Sapphire Rapids: Golden Cove Hits Servers | Hacker News Reader

Sapphire Rapids: Golden Cove Hits Servers

chipsandcheese.com

40 pointsgautamcgoel22 comments

Intel must feel pretty bad about messing up the whole chiplet thing since that advantage allowed AMD to overthrow them in just under a couple of years. And it seems like they still seem to be adjusting into it since compared to Zen 4 SKUs, SPR seems like a horrible design that won't scale well (and doesn't according to pretty much all benchmarks).

AMD can simply make a smaller core and release a new CPU with 128 cores (Zen 4C -- higher density core clusters consisting of smaller Zen 4 cores [with less L3 than a normal one]). On the other hand, for nearly every Intel CPU it is a whole new design/engineering mess.

I hope they don't fall to behind on this, otherwise we might see yet another monopoly (e.g. AMD not releasing a consumer grade Threadripper is the first indication that they kind of feel safe on this side since even their multiple years old SKUs offer far more value brand-new SPRs).

Intel's mesh is superior in some ways and I think they could scale it to 96C (4x24C) if they wanted to. There are rumors that Granite Rapids is moving to a slightly different layout with I/O dies bookending 2-3 compute dies. Ultimately AMD's chiplet architecture is cheaper, simpler, and faster for almost every app though.

The max all-core turbo frequency of Intel's flagship 60-core Sapphire Rapids model is 2.60 GHz. It costs $17,000.

The equivalent AMD 9554P processor has 64 cores, and has a base clock of 3.10 GHz for just $7,100.

The AMD processor's boost clock is 3.75 GHz, which means it is about 50% faster if you factor in the extra 4 cores and assume that both processors have roughly the same IPC.

The price/performance difference is just staggering.

Sapphire Rapids was supposed to compete against Milan where it would have had a chance (with massive discount). It was delayed so long that it's now pointless.

The public clouds will still buy it, mainly because AMD cannot manufacture enough 9004 series processors to meet their demand. Large orgs also like to have two sources for each component to prevent excessive lock-in, price gouging, and other counterparty risks.

There aren't any genoa in public clouds, so if sapphire rapids arrives in general production soon it will indeed by head-to-head against milan. And in my experience even ice lake beats milan.

Azure has 176 vCPU instances as private preview, but only in the "HPC" sizes: https://azure.microsoft.com/es-es/blog/do-more-with-less-usi...

Apparently they're waiting for the v-cache enabled version to be available and they will buy those in bulk. I heard they'll be deploying them this month.

Genoa is delayed due to issues with its memory controller: https://semiaccurate.com/2023/03/13/a-bit-more-on-amds-genoa...

Most clouds do want huge instances.

Performance depends on which SKU the clouds run, which can differ depending on what they optimize for. In the past, most clouds have offered AMD as a "cheaper option" to Intel and so that's how it's been spec-ed. As this changes things might look a bit different.

(If Intel charges $17K when they're losing, imagine what they would charge without competition.)

Did we ever see a definitive rationale for ending consumer grade Threadripper? I’ve always wondered if sales were simply cannibalized by the growing thread counts in Ryzen. A sort of Innovator’s Dilemma, except with more vocally unhappy consumers.

I think it wasn't just the growing core counts on the mainstream platform, but also the death of multi-GPU gaming and steady growth in DDR4 speeds and NVMe SSD capacity and performance—all of which combined to eliminate the need for a larger socket with all the extra IO.

It's not definitive, but AMD also had a capacity shortage so they allocated their capacity to more profitable Epyc.

Threadripper was nearly pointless. What they really need is some SP5 workstation boards, but what does that even get you that you don't get from the server boards? Integrated audio? Just add a sound card.

The real disappointment is the clock speeds on the lower core count Epyc SKUs. Epyc 9124 and Ryzen 9 7950X have the same configuration and Epyc has a higher TDP but it's 3GHz/3.7GHz instead of 4.5GHz/5.7GHz.

And the reason for it is that the Epyc SKUs with more cores have almost the same TDP, which they couldn't hit with twice as many cores at 4.5GHz -- the 9334 has a base clock of 2.7GHz, and probably not artificially. Who is going to pay ~$3000 for a 32-core 2.7GHz 9334 if there was a ~$1000 16-core 4.5GHz 9124? It would be only 20% faster for three times the price. So they artificially limit the 9124 for market segmentation.

Then there is nothing available with the higher clock speeds that has more I/O, 12 memory channels and RDIMMs. Which seems like leaving money on the table. I mean, sell something on SP5 with the 7950X's clock speeds for some price. WTF.

Assuming they had fab capacity to produce a hypothetical Epyc 13912k with 4.5GHz base clock, I think AMD still nailed the revenue sweet spot with their current mix. So maybe they're not strictly leaving money on the table, so much as it is a lack of demand for a high end desktop that doubles as a server.

Intel has to produce that kind of beast because they're currently slightly behind in some areas. AMD might lose money if they followed Intel's playbook right now.

They could sell the 7950X but on SP5 and for $4000, or with twice as many compute dice but the same clocks as the 7950X and the TDP of the 9474F for $8000. Put a 24-core in the middle for $6000 and call these the "workstation" line.

At those prices there would be no money to lose because nothing close to equivalent would be cheaper, but then at least it would exist.

It's really interesting how slow the rollout of these chips has been - there are so many steppings! There really must have been intense engineering challenges. I found A B0 Sapphire Rapids CPU on eBay for 100$ back in September 2021!

Does anybody know if these cloud instances of SPR have QAT and DSA? The accelerator story for this generation seems more interesting than the core CPU story.

Apparently all of the Intel Sapphire Rapids CPUs have the accelerator modules, they're just enabled by a software license. A safe bet is that all of the hyperscalers would have them enabled, otherwise these Intel CPUs are pointless and totally uncompetitive with 9004 series AMD chpis.

The question is whether they are passed through to VMs. Even with accelerators, I have been wondering why cloud providers even bother with Intel unless they're getting the processors basically for free.

They most likely would. There are about half a dozen DB vendors that have announced support, including Microsoft's own SQL Server 2022: https://cloudblogs.microsoft.com/sqlserver/2023/03/09/sql-se...

The other DBs with support are all "cloud era" technologies that are infrequently deployed on-prem.

Possibly does DBMS benefit from SPR L3 cache structure?

Reply on news.ycombinator.com