Read more.Quote:
Chinese tech site says that this 'Rome' processor is an engineering sample.
Printable View
Read more.Quote:
Chinese tech site says that this 'Rome' processor is an engineering sample.
Oh boy, AMD is really pushing.
I think their stock will (continue to) go through the roof.
So, this will be the top Threadripper (64 cores/128 threads) in 2020?
So does it being a 64C/128T lend credence to Zen 2 being designed around 4 core CCX's?
If so how's that going to work with existing sockets only supporting dual, quad, and eight channel DDR.
That works out to a score of 196 per core. That puts the Rome Zen2 on an equal per core performance to a CoffeLake 8700K.
As this is a server chip I doubt it's running at 4GHz let alone the 4.7GHz an 8700k can turbo to, then they must have made some Serious IPC gains over the last 18 months or so.
I read a rumour that they've changed the memory and IO controller design for Zen 2 so that the number of memory channels available isn't based on the number of CCX's wired in; there's a single IO/Memory controller chip surrounded by the CPU core chips on the interposer, instead of each CCX talking to its local memory & IO.Quote:
Originally Posted by Corky34
To be honest in my experience AMD have always seemed to have lower overheads when using multiple cores compared to Intel. Just look at the way threadripper/rizen scales with multiple cores compared with Intel.Quote:
Originally Posted by zaph0d
Even back at uni when I had a dual socket AMD MP1900 machine the fall off from '2x performance' when using both cores was far lower than the Intel options that were available at uni.
So I wouldn't necessarily say the gain is quite as high as you're suggesting but it's clear there's been a gain somewhere.
That had me confused so i had to do a quick refresher on current Zen design and unless I've got it wrong the memory controller isn't tied to the CCX's, it's on the same die but it sits outside the core domains. Then again I've just realised I got the design of the CCX wrong as a single CCX encompasses 4 cores already. :embarrassed:Quote:
Originally Posted by afiretruck
I expect there has been a lot of experience in this area from back in the days where they couldn't compete and were just throwing more cores at the problem. They probably got a lot out of trying to scale cores the FX6300 era chips.Quote:
Originally Posted by LSG501To be honest in my experience AMD have always seemed to have lower overheads when using multiple cores compared to Intel. Just look at the way threadripper/rizen scales with multiple cores compared with Intel.Quote:
Originally Posted by zaph0d
Even back at uni when I had a dual socket AMD MP1900 machine the fall off from '2x performance' when using both cores was far lower than the Intel options that were available at uni.
So I wouldn't necessarily say the gain is quite as high as you're suggesting but it's clear there's been a gain somewhere.
That would make more sense! I've been misusing the term 'CCX'.Quote:
Originally Posted by Corky34
They must be doubling up on the cores per CCX with 7nm, which sounds a little risky to me. They must be confident with their 7nm yields.
Intel ICE lake 64 core to be built on TSMC 7nm process to combat AMD and remain on top.
That sounds pretty far fetched, got a source for that?Quote:
Originally Posted by lumireleon
Because Intels roadmap puts Ice Lake on 10nm.
8c/12t on desktop now, give it 5 yeara and we could see this double within the same power/thermal limits and I think may struggle.
I wonder what the tdp on that beast is. VERY impressive,
Keep up the good work AMD, and perhaps we will have genuine advancements in high-end consumer CPU's.
So the TDP for the 7601 is 180W so it's potentially going to be less than double. I estimate around 250W which would align up with the 32-Core Threadripper and 7nm is meant to have a big thermal profile improvement so I would put money on 250-300w TDPs :)Quote:
Originally Posted by will19565
It would be no denser, thermally, than 2700E as a matter of fact. If they are able to fit 8C/16T at 45w TDP, this would only be a 180+180 watt dual-socket.
PS: Technically, it will act as a 48C/96T 3.9GHz Ryzen, imo.
I am really curious how you can align CPU groups on that with Windows, last time I worked with that they were masked by 64bit integer and you couldn't put more than 64 in one group...
(see https://docs.microsoft.com/en-us/windows-hardware/drivers/devtest/boot-parameters-to-test-drivers-for-multiple-processor-group-support )
From what i understand in that article, you just need to make another group. The max size of one group is 64 processors so it would stand to reason that you could just add another group. I would presume it would operate the same way for a Xeon Phi? Although, someone more versed in this would have to weigh in.Quote:
Originally Posted by HollyDOL
Or they're building the design with 4 4-core CCXes. The whole point about the CCX is that it's a modular block, and I don't see any good reason to redesign it when you can just slap more of them on a piece of silicon...Quote:
Originally Posted by afiretruck
Well, not really - current top-bin EPYC processors *are* essentially 4x Ryzen 7s on a substrate, and the TDPs for those range between 130W and 180W, so that's 32.5W - 45W per die. But that only gets you to 32 cores per socket.Quote:
Originally Posted by mtcn77
With the higher density and improved power efficiency of moving to 7nm it's perfectly feasible that they'll be able to do a 16 core die in that 30W - 45W TDP window and get 64 cores per socket in a 180W TDP, but it's all going to be conjecture until they release more details. Given we've already seen a move to 250W TDPs on Threadripper 2, I can't help thinking that there might to a top-bin EPYC 2 with a 250W TDP...
I think it would need to be 8 4-core CCX's as each die has two 4-core CCX's, either that or they'd need to increase the number of cores per CCX (at least i think that's how it is as I'm getting confused with how many cores, CCX's and dies are where. :))Quote:
Originally Posted by scaryjim
EDIT: Even reading that back I've confused myself, all i know is that each CCX has 4-cores and each die seems to have two CCX's (excluding low core count and low power CPUs) :crazy:
Yeah, current design is 4 cores, each with its own L1 and L2 cache, plus 8MB of L3 cache, per CCX. The base die (which I think is codenamed 'zeppelin') has 2 CCXes (and therefore 8 cores), a 2 channel memory controller, and 32 PCIe lanes. They stick 4 of those dies on a substrate to make a 32 core EPYC processor.Quote:
Originally Posted by Corky34
Assuming they're sticking with 4 dies for a 64 core 7nm EPYC processor, it'll need to have 16 cores per die. They could do that by changing the CCX design to have 8 cores in each CCX, but that would seem like an odd decision when the whole point of the CCX design was that it was modular so you didn't have to redesign the entire CPU to change how many cores it had...! It'll be much easier to build a 16 core die by simply putting 4 CCXes in it.
The interesting thing will be seeing what they do with the 7nm consumer line - will they use the same 16 core die and have huge consumer core counts? Will they do a smaller (i.e. cheaper) 8 core die? Will we see 8 core APUs? Can they afford to tape out multiple different dies for different market segments rather than using the one dies fits all strategy we saw with 14nm?
The 7nm process allows 45% boost in performance over zen1.Quote:
Originally Posted by zaph0d
If AMD is targeting this with clocks alone, then the clocks would have to go up by about 60%. Ryzen 1 needed 15% clock increase to gain 10% performance.
That of course does not take into account things such as IPC gains.
And also, we don't know how many changes the zen2 architecture underwent, so the scaling of clocks and performance could have been equalised for all we know.
Right now its too soon to tell. But the general rule is that 45% performance increase over ryzen 1 is doable at same TDP.
I suspect AMD will be targeting IF improvements,ie,looking to improve power draw so they can run it at a higher clockspeed,and it will be needed if they intend to add more CCX units.
Actually it's taken quite a crap recently, but that is because they have to prove you can make more than 100mil in a Q, and do it for enough times to get somewhere near your PE ratio etc. They are currently priced WAY above their last Q, but I agree with your statement over the long run as 7nm stuff ALL gets out the door. That will give time for it to sink in how much intel is struggling which I expect to go down more for a bit at least until they have a decent response announced and benched. You've lost the fab lead when you're outsourcing your 14nm to 3rd party fabs and talking 10nm LONG after the other guys are mass producing 7nm. Doesn't matter that one might be a bit better (we'll see), what will matter is AMD giving more cores, for about the same price as intel and AMD being FASTER in most stuff (if not all next round at 7nm) and meeting or beating Intel watts/heat. If you win in perf and match in watts/heat hopefully AMD is smart enough to charge exactly the same as Intel and capitalize on shortages! Or, heck, charge MORE than Intel if you win EVERYTHING almost across the board (games and apps). You basically have the next 2yrs or so to make hay while the sun shines then intel will have spent enough on FAB R&D to get back into this game now that they know they've fully blown in blowing 4B+ a year for 4yrs on mobile losses when those 4B EACH YEAR should have been put into FABS. We wouldn't be having this conversation if they had done it, 10nm would be done, and they would have already announced 7nm coming shortly. Or, heck for 16B, would you have needed another 8B for 10nm fixes? I'm guessing 4B (a years worth of direct fixing at first sing of issues?), the other 12 could have gone to 7nm and been done by now. My math could be off a bit, but not much I'd guess.Quote:
Originally Posted by darcotech
If AMD has a Q report that makes more than 300mil the stock will shoot up especially if guidance is even higher for next Q (need to head to 1B a year in income or more NET). But if it drops below ~108mil (whatever the last Q was roughly, IIRC it was 108 or so NET), then it goes down to 20ish. They punish AMD when it doesn't meet expectations usually. It's a great swing stock in many situations. Intel has been raising prices and moving production to just high end (due to shortage) mostly, but that won't work if AMD is BEATING them as noted above. It only works as long as AMD is BEHIND in many things still. I didn't buy because of heat/watts and too many gaming hits in things I wanted now or in the future to play. Too big a pill to swallow, but that won't happen I HOPE at 7nm. I'll kick out the 8700k in Q1-Q2 if any sale hits on AMD's new babies :) I've been waiting for this for a long time as I used to sell AMD and PUSH them in my business last time they were great (~1997-2003 were pretty good times with AMD leading for much of that end time). I tried to take every sale I could from Intel if AMD was a decent match for their work/game ambitions. In my defense, I was looking at no silk screening ASUS boards (no name on them) and then shipped in WHITE boxes only so you'd go Intel (Asus etc had fear of next chipset access loss, so no ASUS label or boxes), so it was me sticking it to the man so to speak...LOL.
Can't wait to buy them again (chips I mean, own the stock currently), but they have to WIN to get me to do that. I don't take the loser to do small company X a favor. My wallet takes winners only even if I have to save a bit longer to get what I want.
My 3cents. ;)
https://www.extremetech.com/computin...m-process-node
Example of what I was saying regarding ~4B to make a new process as noted here. Basically 4-5B to get it done according to Extremetech.
“3nm will cost $4 billion to $5 billion in process development, and the fab cost for 40,000 wafers per month will be $15 billion to $20 billion.”
So 7nm probably 4B and 10nm I guess probably would NOT have costed 4B to FIX it. I'm probably still a bit off here, but point made. Blowing 16B over 4yrs and NOT buying NV 5-6yrs ago both cost this fab loss. Imagine NV socs and gpus from Intel fabs 5yrs ago up to now...WOW. Arm sells 2B gpus a year now (and they game etc), so FTC etc wouldn't blink today like they would have before cell phones etc really took off. Between AMD/ARM there are a TON of gpus sold that are not NV/INTC (consoles in there too for AMD).
Chip prices seem to be rising soon as process costs and tapeouts go way up down to 3nm. Heck 5nm doubles the cost almost to tapeout/validate a chip etc vs. 7nm. Costs are RISING quickly. Maybe they can avoid rising costs by expanding markets. If a few other larger nations get better PC's (education, etc) over time, they might mitigate much of the costs over a larger base. We went from 350-380mil PC units, to ~2.3B mobile+PC (260mil now) which is massive gpu unit growth. Interesting times ahead for sure in tech.
NV reportedly considered themselves the more significant in an NV+Intel merger, so Jensen wanted control over the merged companies. Intel fabs aren't optimised for making GPUs either, so that might have been a big upheaval.Quote:
Originally Posted by nobodyspecial
An i7 with Nvidia graphics would have been nice though, if only for the drivers.
But back to AMD, back with the Athlon they had a problem with people buying Intel because it was what they knew and trusted, even when the P4 was so dire. These days there is a huge market for things like cloud servers, they can win that on competence not brand. I hope they don't blow it.
Edit: Oh an Intel have very deep pockets, the billions they blew on trying to get Atom into phones and tablets I'm sure had no impact on the billions they spent on fab R&D.
This kinda relates to GPUs but it shows that even if AMD make technically better kit, they still seem to lose whereas Nvidia can make relative tat and still turn a decent profit.
The same could equally be said between them and Intel. I'm far too lazy to look up figures from the Athlon 64 days but I expect Intel was still turning a decent profit with the P4 regardless of it being relatively awful.
Yep, I used loads of company P4 machines in those years and a grand total on one Athlon 64 box, and that was only because I demanded it. People in that company remarked on how fast the Athlon box was, and then went out and ordered another P4 from Dell.Quote:
Originally Posted by philehidiot
Lawsuits on why it was strangely hard to buy AMD kit forced Intel to dig out what to them was some spare change. https://en.wikipedia.org/wiki/Advanc..._v._Intel_Corp.
I remember that wherever I was at the time spent a fortune on hot running and slow P4 boxes and I just could not understand for the life of me why at the time. I'd bought an Athlon 64 for home and it was cheaper and better. To me these people were just plain morons. I can understand that it might have been harder to get hold of appropriately priced AMD based systems due to Intel's... bribery but in that case I'd have made the business case as the IT guy for building our own systems pointing out the cost savings to be had as well as the increased longevity of the systems due to higher baseline performance. It doesn't take long to sling together a bog standard PC for office use and if you've got a load of them the same you could easily get a production line going once every so often for a small - medium sized company and just keep a stock.Quote:
Originally Posted by DanceswithUnix
But I'm known for thinking too far outside the box and being told "that's a good idea!" to shut me up.