News Intel-powered Aurora supercomputer fails to dethrone AMD-powered Frontier on Top500 list, again — claims spot as fastest AI supercomputer with HPL-...

[TheHerald]

You have to read those charts like someone who's actually wanting to buy one. In other words, look at the lowest-price sellers. If you do that, you'll see what I'm talking about.

I did, up until almost January the 7950x flatlined at 700$, the 13900k was between 620 amd 650.

To put things in perspective, the new zen 5 9700x might not even surpass the 12700k in MT perfomance. By then it would be a 3 generations old midrange cpu....

If you are happy with that amount of improvement, fine, I'm not.

Or to put things into an even better perspective, 13900k increased MT performance by 40% within a year at same power and same node (okay, technically the power limited increased by 13w). The 7950x also increased Mt performance by 40%, but it took 2 years, a node shrink and almost double the power draw over its predecessor.

Those are the facts.

 

[thestryker]

I did, up until almost January the 7950x flatlined at 700$, the 13900k was between 620 amd 650.

13900K:

7950X:

Those are the facts.

Seeing as you couldn't manage to correctly read price history charts...

Intel did drop their prices first... by about 3 weeks

 

[bit_user]

To put things in perspective, the new zen 5 9700x might not even surpass the 12700k in MT perfomance.

I'm not going to waste time & energy speculating. Once they launch, I'm simply going to look at the competitive landscape and how corresponding products compare with the ones of similar street price, performance, and power.

13900k increased MT performance by 40% within a year at same power and same node (okay, technically the power limited increased by 13w).

The node was Intel 7+, even though Intel doesn't call it that. However, they make a point of saying "Improved Intel 7", suggesting they internally consider it to be a different node.

 

[Avro Arrow]

Isn’t it also AMD?

Yes it is. AMD's supremacy was kinda the point I was making. 😉

El Capitan will be powered by a currently unknown number of AMD Instinct MI300A APUs. The MI300A, while being an APU, is nothing like consumer market APUs because at this level, there's no tradeoffs like weakening both the CPU and GPU to allow them to share a heat spreader. More directly, the MI300A has 24 EPYC Zen4 cores and its IGP is Markham's pride and joy, the Radeon Instinct GPU with 224 CDNA3 compute units.

Tom's Hardware reported two days ago that these 224CUs will have a total of 14,592 CDNA3 stream processors. but I think that's inaccurate. ATi compute units have had 64 stream processors each since GCN 1.0 was introduced and that trend continued through RDNA, RDNA2 and RDNA3. It would make a lot more sense if CDNA also had 64SPs per CU because it would equate to 14,336 (224 × 64 = 14,336) which is within margin of error compared to 14,592. The other reason that it would be extremely unlikely for 224 CUs to have 14,592SPs is the fact that 14,592 cannot be divided evenly by 224 (14,592 ÷ 224 = 65.143?).

Now, I know that those numbers don't mean much to most people, hell, they didn't mean much to me when I first read them either. Without a frame of reference when looking at numbers like this, they seem to lose all meaning. So, I figured that the best I could do to get some idea of what these numbers mean was to use my own RX 7900 XTX as a reference point. It's probably not perfect because I don't know how different RDNA and CDNA really are but it's the closest thing that I could think of and the RX 7900 XTX is the most powerful consumer-grade video card that ATi has ever designed. The GPU on that card is the most powerful Navi 31 variant, the Navi 31 XTX. The Navi 31 XTX has 96 RDNA3 compute units which amounts to 6144 RDNA3 stream processors.

Let that sink in for a second... the MI300A APU's Radeon Instinct IGP has 2.3x as many stream processors as the mighty RX 7900 XTX. As mind-blowing as that already is, consider that each of El Capitan's server blades will have EIGHT of these MI300A APUs. Heaven only knows just how many blades that this colossus will have.


My mind has just melted into sludge trying to fathom the computational power that this supercomputer will have. Maybe the next one will have so many blades that they'll have to call it El Hecatoncheire! 🤣

 

[bit_user]

My mind has just melted into sludge trying to fathom the computational power that this supercomputer will have.

I stopped trying to wrap my head around what kind of performance these GPUs like MI300 and Blackwell can achieve. At some point, the numbers just kinda lose meaning to me, especially when you consider they don't have any graphics rendering capability and I don't have anything else I'd want to run on them.

A couple things I do still look at are the amount of on-die cache they have and their memory bandwidth.

As far as supercomputers go, I'm a little interested in how certain computational tasks are split up amongst the nodes. Unless your workload scales to run on them efficiently, it almost doesn't even matter how many nodes they have.

 

[helper800]

I stopped trying to wrap my head around what kind of performance these GPUs like MI300 and Blackwell can achieve. At some point, the numbers just kinda lose meaning. A couple things I do still look at are the amount of on-die cache they have and their memory bandwidth.

The amount even consumer cards can do is mind boggling. We have super computers from the 80s-90s in our pockets, let alone what our 3-4 slot monster cards in our desktops can do...

 

[bit_user]

The amount even consumer cards can do is mind boggling. We have super computers from the 80s-90s in our pockets, let alone what our 3-4 slot monster cards in our desktops can do...

I think about this a lot, actually. I care about efficiently utilizing what I have, before I worry too much about upgrading it.

It's honestly getting to the point where I'm replacing machines based on age and my concern that the next OS upgrade might start to have issues with such old hardware, rather than really needing more performance.

 

[helper800]

I think about this a lot, actually. I care about efficiently utilizing what I have, before I worry too much about upgrading it.

It's honestly getting to the point where I'm replacing machines based on age and my concern that the next OS upgrade might start to have issues with such old hardware, rather than really needing more performance.

The only things that make me upgrade are if i get a new monitor at a higher resolutions or refresh rate of that resolution, a new game that requires such power that I cannot get satisfactory performance, or my previous card dies. I usually end up getting a new card or monitor every 4-8 years. Most of the time a new monitor does not need a new graphics card. I think my next monitor will be one of my last with these Ultrawide OLED's at 240hz+ and 0.1>ms response times.

 

[eX_Arkangel]

So you think that Intel is having to sell their products at less of a margin due to effort because the margin is lower. That's some absolute nonsense if I've ever seen it. No point in continuing going in this circle.

he's been trying the hole time to flip or turn around Intel's failures and call them "accomplishments", basically every "negative" point you argue about Intel he tries to turn it around as positive and if he can he moves the goal post. Every fact or data you showed of intel failing he tried to turn it around or deflect.
He wants to see (tries to) Intel's fall from grace as a story of ongoing success and a simple shift to more profitable markets (server to consumer client) imagine that, what level of reality distortion he's using on himself to accommodate those assumptions is beyond me.
Having read several articles and several post of him about Intel losing to AMD i have to say is a trend of his not specifically tied to this particular article.