AWS is now shipping its Graviton5 server processor in M9g and M9gd instances, and the chip is not what the company showed at re:Invent in December. Instead of one large die with 96 core pairs, Graviton5 links four chiplets, each carrying 48 Arm Neoverse V3 cores, into a single socket with 192 cores, twelve DDR5 memory controllers and eight PCI-Express 6.0 controllers supporting the CXL 3.0 memory extension protocol.

The design borrows directly from Arm Holdings. AWS appears to have taken the Poseidon Compute Subsystem that Arm built for its own AGI CPU, trimmed each block from 64 cores to 48, and joined four of them with Arm's die-to-die interconnect. That interconnect runs at 420 GB/sec per link, and four links tie the chiplets into one virtual processor.

The chiplet route is a yield decision. A smaller die yields far better than a monolithic design pushing against reticle limits at Taiwan Semiconductor Manufacturing Co, so each chiplet costs less even after the interconnects burn extra energy. AWS also moved from the 4-nanometer process used for Graviton4 to a denser but more expensive 3-nanometer process. The trade nets out: the editor estimates Graviton5 draws roughly 650 watts, about half the performance per watt of Graviton4, but delivers 2.4 times more performance per socket.

Why four chiplets instead of one big die?

The economics of lithography drive the answer. Large dies near the reticle limit fail more often, and failed dies raise unit cost. Four modest chiplets recover yield, and the D2D links restore the appearance of a single processor to the operating system. The approach also lets AWS scale cache. Graviton5 carries 2 MB of L2 cache per core, for 384 MB total, plus 384 MB of L3 cache — twice the L2 figure — which grows faster than core count relative to earlier generations.

The power penalty is deliberate. Databases and agentic AI workloads need low latency more than they need low heat, and AWS accepts the wattage to get responsiveness. Compared with a two-socket Graviton4 NUMA node — which AWS built for the first time with that generation, pairing two 96-core chips to reach single-node scale — a single Graviton5 delivers 25 percent more raw throughput with one chip.

What do the new instances cost?

The M9g family lands between 31.9 percent and 33.6 percent better price-performance than the equivalent R8g instances with matching vCPU counts, on on-demand pricing. The M9gd variants with local flash run 22.1 percent to 32 percent ahead of their R8gd counterparts. AWS has not yet published savings plan pricing for the M9g, M9gd or X8g families.

Memory capacity tells a harder story. The M9g instances carry a quarter to half the memory of their Graviton4-based predecessors, and AWS charges about half the X8g price for them. That gap reflects the DRAM and flash crunch: compute is cheap right now, memory is not, and AWS is less generous with both DIMM capacity and local flash. The compensation is speed — the instances use the fastest DRAM available in the DDR5 form factor, running at 8.8 GHz, so bandwidth-sensitive workloads may not miss the capacity.

What comes next for memory?

CXL 3.0 support on the PCI-Express 6.0 controllers points at the exit from the memory cost trap. The protocol allows memory capacity to extend beyond what the socket can affordably hold, since DIMM prices scale exponentially rather than linearly with capacity per stick. The editor expects heavier-memory X9g instances before long, though at a premium given current DRAM prices, and would not be surprised if AWS built a rack-scale shared memory appliance rather than in-node CXL extenders. For now, Graviton5 makes bandwidth the currency and lets customers pay for capacity only when a workload truly needs it.