AMD 4th-Gen EPYC Genoa 9654, 9554, and 9374F Review: 96 Cores, Zen 4 and 5nm Disrupt the Data Center

AMD’s 4th Gen EPYC Genoa processors are the industry’s first 5nm x86 CPUs for the data center, spearheaded by the flagship 96-core 192-thread EPYC 9654. The $11,805 EPYC 9654 packs an unprecedented amount of compute into slim server designs (up to 192 cores and 384 threads). Additionally, AMD delivers up to 30% more performance per core in both integer and floating point operations than Intel’s Ice due to various advances such as a 14% increase in IPC from the Zen 4 architecture and improved power delivery. said to improve. lake. This is even more impressive due to the core count advantage. Top-of-the-line Genoa processors include Twice 60% more core count than the Ice Lake Xeons and the rumored peak of 60 cores for the yet-to-be-released Sapphire Rapids.
The 9004-series Genoa chips also feature the latest connectivity technology, including support for up to 384MB of L3 cache and up to 6TB of memory across 12 channels of DDR5, 128 lanes of PCIe 5.0, and CXL 1.1+ . Among them, Intel’s Ice Lake product stack is 40-core Intel Xeon Platinum 8380 At $9,400, it looks pretty dated. Sapphire Rapids is loaded with advanced connectivity technology, with a host of built-in accelerators, making it a true competitor to Genoa. However, it won’t arrive until January 2023.
EPYC Genoa also brings many other new additions such as support for AVX-512 and support for VNNI and Bfloat16 instructions for AI acceleration.
However, huge core counts and performance come at a price. Genoa’s flagship models come with his 360W peak default TDP, the highest of any x86 server processor to date, and customers can tune them up to his 400W for maximum performance. I can do it.
As we have seen with GPUs, power consumption is increasing rapidly. This is due to the insatiable demand for more computing packed into the smallest form factor. Genoa is no exception. AMD’s customers are demanding higher TDP limits to improve compute density and total cost of ownership (TCO), and improvements in both processors and cooling technology have allowed the company to use standard air cooling for maximum It is now capable of delivering 400W of performance. However, this comes with secondary power requirements. For example, a single fan in our test system can consume up to 300W. This is before calculating his 300W consumed by 1.5TB of DDR5 memory.
Overall, the platform has a voracious appetite for power, but EPYC Genoa translates that power into an incredible amount of performance and a reduced TCO, unmatched by its x86 competitors. We tested AMD’s Genoa in the lab today with a 96 core EPYC 9654, a 64 core 9554 and a frequency optimized 32 core 9274F. Let’s dive in.
AMD 4th Gen EPYC Genoa 9004 Series Specs and Pricing
As you can see on the left, the Genoa processor is much larger than the previous generation Milan chip next to it, not to mention the consumer AMD and Intel processors we threw in for comparison.
Genoa’s larger chip package houses up to 12 5nm core compute dies (CCDs), each with 8 cores. That’s an increase of four additional CCDs compared to the previous generation Milan, requiring a larger chip package and an integrated heat spreader (IHS) to help improve cooling. The chip also contains his 6nm I/O die in the middle that ties all the chiplets together. More on this on the next page.
The Genoa processors will drop into the new SP5 socket which is not backwards compatible with the socket SP3 found on previous generation EPYC systems. That means the chip needs an entirely new platform. In the future, SP5 will also support Genoa-X processors. It incorporates a 3D stack L3 cache. Milano-X, and the Bergamo Chip with new high-density Zen 4c cores that enable up to 128 cores in a single socket.
model | price | core/thread | Base/Boost (GHz) | TDP | L3 Cache (MB) | cTDP (W) | package |
---|---|---|---|---|---|---|---|
EPYC Genova 9654 | $11,805 | 96/192 | 2.4/3.7 | 360W | 384 | 320-400 | 12+1 |
EPYC Genova 9634 | $10,304 | 84 / 168 | 2.25/3.7 | 290W | 384 | 240-300 | 12+1 |
EPYC Genova 9554 | $9,087 | 64/128 | 3.1/3.75 | 360W | 256 | 320-400 | 8+1 |
EPYC Milano 7763 | $7,890 | 64/128 | 2.45/3.5 | 280W | 256 | ||
EPYC Genova 9534 | $8,803 | 64/128 | 2.45/3.7 | 280W | 256 | 240-300 | 8+1 |
EPYC Milano 7663 | $6,366 | 56/112 | 2.0/3.5 | 240W | 256 | ||
EPYC Genova 9454 | $5,225 | 48/96 | 2.75/3.8 | 290W | 256 | 240-300 | 8+1 |
EPYC Milano 7643 | $4.995 | 48/96 | 2.3/3.6 | 225W | 256 | ||
Xeon Platinum 8380 | $8,099 | 40/80 | 2.3/3.2-3.0 | 270W | 60 | ||
Zeon Platinum 8368 | $6,302 | 38/76 | 2.4/3.4 – 3.2 | 270W | 57 | ||
EPYC Genova 9354 | $3,420 | 32/64 | 3.25/3.8 | 280W | 256 | 240-300 | 8+1 |
EPYC Genova 9334 | $2,990 | 32/64 | 2.7/3.9 | 210W | 128 | 200-240 | 4+1 |
EPYC Genova 9254 | $2,299 | 24/48 | 2.9/4.15 | 200W | 128 | 200-240 | 4+1 |
EPYC Genova 9224 | $1,825 | 24/48 | 2.5/3.7 | 200W | 64 | 200-240 | 4+1 |
EPYC Genova 9124 | $1,083 | 16/32 | 3/3.7 | 200W | 64 | 200-240 | 4+1 |
EPYC Genova 9474F | $6,780 | 48/96 | 3.6/4.1 | 360W | 256 | 320-400 | 8+1 |
EPYC Genova 9374F | $4,850 | 32/64 | 3.85/4.3 | 320W | 256 | 320-400 | 8+1 |
EPYC Milano 7F53 | $4,860 | 32/64 | 2.95 / 4.0 | 280W | 256 | ||
EPYC Genova 9274F | $3,060 | 24/48 | 4.05/4.3 | 320W | 256 | 320-400 | 8+1 |
EPYC Genova 9174F | $3,850 | 16/32 | 4.1/4.4 | 320W | 256 | 320-400 | 8+1 |
The entire EPYC Genoa 9004 series family spans 18 models in 3 categories: Core Performance, Core Density, Balanced and Optimized, far more than the Ice Lake Xeon which has a total of 56 models with various feature sets. Create a simple product stack in .
AMD has made a concerted effort to limit the product stack to critical swimlanes. The majority of the Genoa family are general purpose chips that slot into the “Core Density” with the highest number of cores and the “Balanced and Optimized” category for a combination of performance and TCO. On the other hand, F-series chips with higher frequencies and higher core-to-cache ratios fall into the “core performance” tier. AMD also has some P-series models such as the 9354P designed for single processor (1P) systems (listed in the slide below).
Genoa chips range from 16 to 96 cores, and I noticed that 8, 28, or 56 core variants are gone. At least for now. Peak clock speeds range from 3.7 GHz to 4.4 GHz, giving you the best boost from the frequency-optimized F-series models.
Genoa’s TDP ratings range from 200W to 360W, so the lowest TDP has increased by 45W and the highest TDP has increased by 80W. Configurable TDP (cTDP) is a customer/OEM tunable parameter that improves system performance with robust cooling. It now reaches 400W, a significant 120W increase for him over the previous generation chip.
As you can see from the previous generation 7003 series Milan chips we’ve added to the table, Genoa’s flagship will cost $3,915 more than the previous generation Halo part, the EPYC Milan 7763, with an extra 32 cores. AMD has also increased the prices of the 64-core models by $1,200 and $1,750 over the previous generation models. many Quieter prices move further down the stack. For example, the two 32-core models went up by $341 and $150, while the 48-core model went up by only $200. Also keep in mind that, like Intel, AMD’s server chip pricing is just a guideline, so the actual prices can vary significantly, especially for larger customers.
We’ve added a few Ice Lake Xeon models to the table above, but we kept the additions to a minimum due to the large number of Intel SKUs. On top of that, Genova will mostly face off against the upcoming Sapphire Rapids – as you can see in the benchmarks, we’ll have to wait for that launch for a fair comparison. It’s underwhelming with cores and 60MB of L3 cache, but the Genoa stack tops the list with 96 cores and 384MB of L3 cache. AMD has 6 SKUs with higher core counts, with 9 SKUs claiming to offer higher performance on integer workloads (last slide of the album).
All Genoa chips support:
- Simultaneous Multithreading (SMT)
- 12 channels of DDR5-4800 memory in 1DPC configuration (2DPC speeds to be announced in Q1 2023)
- 6 TB of memory per socket
- 128 lanes of PCIe 5.0 (64 lanes support CXL 1.1+)
- AVX-512, VNNI, Bfloat 16
Genoa processors mark the debut of several new technologies for x86 servers such as DDR5 and PCIe 5.0. The former currently costs more than his existing DDR4 memory, while the latter costs more on the motherboard. Other advancements such as his 5nm/6nm production node used within the chip and increased power and cooling requirements for the top-of-the-line models also add cost. As such, the two families will coexist in the market for some time, as AMD admits that many customers will continue to deploy his EPYC Milan chips in low-priority systems.
Meanwhile, the most compute-intensive and memory-bandwidth and memory-capacity-hungry workloads move to Genoa. Genoa has a higher initial cost, but the TCO benefits are fully realized due to the higher performance per watt and rack density shown in the slide above.
New SP5 sockets, platforms and chip designs are required to support these new features. Let’s move on to technical details, platform overview, and test results.