Lets start out with what just happened, Intel dropped the price of some of their most expensive Xeons by 28%, more as you go down the stack. Before you get excited about your server prices in Q1 doing the same after a phone call to your OEM and a re-bid, lets just say this won’t happen unless you are a select few customers. Which customer? Those who bought the -M and -L large memory SKUs, roughly 2% of sales according to informal numbers given to SemiAccurate at the Skylake launch.
You might recall with Skylake era Xeons, Intel introduced the reviled crippled memory SKUs with the -M prefix. This meant if you needed more than 1.5TB of memory per socket, you had to pay a >$3000 tax per socket for Intel not to blow a fuse on your CPU. This obviously went over like a lead balloon with customers but the idea was to juice sales for the moribund Optane DIMMs which no one other than one customer SemiAccurate can find actually wanted. The kicker to all of this was that Intel took the pain of the blatant gouging but didn’t actually launch the DIMMs that took advantage of it until Cascade Lake launched. In the history of own goals, this one is right up there.
Then we get to Cascade Lake, a bug fixed Skylake CPU that came out about two years later. This time there were two ‘large memory’ SKUs, the first being -M for those who need >1TB of memory that the non-suffixed Xeons support but <2TB that the -Ms max at. This adds a mere $3003 to a base Xeon. The -L up the limit to 4.5TB, effectively the max possible with a 256GB DRAM and a 512GB Optane DIMM. For the privilege of not blowing two fuses, Intel charged customers a mere $7897, it takes a lot of work to not blow those two fuses AND add a -L to the etchings on the heat spreader you know.
Strangely enough these two models don’t seem to be selling well, and by well we mean at all. The thriving ecosystem that is taking time to mature seems to have one real customer SemiAccurate can find at the moment, the rest seem to have kicked the tires and walked away. While it is possible that we are missing a huge market for these DIMMs, none of the major players are taking it up. Today’s price cuts are not a reflection of market strength either.
If the CPU pricing isn’t enough of a nail in the coffin for Optane, think about this, Intel can’t actually cut the price on the memory at this point. SemiAccurate explained why in detail here so the only lever left is reducing the price of the Xeons. (Note: Since that article last July, DDR4 prices have increased but this seems to be a short term issue due to fab issues in Asia. Then again, DRAM is volatile and it could stay this high forever.) That is fine but those Xeons are still higher cost than the AMD Epyc alternatives which have unlocked memory capacity, are significantly cheaper, is a multiple faster, and supports 4TB of DDR4 per socket vs 3TB of DDR4 for the Xeon. (Note: The 4.5TB is for DDR4 + Optane). Slashing the price of a Xeon still puts it at a multiple of the cost of an equivalent Epyc system.
Until Intel can get Xpoint to outshrink the entire DRAM industry, you know the cutthroat guys who live on meager margins in a commodity space, this situation won’t change. Cutting the price on Cascade is a welcome thing but this memory tax is a self inflicted wound of Epyc proportions. Yes we had to go there, sorry. Back to the point when it was launched with Skylake it was sheer gouging that did nothing other than to piss off customers. Even if the ~2% of user hitting that cap was a real number or not, all it meant was that the other 98% were being told they were buying an inferior product. Would you want the crippled Xeon or the ‘real’ one? If you were a CTO or CIO, I’ll bet it is gnawing a little in the back of your head every time you look at pricing.
Then with Cascade it got even sillier with the tax going down to those who needed 128GB DIMMs in their system, a much more common occurrence when Cascade launched versus when Skylake came out. Back when SemiAccurate last researched DRAM pricing, 1x 128GB Optane DIMM was on price parity with 2x 64GB DDR4 modules. That gap grew to about 2x when comparing 256GB Optane to 128GB DDR4 and the 512GB modules were possible to reach with 256GB DDR4 but at a significant premium.
This is not counting the Xeon tax though. If you needed the -M SKUs that would add ~$250 to each DIMM slot meaning at the 256GB Optane DIMM level, DRAM is cheaper. The $7897 tax for the -L fuses equate to a little over $650 per DIMM slot which totally obliterates any advantage the 256GB Optane DIMM has over DRAM and lessens the pain at the 512GB level. In any case if you go the DRAM route you don’t have to re-write your software making the decision to just say no much easier. It also make the decision to go AMD or not much easier if you need >1TB of memory in your system.
That brings us back to the own goals. When Intel announced the memory tax, it basically pissed everyone off for no up side. Optane wasn’t launched until Cascade lake over a year later so why did they bother? It makes no sense to annoy customers at a time when a vastly cheaper and slightly lower performing competitor, AMD’s Naples Epyc in this case, is coming on the market to fight against Skylake. Upping the ante when the vastly better in every way and still significantly cheaper AMD Rome is launching is not a winning strategy. It still pisses off customers even more but this time they have a viable alternative.
On top of this the tax on Optane is just that, it drives the cost of a 2S system up significantly, between $6000-15,000 which is enough to hurt sales. When you are trying to seed a market and grow a technology that requires significant software changes to take root, making the cost painful for early adopters is not a route to success. If you look at the games Intel played with TSX for example, it’s stunted use years later is a good roadmap on how not to do things. Today’s price cuts show that Intel hasn’t really learned this lesson. That is the first black eye.
The other is the climbdown on Xeon pricing. When Intel announced the memory tax, everyone in the room knew it wouldn’t last. SemiAccurate is honestly surprised it took this long for Intel to backpedal, hence the second black eye. By ending the -M SKUs and replacing it with the -L SKUs at -M pricing, Intel has cut pricing on their most expensive models by $4894, the top Xeon 8280L is now ~76% of the price it was yesterday while the lowest end -L, the 6240L, is now ~53% the cost it was hours earlier.
So what will the headlines read? Intel drops prices on top Xeons by 24%. Sure it was an artificially high price based on an untenable memory tax, but it was the MSRP for the halo Xeon product. If you want to compare against an AMD Epyc system with similar features, this is what you have to choose from. Black eye number two right there, three if you count driving customers to the competition.
In the end this climbdown had to happen. Optane is simply not selling and Intel can’t cut the price there for all of the reasons explained in our earlier article. The only thing left was to cut the price of the base Xeon which has optics that, well, look horrid. It also reminds customers how they are being gouged, picking at scabs is never a good business strategy. And all of this was meant to protect a market that literally didn’t exist for the first generation it was out and is a rounding error now.
While SemiAccurate doesn’t think today’s Xeon price cuts are a response to AMD’s Rome pressure, you are seeing that in other ways, it is a response to Optane’s failure in the market. The only down side to today’s price cut is Intel didn’t get rid of the memory tax entirely so there is time for the eyes to heal before they are blackened again.S|A
Latest posts by Charlie Demerjian (see all)
- Intel is close to honest on Ice Lake server dates - Oct 22, 2020
- A key market shifts away from Intel - Oct 21, 2020
- Intel force bundles flash divisions - Oct 20, 2020
- Qualcomm may be launching infrastructre silicon today - Oct 20, 2020
- Intel talks about Ice Lake security - Oct 14, 2020