Dell Precision 7920 vs 7960: Two Generations of Big Iron

The Precision 7920 Tower and the 7960 Tower occupy the same position in Dell's lineup ? the largest tower they build ? but they solve the core-count problem in opposite ways. The 7920 uses two Xeon Scalable sockets. The 7960 uses one Xeon W-3400. Both top out at 56 cores. How they get there changes memory topology, PCIe generation, NUMA behavior, and per-socket licensing, which is a longer list of consequences than the matching core count suggests.

The 7920: two sockets, twenty-four DIMM slots

The 7920 Tower is built on dual LGA 3647 and takes first-generation (Skylake-SP) and second-generation (Cascade Lake) Xeon Scalable processors, up to 28 cores per socket. Each socket contributes six DDR4 channels and 48 PCIe 3.0 lanes, so a fully populated machine offers 12 channels across 24 DIMM slots and 96 CPU-attached Gen3 lanes. Top-bin second-generation parts run memory at 2933 MT/s with one module per channel; fill the second slot on each channel ? which is what using all 24 of them means ? and Cascade Lake steps back to 2666. Silver and Bronze SKUs run slower than that no matter what you install.

The two sockets talk to each other over two or three UPI links depending on the SKU tier, at 10.4 GT/s on Platinum and the upper Gold parts and 9.6 GT/s further down the range. Those links are the thing to keep in mind: the operating system sees two NUMA nodes, and a thread reaching memory attached to the other socket pays for it in latency. Dell also sold this platform as the Precision 7920 Rack, a 2U variant, which is useful if you want workstation-class graphics inside a cabinet.

The 7960: one socket, eight channels, PCIe 5.0

The 7960 Tower moves to a single LGA 4677 socket and the Xeon W-3400 family, which reaches 56 cores in one package at 350 W TDP. Memory becomes DDR5 registered ECC at 4800 MT/s across eight channels and 16 DIMM slots ? sixteen 64 GB modules gets you to 1 TB. The same two-per-channel rule applies here: filling the second module on each channel drops the rated speed from 4800 to 4400 MT/s, so plan capacity against the speed you actually need rather than filling every slot by reflex.

Expansion is the bigger jump: the W-3400 provides 112 PCIe 5.0 lanes. A Gen5 x16 link carries roughly four times the bandwidth of a Gen3 x16 link, which matters for NVMe arrays, 100 GbE adapters, and modern accelerators far more than it does for a graphics card driving displays. The W-2400 series, with four memory channels and 64 lanes, lives in Dell's 5860 Tower rather than this one.

Side by side

Specification	Precision 7920 Tower	Precision 7960 Tower
Sockets	2 ? LGA 3647	1 ? LGA 4677
CPU family	Xeon Scalable gen 1?2 (Skylake-SP / Cascade Lake)	Xeon W-3400 (Sapphire Rapids)
Maximum cores / threads	56 / 112 (2 ? 28c)	56 / 112 (1 ? 56c)
Memory	DDR4 ECC RDIMM/LRDIMM, 24 slots	DDR5 ECC RDIMM, 16 slots
Channels	6 per socket, 12 total	8 total
Rated speed	2933 MT/s at one DIMM per channel, 2666 MT/s with both slots filled	4800 MT/s at one DIMM per channel, 4400 MT/s with both slots filled
Theoretical peak bandwidth	~281 GB/s with 12 modules, ~256 GB/s with all 24 ? split across two nodes	~307 GB/s with 8 modules, ~282 GB/s with all 16 ? single node
PCIe	Gen3, 48 lanes per socket (96 total)	Gen5, 112 lanes
NUMA nodes	2	1
Top CPU TDP	205 W per socket	350 W
Notable ISA	AVX-512 (VNNI on gen 2)	AVX-512 plus AMX

Which one is actually faster

Look at the bandwidth row again. At one DIMM per channel the two machines are close on paper ? roughly 281 GB/s against roughly 307 GB/s ? and with every slot filled they land at roughly 256 GB/s and roughly 282 GB/s. But the 7920's figure is split between two NUMA nodes joined by UPI, while the 7960's sits in one flat address space. For a single memory-bound process that cannot be partitioned cleanly, the single-socket machine usually wins by more than the raw numbers imply, because it never pays a remote-node penalty. For workloads you can pin per socket ? a batch renderer, several VMs, parallel solver instances ? the dual-socket topology stops being a liability.

Per-core throughput is not close. Sapphire Rapids delivers substantially more work per clock than Skylake-SP or Cascade Lake, and it adds AMX for low-precision matrix math. If your software is licensed per core, the newer platform returns more compute per license. If it is licensed per socket, the 7920 counts twice.

The 7920 answers back on slot count. Twenty-four DIMM slots means very large capacity is reachable with smaller, widely available DDR4 modules instead of the largest DDR5 parts ? at the 2666 MT/s grade that a fully populated board implies. For memory capacity as the primary requirement ? large in-memory datasets, many concurrent VMs, big scenes ? that is the practical advantage, and DDR4 registered memory remains easy to source.

Buying either one used

On a 7920, verify the CPU pair matches exactly and that the second socket ships with its heatsink and air shroud; a second CPU that arrives without them is an incomplete parts list, not a working socket. Check DIMM population order ? one module per channel across both sockets before doubling up ? and confirm the installed power supply against your GPU plans and the circuit you will run it on. Riser cards and drive carriers are the parts most often missing.

On a 7960, the platform is DDR5 registered ECC only; unbuffered desktop DDR5 will not post. Confirm the cooling assembly is the one specified for a 350 W part rather than a lower-TDP variant, and check that the BIOS revision matches the processor you intend to run.

Where to go from here

Both families sit in our refurbished workstations collection, usually in several core-count and memory configurations. If you already know the shape of the machine you want ? socket count, DIMM population, storage layout, accelerator ? specify it directly on the Build Your Server page and request a quote. If you are still deciding between one big socket and two, tell us what the software does and how it is licensed; that pair of facts settles it faster than a spec sheet will.