China's 96-core x86 CPU taps chiplet design to rival AMD EPYC and Intel Xeon
4 hour ago / Read about 17 minute
Source:Tomshardware
Zhaoxin announces the company's latest KH-50000 series of server processors, leveraging a chiplet design with up to 96 cores.

(Image credit: Zhaoxin)

Zhaoxin may not produce the best CPUs for gaming, but the leading Chinese fabless semiconductor enterprise is undoubtedly preparing to unleash a highly impressive server chip. Zhaoxin has unveiled its next-generation Kaisheng KH-50000 processors, which the company describes in its press release as "presenting a 'technological gift' on the eve of the 76th anniversary of the founding of the People's Republic of China."

The KH-50000 utilizes Zhaoxin's latest Century Avenue architecture, named after a famous road in Shanghai. The company is fond of naming its architectures after famous locations within Shanghai because, after all, Zhaoxin is a joint venture between VIA Technologies and the Shanghai government. Century Avenue is the current architecture used by the company for its mainstream KaiXian KX-7000 processors; consequently, it is logical for Zhaoxin to align its latest server processors accordingly. Although Century Avenue is an in-house developed architecture by Zhaoxin, speculation suggests that Century Avenue may be a derivation from Centaur Technology's CNS core, prior to the company's split from VIA Technologies in 2021.

Zhaoxin utilizes a chiplet design for the KH-50000, similar to AMD's Ryzen and EPYC processors, with a greater emphasis on the latter, given the high number of cores. A chiplet design would enable Zhaoxin to push the core boundary on the KH-50000, effectively matching AMD's EPYC 9004 (codenamed Genoa) series that tops out at 96 cores. Zhaoxin has planned two variants of the KH-50000: the flagship 96-core SKU and a more affordable 72-core SKU, both of which lack simultaneous multithreading (SMT). The KH-50000 represents a monumental leap forward for Zhaoxin, as it provides 3X more cores than the company's existing KH-40000.

Zhaoxin's photograph of the KH-50000 reveals that the chipset layout exhibits minor differences from that of AMD; however, the core design remains consistent. The gargantuan I/O die is centrally positioned on the processor, encircled by four clusters of compute dies. Each cluster contains three compute dies, totaling twelve. Each die incorporates eight cores and 32MB of L3 cache. When assembled, the resulting processor comprises a 96-core configuration with 384MB of L3 cache.

Zhaoxin Kaisheng KH-50000 Specifications

Swipe to scroll horizontally

Processor

Architecture

Cores / Threads

Base / Boost Clock (GHz)

L3 Cache (MB)

Memory Support

PCIe Lanes

SATA 3.2 Ports

USB Ports

Socket

Package Size (mm)

KH-50000

Century Avenue

96 / 96

2.2 / 3.0

384

12 Channel DDR5-5200

128 PCIe 5.0 + 16 PCIe 4.0

12

4 x USB 3.2 Gen 2

LGA

72 x 76

KH-50000

Century Avenue

72 / 72

2.6 / 3.0

384

12 Channel DDR5-5200

128 PCIe 5.0 + 16 PCIe 4.0

12

4 x USB 3.2 Gen 2

LGA

72 x 76

KH-40000/32

Yongfeng

32 / 32

2.5 / N/A

64

8 Channel DDR4-3200

128 PCIe 3.0

16

8 x USB 3.2 Gen 1

LGA

77.5 x 56.5

The clock speeds on the KH-50000 aren't too shabby and fall in line with what you'd expect from a server chip. The 96-core variant has a 2.2 GHz base clock and 3.0 GHz boost clock. Since the 72-core chip has fewer cores, Zhaoxin could push the base clock to 2.6 GHz but maintained the same boost clock. Although the company has taken the wraps off the KH-50000, it didn't reveal the TDP or other power metrics for the upcoming server chip. The thing with a chiplet design is that Zhaoxin can effectively utilize older process nodes for the KH-50000. Sanctions don't hurt as much if you don't care about power consumption.

For comparison, AMD has historically kept its top EPYC chips around the 300-350W range, and that's with SMT. Nonetheless, the chipmaker has recently pushed the power envelope up to 500W, which is understandable when its EPYC processors are maxing out at 192 cores.

In addition to core count, the KH-50000 advances the development of Chinese server processors. It now supports up to 12 channels of DDR5-5200 RAM, allowing for a maximum of 3TB of memory, in contrast to the 2TB supported by DDR4-3200 on the KH-40000. Zhaoxin has added Compute Express Link (CXL) interconnect support. Furthermore, the expansion capabilities of the KH-50000 have enjoyed an upgrade to include 128 PCIe 5.0 lanes and 16 PCIe 4.0 lanes, compared to the 128 PCIe 3.0 lanes available on the KH-40000.

The SATA and USB ports experienced a slight decrease in numbers when comparing the KH-50000 to the KH-40000. However, Zhaoxin has upgraded the latter to support the latest USB 3.2 Gen 2 specification.

(Image credit: Zhaoxin)

The KH-50000 supports x86 32-bit and 64-bit instructions, including SSE4.2, AVX, and AVX2. Support for virtualization is also present. To adhere to China's security guidelines, the KH-5000 supports the country's proprietary SM2, SM3, and SM4 encryption standards. Notably, Zhaoxin has integrated National Technology's fourth-generation trusted computing chip (likely the NS350) beneath the KH-50000, where the contacts are situated. This chip meets the security requirements of China's GM/T 0012-2020 cryptographic module standard and complies with the international TPM 2.0 (SPEC 1.59) standard.

The footprint of the KH-50000 measures 72 x 76 mm, which is considerably larger than that of the KH-40000. Notably, it shares dimensions with AMD's Genoa and Bergamo processors, which measure 72 x 75.4 mm and are compatible with the socket SP5. Therefore, the size of the KH-50000 is precisely the same as that of AMD's more recent EPYC chips.

The KH-50000 slots into a socket with a Land Grid Array (LGA) design, meaning the pins are located on the motherboard rather than on the processor. Zhaoxin's latest server chips are scalable, similar to AMD's EPYC and Intel's Xeon chips. The KH-50000 embraces 2S and 4S systems, where you can accumulate up to 384 cores on the latter. Zhaoxin built its own ZPI (Zhaoxin Processor Interconnect) 5.0 for inter-chip communication.

Contrary to AI GPUs, companies in China can still acquire server chips without significant difficulty, albeit potentially at increased costs. Nonetheless, Zhaoxin continues to make considerable progress in the domestic market, and with Chinese authorities firmly committed to utilizing domestically produced technology, the company could achieve success even if the KH-50000 does not rival AMD or Intel's latest server chips.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!