AMD just obsoleted Intel’s server line with their new Rome CPU. While SemiAccurate won’t make an Epyc/epic pun, the industry hasn’t seen this level of step change since AMD’s dual core Opteron almost two decades ago.
Last week AMD launched the 2nd Gen Epyc CPU which was code named Rome, a 64-core monster which SemiAccurate exclusively told you about over a year ago. A year to the day before the Rome launch, we wrote an article titled, “Intel has no chance in servers and they know it“. Some thought this was tongue in cheek, a lark to annoy some executives, or many other trolling theories. They ignored the specific numbers we put in for the then current platforms, Intel’s Purley/Sky Lake-SP, AMD’s Naples/Epyc, and several future generations.
Stepping forward a year or so there have been two releases that hit the numbers predicted in our article, Cascade Lake/Xeon x200 and Rome. Both were dead on which makes you wonder how good our, “made up fantasies” were for the next three, Intel’s Cooper Lake and Ice Lake plus AMD’s Milan. And do note we specifically pointed out those numbers were, “the worst case for AMD vs the best case for Intel, fudged a lot in Intel’s favor on top of that.” Last week the laughing seems to have died down. Why? Performance, price, and features mainly, all of which we will look at now.
AMD’s Rome lineup is pretty simple 14 2S SKUs ranging from 8-64 cores with no more than three models at any given core count. All models have all features enabled, no games, no abusive pricing to turn on features arbitrarily turned off, all in for between $475 and $6950. Compare and contrast this with Intel who at launch had 41 public Cascade SKUs ranging from $275-$10,009 with a veritable minefield of abusive price gouging ‘options’.
A good example of this is memory, if you want to have the full capacity of the socket, buy an Epyc at MSRP or buy one of seven Cascade SKUs and add $7879 for the -L suffix. If you don’t think this is laughable, consider that on the top $10,009 8280 SKU this is about an 80% gouge but on the lowest end $1221 5215 you pay >6x the cost of the CPU to get what AMD offers out of the box. Unfortunately the closest AMD Rome to the 5215 has 1.6x the cores, significantly higher base clocks, 100MHz lower boost clock, and costs $243 less. Any questions so far? If not consider that the memory unlock upcharge alone on a Xeon will buy you the highest end Rome and still leave you with $800 for LEDs and case fans.
With that in mind, lets take a look at what AMD has to offer. There are the aforementioned 14 2S SKUs and five 1S SKUs with a -P suffix. The rest is as it looks, you get what you pay for, no games, no hidden gotchas, no artificial crippling via fuse to sell back later. It just works. And if you want things like Intel’s excellent Service Assurance Administrator (Basically cache QoS) you tack on thousands per socket for the software, AMD’s version is free. In short the price disparity only gets worse for Intel when you look at the details.
AMD’s 2nd Gen Epyc SKU table
So there you have it. And to recap, here is the Intel Cascade launch table, there have been more added since then, most notably the 240W, $15,460 8284 which doesn’t show up on ARK unless you search for it and even then doesn’t list the MSRP. Also do note that the 9200 line of CPUs isn’t really a CPU line, it is an expensive PR stunt that Intel refuses to disclose price for.
Intel’s Cascade SKUs (At launch)
Compare and Contrast:
So now you know the basics, lets move on to a bit more advanced analysis. AMD did some nice work in comparing Rome to Cascade using SpecInt rate, a good benchmark for many enterprise workloads. These charts will move a bit if you use other benchmarks like SpecFP or SpecJBB, tweak the settings using trick compilers like Intel’s ICC, or play similar games. What doesn’t change is AMD’s lead of 50-200% in anything but a few pathological cases. If you consider an AMD Rome core to be ~= to an Intel Cascade core at the same clocks, you won’t be far off the mark.
AMD Epyc 2S vs Intel Xeon 2S
As you can see from the chart above a 32C Epyc, probably the 7452, beats the best of Intel’s Cascade line by a substantial margin. The Intel 8280L (Note that we will use the L/M variants from here on out because the crippled -nothing parts are not comparable to AMD’s Epyc line in our eyes) is a $17,906 part where the 7452 costs $3400. And has more features like 8-channel DDR at higher speeds, PCIe4, more than 2x the PCIe lanes, etc etc.
On the down side the 7452 consumes 20W more to do so. Depending on your TCO calculations, usually $1-3/W/year, this could add almost $300 to AMDs tab reducing the price differential to a mere $14,506 per socket. Do realize the performance delta tips things back toward AMD but we are trying to throw Intel every bone we can. Now lets look at the 1S situation.
AMD Epyc 1S vs Intel 1S
As you can see, AMD’s claim that a single socket Rome is faster than a 2S Cascade system is on pretty solid ground. A single $2300 7502P beats anything Intel has to offer and a $4425 7702P doubles that lead. Once again we will include the caveat that this number will vary with the benchmark used but it won’t change the general trend. This is abusive but it does get worse for Intel. Lets now look at a 2S Xeon system vs a 1S Epyc.
AMD Epyc 1S vs Intel 2S
Yes you are reading that right, a single socket AMD Epyc 7742 beats a dual socket Intel 8280L system, not by much but it does win. The Intel box does have more memory capacity, if you use the -L parts but not if you use the -M or -nothing, and does have higher memory bandwidth either way. If you need high memory bandwidth as your primary performance parameter, a dual socket Intel box does beat a single socket AMD Epyc. Then again you can always add a second socket to the Epyc system but that changes the cost by quite a bit.
Those AMD 7742 CPUs cost $4425 for the single socket 7702P and compare to 2x $10,009 or 2x $17,906 for the Xeons. This is more than a 4x price advantage for AMD if you take the crippled memory Xeon, ~8x advantage if you compare like for like on features. That however is before you take in to account the added costs of a dual socket system, DIMMs, PSU, board, rack space, power, and other TCO ‘gotchas’. Calling AMD’s TCO advantage here a high single digit multiple is entirely fair and higher multiples are quite defensible.
When you look at dual socket systems the same numbers hold on the Intel side but AMD’s numbers get a bit worse because a 2S capable Epyc 7742 costs $6950, a mere 30% price advantage for Epyc using the best case numbers for Intel. Then again the AMD system is about twice as fast as the Intel one in most benchmarks so as long as you only look at purchase price, AMD only wins by a lot. When you look at performance per dollar things get abusive and real TCO comparisons mean Intel has, wait for it, no chance.
All Together Now:
Cascade and Rome in one painful chart
In the end we get to the painful bit, combining the AMD and Intel price charts. We have removed the Intel -M and -L variants for sanity so add $7879 to the Intel prices if you want a fair comparison. Also keep in mind that the top priced Epyc 7742 beats the top priced Xeon 8280 by a large margin and absolutely trounces the comparably priced 8256. Unlike other recent releases this performance delta carries on down the stack, it isn’t just cherry-picked SKU comparisons. Intel has absolutely no answer to AMD’s Rome line up, period.
So there you have it, Intel has no chance in servers and they know it. We wrote that a year ago, backed it up with specific numbers for the now current generation and several generations of future products too. We meant it and the chips released since then back up what we said exactly. While Intel has made some changes to their 2020 line of server products since that article, (1, 2, 3) we don’t think the basics will change, Intel will just lose by a little less for a lot more OEM pain. As we said last year, “One of the documents says in no uncertain terms that the company understands they will not be competitive in the server market until AFTER Sapphire Rapids, the 2022 server part.” SemiAccurate meant it.S|A