AMD dreams of being the king of hybrid components for servers

It’s been 5 years since AMD returned in force to the field of consumer processors with its Ryzen which turned the market upside down and put Intel in difficulty. The objectives then set have been achieved and the company must now drive the point home.

AMD and Intel are no longer alone in servers

And it intends to do so, multiplying announcements in recent months to convince its partners and investors: Zen 4 is coming, new Radeons will follow. The growth of the next few years will come in part from its efforts in the field of servers, a growing market, with an estimated potential of no less than 42 billion dollars in the long term.

The server market, an opportunity that remains to be seized: 26 billion dollars in the clouds

And if the share won by EPYC processors has never been as strong as it is now (around 12%), they remain a minority despite all the work done by AMD teams. The inertia of this sector is stronger and the ultra-complete ecosystem of Intel has played a lot in its favor, both at the hardware and software level. Pat Gelsinger, new boss of the Santa Clara giant, has also proclaimed it loud and clear: he intends to not let it go and will come back to the score.

But for the two longstanding competitors, the threat is now elsewhere: ARM solutions are gaining momentum, with high energy efficiency, very good density and a certain aggressive pricing; strengths that were until now those of AMD against Intel.

We thus see companies like Ampere Computing gaining ground, with NVIDIA in ambush working on its own CPUs. At Computex, the latter was clear: by 2024 he intends to position his Grace (Hopper) as alternatives allowing him to fulfill the mission set by Jen Hsun Huang: reinvent the datacenter.

Just like Intel, the father of the GeForce understood that it is with a complete offer that one convinces the professionals. In addition to its CPUs and GPUs, it has been campaigning for years for the adoption of its next-generation network solutions (DPU) to deport certain calculations and lighten the central processor. It now offers its own turnkey switches, servers and even datacenters. All with a good dose of software ready for (almost) any situation.

A new era for EPYC, that of variations

It is to all these threats that AMD must now respond, and not go in a few years from the ambition of getting closer to number 1 to that of not finishing on the podium.

A battle plan was therefore drawn up, with a strategy at several levels. The first consists in reviewing the offer of EPYC processors to adapt it to a greater diversity of needs. Finished the strategy one die fits all From the beginning, we stick to the needs of customers and the competition.

Zen 4 will thus be available in four segments. First of all the classic EPYC processors (Genoa) which are expected for the fourth quarter, with classic and full cores. As a reminder, they will be up to 96 (compared to 64 on the EPYC 7003), engraved in 5 nm with support for DDR5, PCIe 5.0, Compute eXpress Link (CXL) and instructions such as AVX-512 or to accelerate the AI.

AMD mentions an improvement in the IPC of 8 to 10% over this generation, but performance up to 35%, and 25% on the energy efficiency side (for a Ryzen 16C / 32T). Evolutions allowed by the improvement of the fineness of engraving and frequencies.

Zen 4 will have twice as many variations… and Zen 5 (Turin)?

In a second step will come the “X” version with 3D V-Cache to improve performance in very specific cases, which already exists on the current offer. In 2023 we will discover Bergamo (Zen 4c), AMD’s answer to ARM chips: up to 128 cores can be used in the same socket. They will therefore be reduced, even if the manufacturer promises that they will manage the same sets of instructions, avoiding different levels of optimization over the same generation.

This conference was finally the occasion to discover Siena, code name of processors which must target the telecom and embedded market. They will thus be limited to 64 cores, more accessible, with a focus on the performance/watt ratio. An interesting strategy with the rise of “edge” uses but also of “software defined” network solutions, operating on conventional processors, particularly around 5G.

Unify the offer after an external growth phase

These choices take on all the more meaning when viewed through the two recent takeovers of AMD: Xilinx, a giant of FPGAs and turnkey AI solutions, then Pensando, which offers Smart NICs and other DPUs. in order to speed up more and more things directly through the network.

From now on, AMD no longer offers only CPUs and GPUs, but an ecosystem

This is enough to open up markets to the manufacturer that were previously closed to it, by offering a very integrated offer where each brick is designed to work with the others, as its competitors Intel and NVIDIA are also preparing.

It now remains to be seen who will have the most convincing offer, but above all the best execution. Because having an interesting offer on paper is one thing, you still have to unite customers, developers, be selected and win big contracts. Everything remains to be done, especially since the local sales force is not really AMD’s strong point.

The future will be hybrid, MI300 will be an APU

Last part of the strategy and not the least: the hybridization of architectures. This too is a movement that we see elsewhere, but which seems to constitute a clear market trend for the years to come: CPUs, GPUs and various accelerators will no longer be mixed only within the motherboard, but also through the different chips interconnected in the same SoC.

AMD says it is already ready for this little revolution, since it already offers processors made up of different chiplets and arrives at the fourth version of its in-house interconnection (Infinity architecture), with a cap that should allow it to extend to the CXL 3.0 and UCIe standards.

MI300: Zen 4, CDNA 3 and HBM are in one package, is the competition falling apart?

We can thus imagine seeing tomorrow components with CPU, GPU, AI accelerator and why not chips manufactured by third parties, side by side or stacked (3D stacking).

Until then, a first test will be conducted with MI300. This computing solution for servers will be an APU engraved in 5 nm, composed of a CPU (Zen 4) and a GPU (CDNA3) accessing the same HBM memory, all within the same packaging.

AMD promises eight-fold performance over an MI250X in AI-like workloads, with a five-fold performance-to-watt ratio between CDNA 2 and CDNA 3. See you next year to find out what it’s all about. will actually return.