160 Cores per Socket, or 8 TB of Memory. Not Both.

Written and maintained by Haink's infrastructure team · Verified against Lenovo Press SR650 V3 and SR665 V3 product guides, 29 August 2026 · authorized-channel, MTM and serial verified

Two 2U dual-socket servers, same generation label, same chassis role, same price bracket. The SR650 V3 is Intel, the SR665 V3 is AMD, and they are not two versions of the same machine — they bind on opposite resources.

The AMD platform gives you two and a half times the cores. The Intel platform gives you a third more memory and eight more DIMM slots. Which of those constraints your workload hits first is the entire decision, and it is answerable from your own utilisation data in about ten minutes.

Side by side

Memory per core is where the two diverge

The headline numbers hide the constraint. Divide memory by cores and the platforms look completely different:

• SR650 V3: 8 TB across 128 cores — roughly 64 GB per core available.
• SR665 V3: 6 TB across 320 cores — roughly 19 GB per core available.

For general virtualization, where memory is more often the binding constraint than CPU, this is the whole story. A fully populated AMD node runs out of memory well before it runs out of cores at any realistic consolidation ratio: at six virtual CPUs per physical core and 12 GB per virtual machine, 320 cores wants something close to the full 6 TB. Push the ratio or the VM size any further and the cores sit idle because there is nothing to put on them.

The Intel node has the opposite problem in mirror image: memory headroom it cannot use, because 128 cores run out first. Neither is a defect. They are different machines for different shapes of workload, and the shape is in your utilisation data — committed memory divided by committed vCPU, which is the ratio to bring to the conversation.

The licence changes the answer more than the silicon does

Every major virtualization stack meters on physical cores. VMware bills per core with a 16-core minimum per processor, Nutanix per core, Azure Local as a per-core monthly host fee. A fully populated SR665 V3 carries 320 billable cores; a fully populated SR650 V3 carries 128.

That is two and a half times the annual licence for the same rack unit, and it recurs for the life of the platform. On a core-licensed estate the AMD consolidation advantage largely transfers to the licence vendor rather than to you — which is why the honest answer for most virtualization refreshes is a mid-range processor in either platform rather than the top SKU in the AMD one. We work that arithmetic through in choosing a generation under per-core licensing.

Where the workload is not licensed per core — containers on open-source orchestration, HPC, scale-out application tiers, anything licensed by socket or by user — the calculation inverts completely and the AMD platform is the obvious buy.

Three smaller differences that decide real designs

1. Memory speed favours AMD. 6400 MHz on 5th Gen EPYC against 5600 MHz at one DIMM per channel on the Intel platform. For bandwidth-bound workloads — analytics, in-memory databases, simulation — that is a real advantage that the core-count argument tends to drown out.
2. DIMM population works differently, not just tighter, on AMD. Twelve slots per processor is one per memory channel, so a balanced AMD node is already fully populated and capacity grows only by replacing modules. On Intel, eight per socket is the full-bandwidth population and sixteen is available later at a bandwidth cost. The population rules and bandwidth tables cover both.
3. The NVMe lane story is documented on AMD. The SR665 V3 guide states up to 32 NVMe drives without oversubscription. The Intel guide does not give a comparable figure, which is not the same as saying it is worse — it means you should ask for the lane map rather than assume parity.

Choosing

One procurement note that applies to both: base warranty is encoded in the machine type on this generation as well. The SR665 V3 uses 7D9A for three-year and 7D9B for one-year base warranty — the same server, two machine types. Compare machine types rather than product names, as set out in MTM and part numbers.

Frequently asked questions

Which has more cores, the SR650 V3 or the SR665 V3?

The AMD SR665 V3, by a wide margin. It supports up to 160 cores per socket on 5th Gen EPYC "Turin" — 320 cores and 640 threads per node — against up to 64 cores per socket and 128 per node on the Intel SR650 V3.

Then why would anyone buy the Intel platform?

Memory and licensing. The SR650 V3 has 32 DIMM slots against 24 and reaches 8 TB against 6 TB, which matters because memory binds before CPU on most virtualization. And because every major stack licenses per physical core, the AMD node's 320 cores carry roughly two and a half times the annual licence of the Intel node's 128.

Does the SR665 V3 support 192-core EPYC processors?

The EPYC 9005 family reaches 192 cores at the top of the range, but the SR665 V3 platform supports up to 160 cores per socket. Always confirm the specific processor SKU on the quote rather than working from the family maximum.

Which platform has faster memory?

The AMD one. The SR665 V3 runs DDR5 at up to 6400 MHz with 5th Gen EPYC, against 5600 MHz at one DIMM per channel and 4800 at two on the Intel platform. For bandwidth-bound analytics and in-memory workloads that is a meaningful advantage.

How do we decide between them from our own data?

Divide committed memory by committed vCPU across the estate. A memory-heavy ratio points to the Intel platform, which offers roughly 64 GB per core against about 19 GB per core on a fully populated AMD node. A core-hungry, memory-light ratio points the other way — provided the workload is not licensed per core.

Is a V3 server outdated?

Not on the AMD side. The SR665 V3 carries 5th Gen EPYC "Turin", which is the current silicon for that platform — the AMD line took the new processors without a version bump. On the Intel side, V3 is the previous generation and V4 is current.

Do these two have different machine types?

Yes, and each has more than one. The SR665 V3 uses 7D9A for three-year base warranty and 7D9B for one-year. Two quotes for the same server with different machine types are not the same purchase.

Related

• ThinkSystem V3 vs V4 — why the AMD platforms carry a V3 label while running current silicon
• SR630 V4 vs SR650 V4 — the 1U against 2U question on the Intel line
• Model numbers decoded — why the last digit is what separates these two
• MTM and part numbers — 7D9A against 7D9B, and what else the machine type encodes
• How to choose Lenovo ThinkSystem servers — workload-to-model selection and host sizing
• Memory population rules — why twelve DIMMs per socket is optimal on AMD and a mistake on Intel
• Sizing an HCI cluster — where the memory-to-core ratio actually gets used
• Lenovo vs Dell vs HPE — the same chassis role across three vendors
• Lenovo stock, lead times and pricing · stock vs BTO/CTO

Sources

• Lenovo Press — ThinkSystem SR665 V3 product guide (5th Gen EPYC 9005 up to 160 cores and 320 threads, 400 W TDP, 24 DIMM slots, 6 TB, 6400 MHz, 32 NVMe without oversubscription, 12 PCIe slots plus OCP, machine types 7D9A and 7D9B)
• Lenovo Press — ThinkSystem SR650 V3 product guide (4th and 5th Gen Xeon Scalable up to 64 cores, 385 W TDP, 32 DIMM slots, 8 TB, 5600 MHz at 1 DPC, 12 PCIe slots plus OCP)
• Lenovo Press — ThinkSystem servers with 5th Gen AMD EPYC (which platforms took Turin; the 9005 family range up to 192 cores)
• VMware licensing changes (per-core subscription, 16-core minimum per CPU)