V3 or V4? On Lenovo the Generation Number Does Not Tell You How New the Silicon Is

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

A quote lands with "ThinkSystem SR650 V3" on it and the immediate reaction is that someone is trying to sell last year's hardware. Sometimes that is exactly what is happening. Often it is not, and the reason is that the V-number on a Lenovo server describes the platform generation, not the vintage of the processor inside it — and the two lines have drifted apart.

On the Intel side, V4 carries the newest silicon. On the AMD side, as of August 2026, the newest silicon still ships inside machines labelled V3. Someone comparing two quotes on the strength of the version number alone will reach the wrong conclusion roughly half the time.

What the label actually means on each side

Read that table once more, because the consequence is counterintuitive: if you want the highest core count Lenovo currently sells in a two-socket rack server, you buy a V3, not a V4. The EPYC 9005 family spans 8 cores at the bottom to 192 at the top, across Zen 5 for per-core performance and Zen 5c for density, and it went into the existing V3 chassis rather than triggering a new generation label. Platform ceilings sit below the family maximum — the SR665 V3 supports up to 160 cores per socket — so check the platform, not the family.

So "V3 or V4" is not really the question. The questions are which processor family suits the workload, what the licence model does to that choice, and what is actually available in the timeframe you have.

SR650 V3 against SR650 V4, side by side

Same chassis role, one platform generation apart. This is the comparison that decides most standard virtualization refreshes:

What actually improved, ranked by how much it will matter to you

1. Storage topology, not core count. Thirty-six onboard NVMe ports and thirty-two NVMe drives without oversubscribing PCIe lanes is the change with the largest practical effect on dense virtualization and database hosts. On the previous platform, high NVMe counts meant retimers, expanders and a conversation about which drives share bandwidth. That conversation goes away.
2. Cores, +34%. Real, and it does change consolidation ratios — but see the licensing section below before treating it as a straight win.
3. Memory speed and MRDIMM. Faster at both 1 DPC and 2 DPC, with MRDIMM reaching 8000 MHz. Note what did not change: maximum capacity is 8 TB on both. If your constraint is memory capacity rather than bandwidth, V4 does not solve it.
4. Two OCP slots and a 400 GbE ceiling. Matters if the node is doing storage networking or east-west traffic at scale; irrelevant on a general-purpose host with a couple of 25 GbE links.
5. CXL 2.0 memory expansion. Real capability, thin production track record. Interesting rather than decisive for most estates today.

The trade nobody puts in the quote

The Neptune Core liquid-cooled compute complex on the V4 halves the DIMM slots, from 32 to 16. On a memory-heavy virtualization host that is not a footnote — it is the difference between hitting your RAM target with sensible DIMMs and being forced onto the largest, most expensive modules to compensate. Liquid cooling on this platform buys density and power efficiency at the cost of memory capacity — but not, as it happens, memory bandwidth: sixteen slots on a two-socket machine is eight per socket, which is the balanced full-bandwidth population. The memory population rules explain why. Decide which constraint binds before you tick the box.

Per-core licensing inverts the "more cores is better" instinct

This is the part that changes the answer for most buyers, and it has nothing to do with the hardware datasheet.

Every major stack now meters on physical cores. VMware bills per physical CPU core with a 16-core minimum per processor, so billable cores are the sum of the greater of actual cores or sixteen across every socket. Nutanix licenses per core. Azure Local charges a host fee per core per month. The licence follows the silicon, every core, every year, for the life of the platform.

Two consequences pull in opposite directions, and the right node sits between them:

• The floor punishes small processors. A pair of 12-core CPUs bills as 32 cores. Buying low-core parts to save on licensing simply does not work — you pay for sixteen per socket whether they exist or not.
• The ceiling punishes maximum processors. A 160-core EPYC delivers extraordinary consolidation, and then charges you a licence on all 160 cores annually. Over a three-year term that recurring cost routinely exceeds the difference in hardware price between that part and one two tiers down.

So there is an optimal band rather than an optimal maximum, and it moves with your licence cost per core. The arithmetic worth doing before the CPU model is fixed: take the licence cost per core per year, multiply by the term, add it to the server price, and divide by the workloads that node will actually carry. Run it for three candidate SKUs. On most estates the winner is not the top part, and it is not the cheapest one either — and the gap between the naive choice and the right one is frequently larger than any discount you will negotiate on the hardware.

We work through the licence side of this in more depth, including what it means for storage-heavy clusters, in the hardware decision behind a VMware exit.

What to buy now

Availability decides more of this than the spec sheet does

Our published position as of August 2026: ThinkSystem SR630 and SR650 V3 are in stock in Hong Kong and Dubai, typically shipping in days to a week; V4 is to order at two to four weeks; ThinkAgile and storage lines run three to six weeks. On a refresh with a fixed maintenance window, that gap frequently settles the generation question on its own — and the honest version of the advice is that a V3 node installed on schedule beats a V4 node that arrives after the window closed.

The difference between stock and BTO/CTO is worth understanding before you specify, because a configuration that is one component away from a stock build can often be met from stock, while an insistence on a specific SKU pushes the whole order into a factory cycle. At project quantities, special-bid pricing changes the mechanism again — it is registered against the opportunity rather than applied as a discount, so the structure has to be settled before the quote.

Four ways this goes wrong

1. Rejecting a V3 quote on the version number. On AMD, V3 is current silicon. On Intel it is the previous generation and frequently the better commercial answer. Ask which processor family, not which V-number.
2. Buying the top SKU because it is the top SKU. Under per-core licensing, the highest core count is a recurring cost decision disguised as a hardware decision.
3. Specifying liquid cooling and a large memory footprint together. On the V4 Neptune Core complex those requirements fight each other — 16 DIMM slots, not 32.
4. Assuming V4 solves a memory-capacity problem. It does not. Both generations stop at 8 TB. If you need more, the answer is a different chassis class, not a newer one.

Frequently asked questions

Is a ThinkSystem V3 server outdated in 2026?

It depends which processor is in it. The AMD V3 platforms — SR635 V3, SR645 V3, SR655 V3, SR665 V3 and others — carry 5th Gen EPYC "Turin" 9005 processors — up to 160 cores per socket on the SR665 V3, the newest silicon Lenovo ships in those chassis. The Intel V3 platforms carry 4th and 5th Gen Xeon Scalable and are one generation behind V4. Same label, different answer.

What is the actual difference between SR650 V3 and SR650 V4?

Cores rise from 64 to 86 per socket, memory runs at 6400 MHz with MRDIMM support to 8000 MHz, CXL 2.0 memory expansion appears, and storage gains E3.S support with 36 onboard NVMe ports and 32 NVMe drives without PCIe lane oversubscription. Networking gains a second OCP 3.0 slot and a 400 GbE ceiling. Maximum memory capacity is unchanged at 8 TB on both.

Which Lenovo server has the highest core count?

An AMD V3 platform, not a V4. The SR665 V3 supports up to 160 cores and 320 threads per socket on 5th Gen EPYC, against 86 cores and 172 threads on Xeon 6 in the V4 line. The EPYC 9005 family reaches 192 cores at the very top, but platform support is what matters — confirm the SKU against the specific server. Whether high core count is the right purchase depends entirely on your licensing model.

Does more cores per socket save money?

Only up to a point. Every major stack licenses per physical core, with VMware applying a 16-core minimum per processor. Low-core CPUs are billed as if they had sixteen; very high-core CPUs carry a licence on every core for the life of the platform. The right choice is a band in the middle, and the arithmetic should be run over the full term before the CPU model is fixed.

Does liquid cooling on the V4 have a downside?

Yes, and it is rarely mentioned. The Compute Complex Neptune Core configuration reduces the SR650 V4 from 32 DIMM slots to 16. On a memory-heavy host that forces you onto larger, more expensive modules. Decide whether density or memory configuration is your binding constraint before specifying it.

Should we wait for V4 or buy V3 from stock?

If the workload is standardised virtualization or databases and the maintenance window is fixed, V3 from Hong Kong stock ships in days against two to four weeks for V4 and generally prices better. If the node is storage-dense, or you want the longest support runway on a new build, V4 earns the wait. A node that arrives after the window closed has no specification advantage worth having.

Is V4 available with AMD processors?

As of August 2026 the AMD platforms carry the V3 designation, including those running 5th Gen EPYC "Turin". The V4 label currently applies to the Intel Xeon 6 line. Ask for the processor family on the quote rather than inferring currency from the version number.

Related

• How to choose Lenovo ThinkSystem servers — workload-to-model selection and host sizing
• HPE ProLiant Gen11 vs Gen12 — the core-variant and memory-speed decision
• VMware exit — the hardware decision — how per-core licensing reshapes node design
• Which ThinkAgile series for which hypervisor — if the refresh is going hyperconverged
• Lenovo vs Dell vs HPE — the same generation compared against PowerEdge 17G and ProLiant Gen12
• Lenovo stock, lead times and pricing · stock vs BTO/CTO · project pricing
• Server refresh and consolidation — the rack, power and licence arithmetic behind the generation choice
• SR650 V3 vs SR665 V3 — Intel or AMD in the same 2U slot, and what per-core licensing does to that choice
• Memory population rules — why eight DIMMs per socket beats sixteen on Intel
• SR630 V4 vs SR650 V4 — the chassis question once the generation is settled
• Lifecycle and end of support — which platforms are withdrawn and what replaced them
• ITAD and the secondary market — recovering value from the fleet being replaced
• Sovereign cloud infrastructure — 40× AMD EPYC 9654, 7,680 vCPUs, 96 TB NVMe

Sources

• Lenovo Press — ThinkSystem SR650 V4 product guide (86 cores, MRDIMM 8000 MHz, CXL 2.0, 32× E3.S, 36 onboard NVMe ports, two OCP slots, Neptune Core and its 16-DIMM configuration)
• Lenovo Press — ThinkSystem SR650 V3 product guide (4th and 5th Gen Xeon Scalable, 64 cores, 5600 MHz, 8 TB, PCIe 5.0 slot layout)
• Lenovo Press — Introducing the ThinkSystem V4 servers with Intel Xeon 6
• Lenovo Press — ThinkSystem servers with 5th Gen AMD EPYC "Turin" (which V3 platforms take EPYC 9005; 8 to 192 cores; Zen 5 and Zen 5c)
• Lenovo Press — ThinkSystem SR630 V4 product guide
• VMware licensing changes — per-core subscription and the 16-core minimum per CPU