AMD「Venice」Epyc 與歷代 AMD 伺服器處理器相比表現如何
How The 'Venice' Epycs Measure Up Against Prior AMD Server CPUs
這篇分析比較 AMD 第六代「Venice」Epyc 與前五代 Epyc 及 Shanghai Opteron,聚焦頂級和中階型號的效能及性價比。Epyc 9996 SP7 相較 Shanghai Opteron 2393 SE 的效能提升 190X、單位效能成本低 14.9X;Epyc 9746 SP8 的效能提升 108.2X、性價比改善 10.8X。
Over the past decade, each generation of Epyc processors from AMD is more complex and that complexity is leveraged not because the company’s top techies are baroque, but because it takes actual engineering distinctions to address the widening number and kind of workloads in the modern datacenter.
With the “Venice” generation of Epyc chips that launch today – the sixth generation and one that will once again put more water between itself and X86 CPU rival Intel – AMD is bringing to bear two different server sockets, two different I/O dies, and two different compute dies to create a product line that initially has 31 different CPUs initially, spanning from 8 cores to 256 cores, 130 watts to 600 watts, and from $700 to $14,904 when purchased in 1,000-unit trays from the company.
NEXTPLATFORM AD
We have already discussed how AMD believes the Venice Epycs address today’s workload needs and also very generally how they compare to alternatives from X86 rival Intel and the Arm collective. This story walks through the Venice SKU stack and how the top bin and middle bin parts compare to the prior five generations of Epycs and the touchstone “Shanghai” quad-core Opteron processors were launched way back at the beginning of the Great Recession in November 2008.
The Salient Feeds And Speeds
Before we dive into the SKUs, we need to talk a little tech.
Rather than keep the Zen 6 and Zen 6c designs in distinct processor families within the Venice line, as AMD has done during the prior “Genoa” and “Turin” generations, the company has decided to blur the distinction between the Zen 6 and Zen 6 cores by keeping the cache hierarchies the same but in optimizing the schedulers inside the two cores in different ways – one optimized for higher core counts and lower clock speeds on the core compute dies (CCDs) and the other optimized for lower cores and higher clock speeds. With the prior Turin Zen 5c cores, AMD cut the L3 cache on each CCD in half and that allowed it to put more cores on the die but also crimped its performance on cache-sensitive workloads. This new approach seems sensible. Both the 12-core CCD, which has 48 MB of L3 cache, and the 32-core CCD, which has 128 MB of L3 cache, are etched using 2 nanometer processes from Taiwan Semiconductor Manufacturing Co.
NEXTPLATFORM AD
The other big change is that there are two different sockets, and both kinds of CCDs are used to make some of the CPU SKUs in both sockets.
The SP7 socket, which is shipping now and which makes heftier Venice processors, has a pair of I/O dies that deliver 16 channels of DDR5 memory that can deliver up to 1.6 TB/sec of bandwidth using MRDIMMs running at 12.8 GT/sec and 1 TB/sec using regular RDIMMs running at 8 GT/sec. Each channel can have one memory stick and they can be up to 512 GB in capacity. (If you think you can afford that. . . .) The SP7 socket has 96 PCI-Express 6.0 lanes running at 64 Gb/sec each with eight PCI-Express 4.0 lanes running at 16 Gb/sec; in a two socket configuration, 128 PCI-Express 6.0 lanes plus sixteen PCI-Express 4.0 lanes are exposed to the outside world and 64 of those PCI-Express lanes are used to make a NUMA link between a pair of CPU sockets.
This SP7 socket I/O (and memory) die is made using 6 nanometer processes from TSMC, just like the I/O (and memory) die for the Turin generation was, but this is clearly a different design. As we have pointed in the past, there is not a lot of benefit from using the most advanced processes to etch I/O and other signaling chiplets because I/O does not shrink as well as compute does. Such a shrink makes signaling more difficult and is also more expensive – two good reasons to hang back on the manufacturing process.
Here are the nine Venice chips that plug into the SP7 socket:
NEXTPLATFORM AD
As we have done with past Epyc processors, we have calculated the relative performance of the Venice chips based on core count, clock speed, and improvement in instructions per clock (IPC) from generation to generation for integer workloads. This time around, the average performance uplift per core at the same clock speed for the Zen 6 core compared to the Zen 5 core, according to AMD, is 20 percent. This is a relatively big jump and is a testament to the core designers who are always coming up with new ways to get the core to do more work each cycle.
By the way, the Epyc 9G76 processor highlighted in blue font and red background in the SP7 socket processor table above is the one AMD has made specifically for the Helion AI racks. It is not particularly different from the Epyc 9686F chip aimed at HPC workloads, but I think they are distinct so they can be discounted differently. The hyperscalers and cloud builders are no doubt going to get better deals than the HPC centers of the world, given the high volumes they are purchasing.
Moving on to the SP8 socket, which is used for more mainstream Venice Epyc processors. And as such, it has only eight DDR5 memory channels per socket (four per I/O die) with two DIMMs per channel interleaving to boost capacity. The I/O die for the SP8 socket has 128 PCI-Express 6.0 lanes and eight bonus PCI-Express 4.0 lanes. So, half the memory bandwidth but 33 percent more I/O. In a two-socket NUMA setup, 64 PCI-Express 6.0 lanes are used to make the NUMA coherency and 192 lanes are exposed to the outside world plus those sixteen PCI-Express 4.0 bonus lanes. (That is for the pair of I/O dies, of course.)
AMD has confirmed to us that the SP8 I/O (and memory) die is etched with TSMC 6 nanometer processes, as we expected.
NEXTPLATFORM AD
Here are the 22 variants of the Venice Epyc CPUs outfitted for the SP8 socket:
The SP8 socket versions of the Venice CPUs will not ship until next year, and it is my guess that it will happen fairly early in the year although officially all that AMD is saying is that they will ship sometime in the first half of 2027.
Comparisons Are Absolutely Not Odious
With that out of the way, lets see how the top bin parts and middle bin parts across the history of the Epyc line stack up in terms of performance and price/performance, and as usual we are using that Shanghai Opteron line as the baseline against which all comparisons are made. These were pretty much the last competitive Opterons that AMD shipped, which came out in the wake of the “Nehalem” Xeon 5500 chips where Intel copied an awful lot off of AMD’s homework and finally got a decent multicore chip out the door just as the Great Recession was starting and AMD was having issues with the Opterons.
Bad timing for AMD, good timing for Intel. But that is all ancient history now, and this time around AMD is not leaving any crack to exploit as AMD just keeps making better and better server CPUs and eats a little more market share with every passing quarter. It is a wonder that Intel has been able to hold onto the market share is does get, to be honest.
Let’s start with the top bin parts:
In these comparisons, I ignored the special HPC variants that AMD has shipped from time to time, which are denoted with a F designation for frequency optimized and an X designation of they have extra vertical cache on the package. So these are top bin standard parts.
Also, I put top bin parts in the comparison for both the SP7 and SP8 variants in the Venice lineup.
As is immediately obvious in the chart above, the relative performance of the top bin Epycs have scaled linearly, more or less, from the “Naples” Epyc 7001 series all the way through the Venice 9006 series with the SP8 socket. The SP7 socket with 256 cores is a total outlier in terms of the performance boost, and it takes the throughout performance up an exponential scale as you can see with your own eyes.
What also pops out instantly is how the bang for the buck of the first generation Naples processors improved compared to those Shanghai Opterons, but you have to remember that the gap in time is nine years – nine IT years where Moore’s Law was still available to boost performance and improve economics. (The Shanghai Opterons were etched in TSMC 45 nanometer processes, the Naples chips used GlobalFoundries 14 nanometer processes.)
What you can also see is how aggressively AMD moved in terms of price/performance with the “Rome” second generation Epycs, which is when the company first moved to distinct compute tiles and an I/O die linking them together and to the outside world. With the Milan Epycs, performance went up for the top bin part, but value for dollar improvement slowed and the same thing happened again with the fourth-gen “Genoa” top bin parts. But with the Turin and Now Venice parts, AMD is once again pushing the price/performance envelope, and it is particularly awarding customers who go with the throughput-oriented SP7 variants. AMD is charging a slight premium for the SP8 versions of Venice, which have half as many cores that run a little bit faster.
Clock speed is always going to come at a premium because that adds heat faster than increasing cores after a process shrink.
The net result is that the Epyc 9996 SP7 part has 190X more oomph than the Opteron 2393 SE, and if offers that performance at a cost per unit of performance that is 14.9X lower. For AMD, the benefit is that the chip costs 12.8X as much and companies are happy to pay for it for server consolidation and agentic AI and traditional HPC workloads. For the Venice Epyc 9746 SP8 chip, the performance jump compared to the Shanghai Opteron 2393 SE is 108.2X, the bang for the buck has improved by 10.8X, and the price per chip is 10X higher.
The curve for middle bin server CPUs from AMD looks similar if a little flatter in some ways:
But once again, the value for dollar with Venice is a lot better than it was compared across prior Epyc CPU generations.
AMD is not only keeping the heat on Intel. It is turning it up. And it has to if it is to compete against the Arm collective that is rising among the hyperscalers and cloud builders who account for half of the server market these days.
來源:The Next Platform · nextplatform.com