AMD's EPYC Milan Breaks Cinebench Record, Here's a 10nm Ice Lake Xeon Comparison | Hacker News Reader

AMD's EPYC Milan Breaks Cinebench Record, Here's a 10nm Ice Lake Xeon Comparison

tomshardware.com

105 pointsItsTotallyOn63 comments

Cooling: "Liquid Nitrogen" reminds me of the amazing day when somebody first overclocked a slot based AMD Athlon to 1.0 GHz using a custom refrigerant loop system.

Cooling with Liquid nitrogen is in some ways simpler than a refrigerator cycle since it’s completely impractical to compress & recondense it. A cryogenic cooling system usually just vents the gas into the room, it may just be a little well around the part which gets filled with ln2 from a dewar.

> A cryogenic cooling system usually just vents the gas into the room, it may just be a little well around the part which gets filled with ln2 from a dewar.

It depends on the scale, and whether the cooling is only required for a short while.

Large scale cryogenics (such as at CERN, an LNG tanker etc) will recompress the vapor as it forms.

If so, the loop basically IS as refrigerator, just a much bigger, higher performing unit.

For an extreme example, see: https://home.cern/science/engineering/cryogenics-low-tempera...

True, I should have qualified the phrase impractical by stating at this scale. it’s an industrial scale process that because of the pressures and temperatures involved isn’t practical to use for anyone who isn’t permanently installing it at significant capital outlay - which, wouldn’t be practical for cooling any cpu other than maybe a supercomputer.

Indeed so, I remember seeing people do temporary "overclocking world records" with things like 900 MHz Pentium 3s, and Athlons, around twenty years ago, bare motherboard on an open table with liquid nitrogen dewar mounted to the CPU socket. Guys pouring more nitrogen into the thing as it evaporates and they run benchmarks. I guess the procedure for that probably hasn't changed very much, since it's a very ephemeral and temporary process.

Youtube channels like Gamer's Nexus are still out there doing this exact kinda thing. I'm still as fascinated by it now as I was back then.

Related from a little while ago: https://news.ycombinator.com/item?id=24280218

Do you need some form of sonicator to keep bubbles from sticking?

Not in practice. I deal with LN2 cooled optical detectors, and the bubbles clear themselves out pretty quickly. Sometimes folks add a little bit of sand to help the bubbles nucleate.

Man what a throwback, memories I forgot I had. I was all about upgrading my K6-2 450 to a 550, but was equally fun to cheer on AMD's Slot A chips that had begun to erase Intel's dominance at the time.

The peak of tech excitement in my entire life was when a friend brought a copy of Computer Shopper to school, and the cover proclaimed a new 90Mhz 486!! No other tech ever blew my mind like that again.

Likewise - I remember someone walking into art class at school with a computer magazine featuring at 100mhz 486 DX4 on the cover. Triple figures!

I have to wonder if computer parts are less exciting now because of the lack of generational improvement or the fact that I'm an old fart that can't spend a bunch of time geeking out over the next generation.

I really followed computer evolution during the GHz wars. Watching a CPU go from 300Mhz being fast to 2GHz being fast was amazing. Then watching as AMD pulled 1.8GHz parts and kicked the pants off of Intel's 2.5GHz parts was equally incredible to young me.

Now I see generation to generation pushing 5->10% improvements and mostly go "meh".

Do you remember being able to run DooM and Winamp at the same time!?

Upgrading from 450 to 550? I once had an AMD Duron 650, which I overclocked very late in its life to 850MHz simply by increasing the FSB from 100MHz to 133Mhz (or so), no extra cooling or anything.

Yeah maybe today 100MHz difference might not seem like much, but it's a >20% increase which is nothing to sneeze at! I did a similar one (K6-2 366MHz to K6-2 550MHz).

Those lower-clocked Durons you mentioned were fantastic value though - I remember buying a 700 MHz part and running it at 900 MHz (after using a pencil to short those traces on the chip that let you manipulate the clock multiplier)

The biggest overclocking story of all time was when people discovered that (many/most?) 300Mhz Celerons where intentionally mislabeled 450Mhz ones, because apparently the yields where such that they couldn't not make mostly 450Mhz ones.

Those werent miss-labeled, as there were never 450MHz Celerons. This overclockability was due to excellent 0.25um lithography process Intel was wielding at the time. Every Intel CPU had same MHz ceiling, even older laptop 200MHz Pentium 1 MMX Tillamook CPUs overclocked to over 400MHz https://www.youtube.com/watch?v=fqB4yCssUCM. This was also one of the reasons for Intel to start fighting with overclockers by locking multipliers.

Thanks for the correction, I remember indeed it was as you said, simply that they were capable of much higher frequencies.

I didn't know that it was the case with other models, I guess I was mainly eyeing Celerons at that time because they were so cheap.

Also reminds me of this old video (early 2000s) that shows what happens when you remove the heatsink. This is what cemented that image in my head of “AMD=cheap knockoff” that AMD reversed only 2-3 years ago (and now I am trying to source an EPYC Milan!).

https://www.youtube.com/watch?v=YYQSHXNFvUk

What you probably never saw was the non-retraction retraction Tomshardware had to publish later, and buried so deep even archive.org doesnt have access anymore. Yes, they manually excluded the link from wayback machine :o

Here is the whole story: https://www.tomshardware.com/reviews/hot-spot,365-4.html somehow still accessible under its original 2001 link unlike most of their old material, wonder why :)

>"AMD did not bless the Thunderbird core with ANY thermal protection whatsoever."

In reality AMD socket certification required thermal cutout, same as Intel for Pentium 3. AMD processors do include thermal diode just like Intel ones.

>"Intel's older processor is also equipped with a thermal diode and a thermal monitoring unit"

is a lie. Pentium 3 thermal throttling is performed by the Bios just like in Athlon case. Serendipitously tomshardware picked broken Siemens motherboard for AMD system as recommended by Intel, imagine that.

They published a non-retraction catching Siemens in a lie and showing proper AMD setup safely shutting down. Of course you cant read it because its buried down, excluded from wayback machine and deadlinked http://www.tomshardware.com/column/01q4/011029/index.html

This exact video had a big impact on my impressions of those companies as well, it's pretty amazing how long those views last.

The Kryotech SuperG!

http://www.thg.ru/cpu/19991115/print.html

That was mindblasting for me at the time and remember I dreamed of buying one.

Now I have a multi-ghz Ryzen 270p.. not even the latest and gratest and take it for granted.

Those were the days...

It's unfortunate that the images seem to be gone now, because I remember it being quite a feat of engineering for its era.

It looks like there's some shots of it here: https://www.anandtech.com/show/432/2

LN2 is pretty standard for extreme overclocking. Requires a lot of preparation due to water condensation and some blow torch heating for the system to start up as it can fail to boot entirely.

It's fairly simple and most pundits can do it at home if they so desire.

Sadly, I’ve had to stick with Intel, even with the performance penalty. Why? The massive AMD core counts were killing us on licensing costs. Until we can renegotiate our contracts, we remain an Intel shop.

Wait, what?! I thought the license/core was a thing in the UNIX/NT days, when running more then one core/system meant you were doing serious mainframe/datacenter business, but you're telling me this is still a thing now in the days when a smartphone has 8 cores?

I wish I could recall which software service this was, but when I was handling infra for a small biotech startup a lot of price evals for metrics or CI/CD tools would still penalize us for core counts.

IIRC AMD offers high frequency lower core count SKUs. Would those not work for your situation?

The 72F3 was made almost specifically for this use case, though it's comparable (and may even fall behind per-core) to the Xeon 8356H.

That was my preference but we honestly had trouble with our first order of them (delivery kept slipping) for testing and then our preferred vendors struggled to get sufficient inventory.

My hope is to get the contracts reworked in such a way that I’m not paying more for less if I go AMD with the next procurement.

100%. I wish you luck in the midst of this chip shortage!

Time to switch to another software stack.

I suspect this is about oracle.

Not an easy migration of you are deep in their ecosystem.

It looks like your vendor is not keeping up with the times.

it doesn't specifically we are a dell shop and they always priced their amd epyc server's higher. I'm not sure why, maybe they also added their board design development costs on top, where as intel they have more knowledge, so it's cheaper? no clue, but sometimes it's really wierd that we get really good prices for intel, but not for amd.

Look up how Dell deals with AMD. Just the other day there was a discussion how Alienware (Dell) is misconfiguring and downplaying their own AMD desktop systems. You'll also be hard pressed to find a high-specced AMD laptop from Dell, and their best AMD machines are well hidden on their own website.

Dell is definitely antagonistic towards AMD.

They did a ton of design work during the original Opteron days when it was performance king, only to have it be wasted as AMD flubbed the next gen and circled the drain for 10 years.

Vendors are noticeably cautious this time around, but the longer AMD maintains the crown the more system integrators are starting to invest. Nobody is trying to actively harm AMD - they've just been burned once before.

To be clear, I don't think they are trying to actively harm AMD. It sure seems like they have a better deal with Intel though, and actively try to sell those. There's few other explanations for self-sabotaging the AMD configurations they already have in the selection.

Isn't agility the key these days? I know very few people who purchase Intel these days if they have a choice. If they can't buy it from one vendor, they go to another one.

Intel still has 60% market share so the majority are still buying Intel, but the trend is very clear. Expect more AMD options to come out in the coming months.

That is the cost of sticking to a single vendor.

One reference point: my home built Ryzen 3700x machine gets almost exactly 100,000 points less than the system they tested.

That's laptop territory for zen3 :(

I urgently need a new computer but have been holding for cheap zen3 machines to become available since the difference in performance is so huge.

I'm looking forward for the scores of the Zen3 AMD threadripper. I'm just not sure if I should upgrade as soon as they are out or if I should wait the next-gen for DDR5...

Funny maths, "As you can see, in Cinebench R23, the dual EPYC Milan 7763's are 34% faster than the dual Ice Lake Xeon 8380's."

Yeah, I also wondered about that, isn't it more like 52% faster?

Indeed, the EPYC is 52% faster than the Xeon, which makes the Xeon 34% slower than then EPYC.

Depends on how you define the math of "x% faster than"

The English language defines the denominator in your equation as the object of the sentence.

If you have 100 blue smarties in a bowl, and there are 20% more red smarties than blue, then you have 120 red smarties. There is no ambiguity.

I think that percentages shouldn't be used for things like this.

Show an absolute time, how many units/second of work.

percentages just (intentionally??) confuse things.

the problem is that %faster and %slower are not equal

(113000 - 74000) / 113000 = 34%

Incorrect. 74k should be in the denominator, that is what's being compared against (baseline).

Yes, but that's the opposite denominator from usual. If I say I got 50% more apples from my trees this year, I can only really mean (this_year - last_year)/last_year = 50% (i.e. 1.5 times as many).

> As a quick reminder, AMD's flagship 7763 server chips come armed with the 64 Zen 3 cores and 128 threads apiece and have a 2.45 GHz base and 3.5 GHz boost frequency. All told, we're looking at a Cinebench run with 128 cores and 256 threads, which you can see in the tweet below:

Can someone clarify this wording?

It has 64 cores which can each do 128 threads? That doesn't seem right... but then later it says it has 128 cores

1 processor has 64 cores with SMT giving 128 hardware threads.

In a dual socket configuration that is:

2 processors, 128 cores, 256 threads.

Dual 7763s, each with 64 cores (128 threads). The "apiece" modifies chips, not cores.

With all the hype around new efficient laptop CPUs I had to check how they compare to this monster...

This thing: 113K

Best AMD laptop CPU: 11K

Best intel laptop: 7.5K

Beat Apple laptop: 7.5K

That is 12x to 15x slower, which honesty was better than I expected.

If you take this CPU with its 280W TDP and scale it down to 18W like an Apple M1, you get a score of ~7200 while the Apple M1 gets 7700. Actually pretty amazing how close the performance per watt is to be honest; totally different form factors, different CPU architecture, different R&D (both having spent billions of dollars), different fabs, but still within 10% performance/watt. Makes you wonder how close we are to the physical limits.

Yeah, but don't forget that creating systems with massive core count is much more than just scaling up. It is a whole different science and art. Just handling the heat generated by 280W requires teams of hundreds of engineers. And IPC for 8 cores is child play compared to IPC for 64 cores.

You can already see this with M1: the 10W M1 reaches 7.5K while the 18W (15?) M1 is IIRC around 8.7K on cinebench

Anandtech says during their testing that pretty much all the chips were in the 220w range.

At 18w and 7,833 points, that's 435 per watt for the M1

At 440w for 2 chips and 113,631 points, that's 258 per watt for the Milan

At around 140w and 10,096, that's about 74 per watt for the 5950x

At around 75w and 4,517, that's about 61 per watt for the 5600x

At around 140w and 5,994, that's about 43 per watt for the 5800x

Frequency scaling is TERRIBLE for PPW (performance per watt). The screaming almost 5GHz 5800 cores lose efficiency to the downclocking 5950 cores by almost 2x. Vs the 2.5GHz Milan, it's right at 6x difference.

Apple is on the right track pushing wider rather than faster, but that seems like a path that is hard for x86 to take due to decoder power and complexity (decode units literally eat up more die space than the actual integer math units). It isn't all roses though. A huge portion (30-50%) of the Milan power budget is the interconnect. If you double the TDP on the M1 to simulate the interconnect, PPW drops to around 217 which is about 16% worse than AMD.

We also haven't factored in how AMD gets a huge performance boost with a quarter-gig of cache. Apple likely has to scale up their cache too. 16mb and 8 cores is 2mb/core. 256mb/64 is 4mb/core. But because it's multicore and all cores don't necessarily do the same job at the same time on a server, M1's deal with the devil starts to fade. It will need even more cache per core to keep up with the ultra-wide execution path (else lose PPW) which will increase total power and further reduce max clockspeeds. M1 is a great chip in laptops, but if scaling up were easy, everyone would be doing it.

https://www.anandtech.com/show/16529/amd-epyc-milan-review/5

https://www.anandtech.com/show/16252/mac-mini-apple-m1-teste...

https://www.anandtech.com/show/16214/amd-zen-3-ryzen-deep-di...

Keep in mind that every vendor defines TDP differently, linustechtips have highlighted this many times and even has a video on it [1].

So comparing TDP across manufactures is not a fair comparison.

[1] https://www.youtube.com/watch?v=yDWO177BjZY

Probably large chunk of that difference is consumed by interconnect (infinity fabric)

Reply on news.ycombinator.com