Xeon 600: New workstation processors with current Intel technology

Intel is under pressure from AMD's Epyc and Threadripper processors but does not want to completely cede the workstation market to the competition. The manufacturer is therefore carving out some chips from its actually fully booked server production and repurposing them for faster workstations than was possible with the predecessor models based on the Sapphire Rapids architecture.

The new Xeon 600 series is based on current server and chiplet technology, but is intended for systems with only one socket and thus primarily for workstations. The processors use the LGA4710-2 socket and, together with the W890 chipset, offer up to 4 TByte of DDR5-6400 memory, 128 PCIe 5.0 lanes, Wi-Fi 7, and CXL 2.0 support. As is common with workstations, the Xeon 600 also uses expensive Registered DIMM type DDR5 memory modules with ECC.

Compared to the server CPUs, which according to Intel do not start in workstation boards and vice versa, they primarily lack multi-socket capability and the specialized accelerator circuits Data Streaming Accelerator (DSA), In-Memory Analytics Accelerator (IAA), QuickAssist (QAT), and Dynamic Load Balancer (DLB).

However, the Advanced Matrix Extensions (AMX) are present, which are particularly suitable for AI applications with low precision requirements. They have also received the architectural generation Redwood Cove improvements from the server and now support not only bfloat16 but also the "real" 16-bit floating-point data format FP16. Intel uses this in its Open Image Denoiser from version 2.4 and achieves a gigantic performance advantage of almost a factor of 4 compared to the older version 2.3.1 – however, how much of this is solely due to the AMX-FP16 option is questionable. Our test server reported only the three known AMX flags amx_bf16, amx_tile, and amx_int8 via lscpu | grep amx. A clarifying answer from Intel to our inquiry is still pending.

At least the new Xeon 600 are said to be not only 9 percent faster in single-threading but also have 61 percent more multi-threading performance compared to the top models 698X versus w9-3595X. Intel claims to have measured this in the Cinebench 2026 rendering benchmark. For Blender, it's said to be 74 percent more performance, and for AI-assisted upscaling in Topaz Labs Video, at least 29 percent.

Thanks to the newer Intel 3 manufacturing technology for the compute die(s), the Thermal Design Power (TDP), i.e., the specified electrical power consumption, remains within the same range. For the top model, it even slightly decreases again from 385 watts for the Xeon w9-3595X to 350 watts for the Xeon 698X – the w9-3495X already achieved this.

Eleven newcomers

The range of newcomers extends from the Xeon 634 for a wholesale price of $499, excluding taxes, to the 86-core Xeon 698X for $7699. The former has only four memory channels, 80 PCIe 5.0 lanes, and cannot be overclocked. The top model, on the other hand, is unlocked for overclockers, who will likely use it to chase records again in typical benchmarks. In addition, the Xeon 698X, along with four other X models, can address the particularly fast Multiplexed-Rank DIMM (also known as MCR-DIMMs) at DDR5-8800 speed. Only the smallest Xeon 600X, the 24-core 658X, lacks MRDIMM support. Speaking of which: MRDIMMs are currently mainly available for end customers as part of complete systems; price comparison currently lists no kits at all.

In addition to clock speed increases, selective clock speed reductions are also possible for the particularly energy-intensive AVX-512 and matrix multiplication instruction chains (TMUL). These can increase the general clock speed potential and only reduce the frequency when corresponding instructions are used, so that heat and energy consumption do not exceed limits.

The eleven new CPUs are expected to go on sale at the end of the first quarter and will primarily be used in complete systems from system integrators and large vendors such as Lenovo, Dell, and HP. Intel names Asus, Supermicro, and Gigabyte as partners for the corresponding W890 boards. Board prices are unlikely to be lower than those of the predecessors and will typically occupy the upper three-digit range.

Intel Xeon 600: New workstation processors with P-cores "Redwood Cove"
Xeon	Cores	Base Clock (Turbo all-core/single-core)	Level 3 Cache	TDP	Memory	MR-DIMM	PCIe 5.0 Lanes	Boxed	Price (US Dollars)
698X	86	2.0 (3.0/4.8) GHz	336 MByte	350 W	8 × DDR5-6400	8 × DDR5-8800	128	-	7699 $
696X	64	2.4 (3.5/4.8) GHz	336 MByte	350 W	8 × DDR5-6400	8 × DDR5-8800	128	Yes	5599 $
678X	48	2.4 (3.8/4.9) GHz	192 MByte	300 W	8 × DDR5-6400	8 × DDR5-8800	128	Yes	3749 $
676X	32	2.8 (4.3/4.9) GHz	144 MByte	275 W	8 × DDR5-6400	8 × DDR5-8800	128	Yes	2499 $
674X	28	3.0 (4.3/4.9) GHz	144 MByte	270 W	8 × DDR5-6400	8 × DDR5-8800	128	-	2199 $
658X	24	3.0 (4.3/4.9) GHz	144 MByte	250 W	8 × DDR5-6400	-	128	Yes	1699 $
656	20	2.9 (4.5/4.8) GHz	72 MByte	210 W	8 × DDR5-6400	-	128	-	1399 $
654	18	3.1 (4.5/4.8) GHz	72 MByte	200 W	8 × DDR5-6400	-	128	Yes	1199 $
638	16	3.2 (4.5/4.8) GHz	72 MByte	180 W	4 × DDR5-6400	-	80	-	899 $
636	12	3.5 (4.5/4.7) GHz	48 MByte	170 W	4 × DDR5-6400	-	80	-	639 $
634	12	2.7 (3.9/4.6) GHz	48 MByte	150 W	4 × DDR5-6400	-	80	-	499 $
-: not supported; all cores with SMT, all X models unlocked for overclockers, all Xeon 600 support vPro, max. memory expansion 4 TByte

(csp)