My predicted 3rd Gen threadripper prices

[dragontamer5788]

..the only thing Intel still has in petto is their Optane/3D XPoint memory stuff...

--
Srdja

There will be a number of people who want to use Intel vTune, and other Intel-specific debugging / performance tools. AMD's performance counters are much weaker than Intel's. In particular, AMD requires IBS mode (instruction-based sampling) for accuracy, which is severely non-intuitive to most programmers. Intel in contrast has PEBS, which more easily tracks branch-prediction, cache-misses, and other events with ease.

With that being said: AMD Fans will do themselves some good by learning IBS-profiling through uProf. It isn't anywhere near as good as Intel PEBS, but IBS still gives good information about "why" your code is fast or slow.

------

Intel also has AVX512, but I think I'd rather have more-cores than wider SIMD. Especially now that AMD caught up to 256-bit SIMD pipelines in Zen2.

[smatovic]

..the only thing Intel still has in petto is their Optane/3D XPoint memory stuff...

--
Srdja

There will be a number of people who want to use Intel vTune, and other Intel-specific debugging / performance tools. AMD's performance counters are much weaker than Intel's. In particular, AMD requires IBS mode (instruction-based sampling) for accuracy, which is severely non-intuitive to most programmers. Intel in contrast has PEBS, which more easily tracks branch-prediction, cache-misses, and other events with ease.

With that being said: AMD Fans will do themselves some good by learning IBS-profiling through uProf. It isn't anywhere near as good as Intel PEBS, but IBS still gives good information about "why" your code is fast or slow.

------

Intel also has AVX512, but I think I'd rather have more-cores than wider SIMD. Especially now that AMD caught up to 256-bit SIMD pipelines in Zen2.

Yes, in both cases AMD could catch up, but afaik AMD has nothing in pipe to counter these NVDIMMs by Intel.

--
Srdja

[Dann Corbit]

Yes, in both cases AMD could catch up, but afaik AMD has nothing in pipe to counter these NVDIMMs by Intel.

--
Srdja

Is there anything that prevents someone from popping a few of those NVDIMMs into their shiny new AMD box?

[smatovic]

Yes, in both cases AMD could catch up, but afaik AMD has nothing in pipe to counter these NVDIMMs by Intel.

--
Srdja

Is there anything that prevents someone from popping a few of those NVDIMMs into their shiny new AMD box?

Yes, Intel. Only Cascade-Lake chips have support for Optane DC NVDIMMs. And currently they aim the Enterprise market.

But who knows, maybe it takes only some years until a third party develops some shiny NVDIMMs for everyone

--
Srdja

[Zenmastur]

I don't see the 24 core higher than $1,000.

There are 1.5 times as many cores in the 24 core TR as in the 16 core Ryzen.
$1000 is only 1/4 more than $750.
You would be nuts to buy the 16 core part, when the far more powerful part (1.5x the core count) is only $1000.
But we should know the answer to that question in just a few hours.

I will be shocked, stunned and amazed if they are $1000, and waiting in line to buy a pocket full of them.

I guess you were right on the CPU pricing. Which leaves AMD with a gaping hole in their product stack. Namely the price points between the $750 Ryzen 9 3950X and the $1400 Ryzen Threadripper 3960X. I suppose they could keep the Threadripper 2950X and 2970X alive to fill this gap in the interim. But I would think that eventually they will have to fill it with newer SKU's. The question is what would you fill the gap with? I suppose they could fill it with a 20-core part. Seems a bit awkward though. I guess we'll have to wait and see. If they don't fill it, INTEL is sure to find a part to slide into the middle of AMD's stack to disrupt it. E.g. drop the price of the i9 10980XE and add a 20-core or maybe a 22-core part (a cut down W-3175X at higher clock speeds) in the $1000 to $1300 range.

Regards,

Zenmastur

[Dann Corbit]

I guess you were right on the CPU pricing. Which leaves AMD with a gaping hole in their product stack. Namely the price points between the $750 Ryzen 9 3950X and the $1400 Ryzen Threadripper 3960X. I suppose they could keep the Threadripper 2950X and 2970X alive to fill this gap in the interim. But I would think that eventually they will have to fill it with newer SKU's. The question is what would you fill the gap with? I suppose they could fill it with a 20-core part. Seems a bit awkward though. I guess we'll have to wait and see. If they don't fill it, INTEL is sure to find a part to slide into the middle of AMD's stack to disrupt it. E.g. drop the price of the i9 10980XE and add a 20-core or maybe a 22-core part (a cut down W-3175X at higher clock speeds) in the $1000 to $1300 range.

Regards,

Zenmastur

I think it will be hard for AMD to fill those slots because the high-end CPUs come in 8 core chunks for those high end systems (maximally populated dies, and all the dies get used, less N, because you get 8 cores per die):
1x8=8 cores with one die
2x8=16
3x8=24 (This one is really the oddity! Since they bothered to make it, the only odd numbered die count after 1)
4x8=32
5x8=40 (This one is possible)
6x8=48
7x8=56 (this one is possible)
8x8=64

I think the 4 and 6 core (3200/3400/3600 stuff) is a different type of die, or some cores are dead on it. Most likely a different die.

I don't think AMD can cough up 28, 30, 36 core ... configurations easily.
But I doubt that it matters.
Intel is surely selling at a loss with their new prices.
And they could discount to the point of giving the chips away.
But that is a loss of many thousands of dollars per chip.
And absurdly low prices will just mean that people will vacuum up the chips like a 10,000 foot tall hoover deluxe.
They would replace all the infrastructure in their server rooms with the new stuff.
Now, during this lull, AMD would have to tough it out.
But then, when Intel's inventory hits bottom, they would be in far more serious trouble that they are now.
Currently, they are already having problems meeting customer demand:
https://www.pcmag.com/news/371532/intel ... nd-maybe-i
And AMD seems to be running like clockwork, though a wee bit behind schedule.
And how long would the Intel shareholders put up with billion dollar losses?

I think that right now AMD is in the cat bird seat.
And the best Intel can do is whack them a few cans of corn, after the initial bloody nose.

I guess my odd metaphors are going to miss with the European audience.
"Cat bird seat" means huge advantage, position of control...
"Can of corn" is a baseball metaphor that means a really easy play (soft fly into the outfield right at a fielder).
"Hoover" is a brand of vacuum cleaner

[dragontamer5788]

3x8=24 (This one is really the oddity! Since they bothered to make it, the only odd numbered die count after 1)

Makes me think its 6x4 instead, with 4x "dummy die" for thermal / mechanical purposes only. No CPU cores on "dummy dies".

[Zenmastur]

I guess you were right on the CPU pricing. Which leaves AMD with a gaping hole in their product stack. Namely the price points between the $750 Ryzen 9 3950X and the $1400 Ryzen Threadripper 3960X. I suppose they could keep the Threadripper 2950X and 2970X alive to fill this gap in the interim. But I would think that eventually they will have to fill it with newer SKU's. The question is what would you fill the gap with? I suppose they could fill it with a 20-core part. Seems a bit awkward though. I guess we'll have to wait and see. If they don't fill it, INTEL is sure to find a part to slide into the middle of AMD's stack to disrupt it. E.g. drop the price of the i9 10980XE and add a 20-core or maybe a 22-core part (a cut down W-3175X at higher clock speeds) in the $1000 to $1300 range.

Regards,

Zenmastur

I think it will be hard for AMD to fill those slots because the high-end CPUs come in 8 core chunks for those high end systems (maximally populated dies, and all the dies get used, less N, because you get 8 cores per die):
1x8=8 cores with one die
2x8=16
3x8=24 (This one is really the oddity! Since they bothered to make it, the only odd numbered die count after 1)
4x8=32
5x8=40 (This one is possible)
6x8=48
7x8=56 (this one is possible)
8x8=64

I think the 4 and 6 core (3200/3400/3600 stuff) is a different type of die, or some cores are dead on it. Most likely a different die.

I understand the dies are the same and they either have a defective core on the CCD or they are disabled. I don't know for sure but I read that the CCD's can't be made to work with odd core counts. I.e if one CCX has only three working cores then the other CCX must have one core disabled. This only allows CCD's with 2, 4, 6, or 8 cores. So in addition to the configurations you listed there are these”

1x6= 6 cores aka R5 3600-3600X
2x6=12 cores aka R9 3900X
3x6=18 cores possible?? I guess (this might make a good filler SKU for the thread ripper line, it's better than a 16-core R93950X and would have much better I/O
4x6=24 cores aka Threadripper 3960X
5x6=30 cores possible??
6x6= 36 cores possible??
7x6= 42 cores possible??
8x6= 48 cores aka Threadripper 3980X

I don't think AMD can cough up 28, 30, 36 core ... configurations easily.
But I doubt that it matters.
Intel is surely selling at a loss with their new prices.
And they could discount to the point of giving the chips away.
But that is a loss of many thousands of dollars per chip.

First, Intel isn't selling their chips at a loss! They have such HUGE margins on their chips it's a JOKE!

Data I've seen suggests that Intel could drop the price on the i9 10980XE to below $500 / 1K units and still make a profit. (I.e. Intel's cost to produce 10980XE compatible chips is <$400 per chip. So, the potential is there to drop the prices enough so that a slot between the 10980XE and the 3960X could be filled with a SKU designed to disrupt the AMD product stack. Or better yet leave the 10980XE the same price and insert a higher performance chip in the $1200 to $1400 range even if they have to sell it at cost. The point would be to kill the sales of as much of AMD's HEDT stack as possible thereby robbing them of the profits of their high margin chips.

The hole AMD needs to fill is between the R9 3950X with only 2-channel memory and limited I/O capability and the Threadripper 3960X with 4-channel memory and double the I/O capabilities. So, core counts higher than 24-core don't seem to fit the bill.

And absurdly low prices will just mean that people will vacuum up the chips like a 10,000 foot tall hoover deluxe.

Intel is smarter than that! If they wanted maximum disruption simply announce a 22-core to 24-core part (as an example) price them less than the 3960X and then limit it's production.

Currently, they are already having problems meeting customer demand:
https://www.pcmag.com/news/371532/intel ... nd-maybe-i
And AMD seems to be running like clockwork, though a wee bit behind schedule.
And how long would the Intel shareholders put up with billion dollar losses?

Apparently they are planning to turn a profit even while spending $3B to disrupt AMD profit potential. They're not in any danger of losing money any time soon.

I think that right now AMD is in the cat bird seat.
And the best Intel can do is whack them a few cans of corn, after the initial bloody nose.

Maybe, but that's not to say Intel can't disrupt AMD's plans just to buy some time. Which is very likely what they plan on doing!

Regards,

Zenmastur

[Dann Corbit]

Wikipedia says the Zen 2 chiplets have 8 cores:
https://en.wikipedia.org/wiki/Zen_2
I think it is only the older Ryzen 5 stuff that has multiples of 6.

As far as production costs, you must also factor in the design and engineering costs, building the silicon fabs, advertising, etc., far more than the cost of making a chip.

I do not believe that Intel is charging ten times too much for their chips. It would make it too easy for others to compete with them.
The ARM guys and IBM would have overtaken them, along with AMD.

[dragontamer5788]

Wikipedia says the Zen 2 chiplets have 8 cores:
https://en.wikipedia.org/wiki/Zen_2

They do by design. But when a core is defective, AMD will sell it as a 6-core, 4-core, or even 1-core chiplet.

For example, the 8-core EPYC 7262 has 8x chiplets, 8x cores (1-working core + 7-defective cores per chiplet) and 128MB of L3 cache. https://www.servethehome.com/amd-epyc-7 ... ge-caches/

The huge aggregate L3 cache is useful in some workloads, and very cheap because 56 of the 64 cores are broken.