s.

The clock speeds of the 10 nm solutions drop

Much was speculated in advance what the problems of Ice Lake-SP could be. After the appearance of the U series, these were obvious: The new 10 nm production brought a penalty in the clock rates, which was only ironed out optimizations at Tiger Lake. Inter Ice Lake-SP is better positioned than Ice Lake-U, but the clock rate drop is still large: When comparing with Cooper Lake-SP with the same number of cores and threads and a fairly similar TDP, the new CPUs show single-core clock speeds of at most 3.7 GHz, a maximum of 3.6 GHz is approved for all cores.

And even small SKUs with few cores will not let Intel off the beat with Ice Lake-SP. With the Skylake architecture in 14 nm, there was still up to 4.4 GHz in the server.

The reason is really primarily that Intel Ice Lake-SP uses the same production status as Ice Lake-U, although this has developed a little further and has been optimized in the last two years. Nevertheless, this is located one step below Tiger Lake, where 10 nm Superfin is already used – and thus the lost tact was back. This is once again due to the extreme delay in the 10 nm technology, because Ice Lake in the server should also have been on the market for a long time.

Manufacturer benchmarks: burins against AMD

A few weeks ago it was AMD’s turn with Milan against Intel’s Xeon, today it’s the turn of the market leader again. As usual, various manufacturer benchmarks are included, which usually apply, but have been specially selected and ultimately the cherry on the cake. This is exactly how they should therefore always be perceived.

Intel’s presentation creates a little deja vu feeling, memories of Skylake-SP against AMD’s Naples are awakened. At that time, dismissed by Intel as “glued together ”, that is, glued together, dismissed, however, the competition is now met more with facts. Interestingly, Intel was able to carry out its own tests on a dual-socket AMD Epyc 7763 “Milan” with 64 cores each, the competitor’s new flagship.

But not only the negative things are brought out and of course brought into focus, in some places even Intel shows how good AMD is, for example with the cache latency, especially in the immediate vicinity. The fact that at Epyc there is a significant deterioration at the latest when accessing the L3 in the worst-case scenario is again homemade through multiple chip sets and their detour via an I / O die.

The list of great advantages, however, can almost be summed up in one word: AVX-512. Various benchmarks mentioned use at least one or more of these instructions, plus DL Boost for certain AI skills. Intel wants to make it clear that the Xeon platform stands for more than just pure parallelization of cores, adapted features and suitable software should offer greater added value for customers.

And this is ultimately also visible in your own home, thanks to more cores, more DRAM channels and DL Boost, the Xeon Platinum 8380 is often clearly ahead of the predecessor Xeon 8280. In special environments, especially in the middle class around the Xeon Gold, this can turn out to be significantly higher.

Assessment

With Ice Lake-SP, Intel’s new CPU architecture after notebook (Ice Lake, Tiger Lake) and desktop PC (Rocket Lake) is also finding its way into the server and, like in the notebook, production takes place in 10 nm. But that’s not all Ice Lake has -SP of the Rocket Lake-S series, in contrast to the desktop PC, there are more cores in the server. Without a doubt, the performance of the 1- and 2-socket Xeon platform increases significantly – it depends on AVX-512, memory bandwidth, PCI Express 4.0, or AI algorithms supported by DL Boost, even very clearly.

How competitive 40 Sunny Cove cores are against 64 Zen 3 cores can still be shown in independent tests, but the duel that is essential for the market takes place a few classes below. Because just as Intel’s “normal customers” don’t buy a Xeon Platinum, they don’t use a 64-core from AMD either. And in the golden mean, there is definitely a duel at eye level on the data sheet.

In addition, the adjusted prices, which are well known only in the list and are often undercut when buying, will help Intel that AMD continues to nibble off only small amounts of the market share. In the important price range of 1,000 to 3,000 US dollars per CPU, Intel has positioned significantly more powerful CPUs with Ice Lake-SP than a generation before. That should depress the margin of Intel in this area and the shareholders will fall less, but Intel can bridge the time until the coming solutions. And these come afloat as Sapphire Rapids, Ice Lake-SP is not predicted to have a long life expectancy.

According to Intel, delivery has been in full swing for a few weeks, over 200,000 units have already been shipped, and more than 250 designs from over 50 partners are ready. Since the introduction of Skylake-SP, Intel has sold over 50 million units of Xeon SP. Intel Ice Lake-SP will have to stretch mightily to leave a footprint.