Cloud Computing, Enterprise, Hardware

AMD Server Strategy Update – HSA and ARM Show Up in 2014

Update: AMD informed us that the ARM CPU Seattle actually supports up to 128GB RAM and provided us with an updated slide. Also they released updated roadmaps on their investor relations website, which we incorporated to give a better overview of the timeline of these products.

Today AMD is announcing an updated server roadmap for 2014, which will introduce a number of new products. AMD will attack the market with new multi-socket x86 CPUs, a HSA-enabled server APU and the debut of their first ARM-based chip. In the following paragraphs we will discuss each of the innovations AMD has in store.

Berlin CPU/APU ? Kaveri going into servers

While the client products of Kaveri are scheduled to enter the arena towards the end of 2013 as AMDs public roadmaps show, in the first half of 2014 the company moves their first HSA-enabled APU into the server space. The differences between the client and the server product will be marginal. The server variant called Berlin will support ECC memory, while Kaveri ? the client version will not. The other features are pretty much in line with the expectations on this product.

The two remarkable improvements are the next iteration of AMDs high-performance x86 cores called Steamroller which allegedly brings significant IPC gains to the table as well as the addition of the HSA. AMD claims a 7.8x improvement of GFlops per watt over the Opteron 6386SE (16 Piledriver cores, 2.8GHz) which is possible thanks to the GPU portion. The HSA will allow easier access to the GPU for programmers due to uniform memory access. Similar to the Opteron X-series, there seem to be different products SKUs with and without the GPU enabled. We expect AMD to position them at different clock speeds, TDPs and pricing.

Also, the chip will be the first from AMD to integrate PCI Express 3.0. We actually expect the client version to have the same functionality. During the briefing Andrew Feldman, who leads the Data Center Server Solutions team at AMD, pointed out that their current I/O hubs are connected via PCIe 2.0 but there is the possibility of an update in a future product. While the chipsets for Berlin were not detailed we expect AMD to come up with variations of the current FCHs. However this allows to interface with AMDs GPUs, which integrate PCIe 3.0 since the end of 2011 at a faster speed.

Warsaw ? updated Piledriver-based 2P/4P CPUs

Servers with 2P and 4P configurations get an update in the first quarter of 2014. AMD will introduce a successor to the current Opteron 4300/6300 series. While we initially believed this would introduce Steamroller to the server space and have some implications on a possible release date of desktop FX processors with Steamroller cores, unfortunately this is not the case. Andrew Feldman explained to us that the introduction of Steamroller would have moved this product further down the road, a decision the company didn’t want to make.

Still AMD claims improved performance/watt over the existing families. What we expect here is efficiency optimizations similar to those we observed in the Richland client APUs, where AMD was able to significantly lower operating voltages. The effect of this is multiplied as the number of modules rises so the power savings and efficiency gains could be significant compared to the previous products.

Also the age of the platform we discussed in the Opteron 6300 launch article is not changed. We believe AMD should invest a little bit to update their chipsets to a more current featureset. Their northbridge could use a PCI-Express 3.0 update whereas their southbridge could use some SATA 6.0Gb/s and USB 3.0 treatment. Oddly enough their current server chipsets are not even linked from their product pages anymore.

The small amount of changes to the platform of course also has benefits. The Warsaw CPUs are socket-compatible with current offerings, so they are basically a drop-in replacement in current socket C32 and G34 platforms. Consequently any software certifications that might be important for some companies are retained. AMD also focuses on their Open 3.0 server designs where the server design is actually open source. The Open Compute Project is actually spearheaded by engineers at Facebook who wanted to optimize their servers in various ways.

Seattle ? AMDs ‘no compromise’ entry into ARM servers

The most exciting server product AMD is going to introduce in the next year is their first ARM-based product. Based on the rough list of specs, AMD aims at introducing a killer part which can be regarded as top of the line in many regards. For starters AMD bases it on the latest 64-bit Cortex-A57 cores and aims at clock frequencies of 2GHz and higher. The chip will be made using 28nm process techology. They will start off with 8 core SoCs with 16 core variants planned a little bit later. Those SoCs support up to 64GB of RAM. Interestingly AMD claims 2-4x the performance of their own Opteron X-series with improvements in compute per watt.

But the impressive specs don’t stop at the cores, because the SoCs will integrate additional functionality. For compression and encryption AMD adds in offload engines that allow these operations to be performed more efficiently. AMD also integrates 10GbE for networking as well as their Freedom Fabric for dense systems. This allows flexible interconnection of many of these SoCs. While nothing concrete is announced at this time, you can also bet on these being used for some future system of AMDs SeaMicro division. Last but not least a high port-count storage interface is integrated as well. We wonder what AMD considers a high port count here.

AMD is actually very open about the fact that there are other companies out there who have some head start in terms of experience with client products based on the ARM architecture. However, AMD believes they have unique strengths in the server space where they can count on a lot of experience. Also they bring lots of IP blocks to the table that are currently not found in ARM designs. In the long run AMD is confident that their ARM strategy will pay off given the momentum around the technology. While sampling is scheduled to begin in the
first quarter of 2014, the actual product introduction will have to wait until the second half of next year.

The new roadmap also makes a clear distinction between the different segments AMD tries to address with their new products. While their recently launched Opteron X-Series and the upcoming ‘Berlin’ server APU and the ‘Seattle’ ARM CPU are aimed at web/enterprise services and cluster workloads, their big iron 2P/4P Opterons cater to traditional server markets. The former focuses on performance per watt per dollar and/or compute density per dollar, the latter on performance and/or throughput per dollar. It is important to realize that these are different dimensions to optimize silicon designs for and that each has strengths in different applications. Interestingly enough, AMD considers the ARM-based SoC the successor to the current Opteron X-series. The rough performance claims AMD gave us move Seattle and Berlin into a similar performance space. Needless to say that this will be interesting to watch.