What To Expect From AMD's Threadripper And EPYC Processors.

Summary

• AMD's EPYC 7601 has been reviewed, and it is a competitive product.
• Threadripper is priced competitively, but it is going after a market that has not been established.
• The server market is very risk averse, and very slow to change, so AMD growth will be slow in this market, but does seem likely.

n the past few days, more information regarding AMD's Threadripper and EPYC Server chips have become available. Although neither are high volume markets compared to the PC market, the margins on these products, particularly server, is very high. Considering how little AMD (NASDAQ:AMD) currently has in each segment, this is a growth opportunity for the company. Or is it?

AMD is again trying a tactic that did not work with the Bulldozer line on the desktop, trying to downplay the importance of single-threaded performance. At first they stated it was really multi-threaded performance that mattered. When this failed, and the Bulldozer line was found wanting, the company tried to convince everyone that APU processing (via HUMA) was really important. It failed to convince. Now AMD is again focusing on multi-threaded performance with the Ryzen line, but at least the single-threaded performance is significantly better than before. Oddly, due to AMD's gimped design, the processor can have performance problems when moving beyond four cores that will impact its acceptance in certain workloads.

This has to do with the L3 cache and the way AMD made the chips. There are four cores per CPU Complex (hereafter referred to as CCX). Each CCX has 8 MB L3 cache, and each of the four cores connects to a 2 MB L3 cache directly, and hits it full speed. Hits to the other 6 MB within the CCX is only slightly slower, but still fast.

Any processor beyond four cores is not monolithic. For example an eight-core processor has two CPU Complexes (CCX), and two different 8 MB L3 caches. Whenever a processor from one CCX needs to access the other CCX, it goes through AMD's "Infinity Fabric" comes with comes with a large latency penalty.

But, these are actually on the same die (called Zeppelin), whereas anything larger than eight cores is on separate dies, though on the same processor module. So, moving off die adds even more latency.

Updated 07/18/17 2:58 PM:

Why did AMD choose this design? Well, it's not necessarily wrong, just a design choice, and they didn't have the resources to do it both ways like Intel (INTC) did (depending on core count). For its four core and below processors, Intel chose a ring bus as this could effectively handle four cores, whereas as one added cores, it became less efficient. For the server line (and the higher end desktop parts which are brought over from the servers, which have more than four cores) Intel chose a fabric architecture that has a higher latency L3 cache than the four core processors, but handles more cores well, with latency being pretty consistent as one adds more processors to that design; L3 accesses are also fairly consistent regardless of where it is.

AMD did not have the resources do to a complete redesign, and rather than extend the ring bus to eight cores, created it as mentioned above - four core processors connected through a CCX, and if more than eight, different dies on the same packaging. It gives nice performance for "near" L3 access, better than Intel's, but "far" access to L3 cache is so poor it is barely better than DRAM access.

Most of the world doesn't need more than four cores anyway, particularly since AMD added "hyper-threading" to its line, which has eight logical processors. AMD's releasing all these high-end parts gives the false impression these are more important that lower core count processors. But the reality is, the company will make most of its money from four cores and below, and so will Intel.

Also, although Intel processors scale much better, they are also much more expensive to make since they are one massive die. Intel's fabric architecture adds additional latency to what it had on the old Broadwell processors, although it scales far better.

In short, AMD's L3 cache is very effective if it's on the same CCX (generally 8 MB), but significantly worse past that. Four core processors, and below, will not have any of these issues. Even where this is an issue, it does not render the processor useless, but there is a performance hit that generally will make it less than ideal for that workload.

So now that we have that issue behind us, we should look at the two markets being addressed by the Threadripper and EPYC lines.

Threadripper will sell into the High-End Desktop (HEDT) market. This market is much less resistant to change than the server market. Of course, it's a lot less profitable as well, but AMD has a pretty good chance here, unless Intel again changes its pricing.

The current list price for AMD's 16-core Threadripper 1950x is $999, and the 12-core 1920x is $799. For reference purpose, we'll note that the Ryzen 7 1820x 8-core has a list price of $499, but sells for $419. Intel's 16-core i9-7960x sets one back $1699, whereas the 12-core i9-7920x sells for a cool $1,199. Below them is the 10-core i9-7900x which goes for $999, and $599 for the i7-7820x.

There are a few challenges here. AMD has clearly lost the 8-core battle - it's $499 chip is selling for $419 on Amazon, whereas Intel's $599 i7-7820x is selling for $679, or I should say sold for $679, since it seems to be sold out everywhere. This is not a fluke, it's better, and worth more than AMD's chip. But, this is also probably the sweet spot for Intel, and I personally think the most attractive processor above their "mainstream" i7-77xx line.

If we assume the i7-7820x will settle near $599 as Intel intended, AMD has an uphill battle. AMD simply has no processor that can match the single-threaded performance, at any price. So, it's got to convince people they need more than 8-core/16-thread processors at a time when only a small market is convinced it needs more than 4-core/8-thread processors.

Of course, there are scenarios where one can actually use 12 or more cores (outside of servers), but that's a really limited market. Intel's pricing kind of falls apart after the i7-7820x though, and AMD can really show some value. You could pay $1,000 for an Intel 10-core, or $700 (assuming similar discounting to what we see now for Ryzen 7 1800x) for a 12-core Threadripper. Assuming I've identified I need the cores (if not, I'd go with the i7-7820x), it's likely the performance between these processors will be pretty close to equal, and I could walk away with $300.

That's still not broadly decisive though, but if the $300 didn't matter much, there's the 16-core 1950x which would surely show much better multi-threaded performance in virtually all scenarios. In my opinion, that's the biggest value in the line, the 12-core 1920x is overpriced at $799, and will have to be deeply discounted to sell well. The problem is it's using the same dies as the 1950x, with disabled cores, so AMD may not be able to discount it much. The 1950x is a real bargain though, with some very real competitive advantages over everything above and below it. But, it's not a huge market at all, and that's the biggest challenge for AMD to try to expand that market. It has not been convincing in the past, and even if it is, how will Intel react?

Servers where the EPYC line is intended are another opportunity for AMD. This is a very resistant market to change, and it's going to be a slow process. But, it's a very lucrative market, and AMD finally has a competitive product, especially considering Intel's hideous pricing. Anandtech did a review comparing Intel's 28-core Xeon 8176 ($8719) processor to AMD's 32-core 7601 EPYC ($4200) processor. Sadly, on single-threaded performance, the 7601 only had roughly 95% of the performance compared to Intel's Xeon E5-2690, if one averages all the results. This is a processor five generations removed from current, with the architecture first released in 2011. By contrast, the Xeon 8176 from Intel averages to 129% of the speed of the old Sandy Bridge based E5-2690. That's quite a difference. Oddly, the author states "we can conclude that the single threaded performance of the 'Zen architecture' is excellent," despite it not equaling the performance of a chip five generations old and being woefully behind the current Intel processors in this metric. Excellent is just not the right word. I'd say poor is correct, which is up from abysmal for the previous generation from AMD, so it's solid progress.

Even so, despite the author's odd interpretations, the benchmarks are revealing, so I recommend reading the article. Truly, the gimped architecture as described above does leave it disadvantaged in some described scenarios, but outside of them the multi-threaded performance is quite close to Intel's 8176, and in some cases surpasses it.

Where AMD's processor thrives is floating point, and it beats Intel's pretty decisively. Even allowing for AMD's four extra cores, it's a clear cut victory for AMD.

So, let's summarize. The 7601 has quite poor single-threaded performance. It will be fighting entrenched Intel systems in a market that is resistant to change. It has uneven performance due to the way the processors are connected to L3 cache. Multi-threaded performance is roughly equal to Intel's, with some wins and some losses, albeit with four additional cores. The cost is MUCH less than Intel's 8176. Floating point performance is significantly better.

Put it all together and you'll get very little uptick this year, as development/validation times for servers is quite large, and deployments very slow. This is also a very risk averse market, so customers make changes very slowly, as even a company wanting to make the change would do a very limited rollout to become comfortable (or not) with the new systems. It should also be noted that Intel's next generation of processors will start with servers, rather than the desktop market like it typically does, and this will almost certainly change the competitive landscape unfavorably for AMD. So, expecting large short-term gains is unreasonable.

But, it is a far more competitive product than AMD has had in a long time, and offers significant advantages in some workloads(particularly floating point). It also offers significant price advantages, and although this isn't a very price sensitive market, the price differences are large enough it will create some wins for AMD. There can be no doubt AMD will gain market share, but it will be very slow, and at least in the near to mid term, and not a lot. But, having only .5% of the market now, and given the profitability in this segment, in a few years this could show significant improvement over where the company is now.

One last point thing to think about is the OpenPOWER foundation is just beginning to take off, and the POWER9 processor will soon be out. Unlike AMD and Intel, which make desktop processors and then tweak them for servers, POWER processors are designed from scratch to be in servers, and do extremely well in many server workloads. Companies interested in moving from Intel will likely also consider these, although it's a bigger change than going to AMD processor by virtue of the instruction set. Even so, this is a sleeping giant, as IBM knows all about selling big iron, and they have help from some pretty formidable friends like Google (GOOG) (GOOGL) and Nvidia (NVDA). On the plus side, ARM servers seem no further now than they did a few years ago, so it should be a three pony race. AMD might not have the best pedigree, but it's finally part of the race, and what it's selling is sure to gain traction, albeit slowly, against a company that has been competing with itself for quite a few years.

Analyst’s Disclosure: I/we have no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.