![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
-
Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.
AMD Zen Performance Preview
- Thread starter HardOCP News
- Start date
whats the point of amd slowly moving away from GF if they are hit with the same problem with samsung.
I thought GF just sucked, but there was news about nvidia shrinking pascal to 14nm with samsung?
SighTurtle
[H]ard|Gawd
- Joined
- Jul 29, 2016
- Messages
- 1,410
Is it possible the deal opens the door for TSMC to contract to AMD? Only thing I can see stopping that is TSMC being too busy with Apple.
they wanna make chips elsewhere for sure. They are willing to pay a flat fee quarterly fee if they don't fulfill wafer requirements at GF. That is telling on its own, looks like they have other plans going forward for their higher end products it seems where GF isn't delivering it for them
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
Samsung and Glofo is the same in this regards. Both are failures. 14LPP actually comes by stealing from TSMC to put it mildly.
TSMC is the only way to get a better foundry for AMD. And had the WSA never existed in the first place, AMD would have made all its chips at TSMC. Xbox S is made at TSMC for example.
Nvidia and Samsung 14nm is nothing but a silly rumour.
But remember, TSMCs better foundry nodes isn't going to magically save everything. But it will add a slight improvement.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
Apple is soon moving to Intel anyway
But capacity was never really an issue
When you look at foundries and their abilities across all metrics. Intel is number one being far ahead. TSMC is a sure second and down far away is Samsung/Glofo. When you do a refresh on the same node, then the foundry quality is even more important.
![Zarathustra[H]](/data/avatars/m/9/9298.jpg?1723407683){kind=avatar}
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
Apple is soon moving to Intel anyway
But capacity was never really an issue
Capacity was the number one reason that was given for why GPU's stayed on 28nm for so long, because the smaller nodes capacity at TSMC, Samsung, etc. was all used up by the likes of Qualcomm and Apple, and Intel didn't used to manufacture for others.
Last edited:
EuphoricRage470
[H]ard|Gawd
- Joined
- Apr 15, 2011
- Messages
- 1,105
Nope.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
It was yield, not capacity. Hence it was a cost issue.
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
Well, yield is one part of effective capacity.
Total capacity x yeild = effective capacity.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
It wasn't a wafer capacity issue. Nothing prevented them from making GPUs. The only thing preventing it was in TSMCs case Nvidias willingness to pay what it would cost at the time.
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
Well, yeah, but that's all related.
It's not as if TSMC's smaller node lines were sitting dormant because Nvidia was unwilling to pay. There was simply someone willing to pay more for that same limited resource. if there had been more of the resource available the, the price wouldn't have been bid up as high.
So, in the end it really boiled down to capacity again. Or rather effective capacity, because the effective capacity was unexpectedly low due to yield issues.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
You can add capacity as needed.
Intel always had the tech to beat AMD, PIII was the IP they fell back on for the Core line of processors, they just took two big misstep with P4 and Itanium. Prior to Athlon 64 AMD was just reverse engineering Intel chips and making modifications to improve performance.
For them to create a CPU that will take on Intel's best minds when AMD's resources are limited, that is a tough call,
Kyle even though they might have gotten their IPC fixed up, there are other areas where Intel might still have a significant advantage. Branch prediction is an area were Intel has certainly taken a lead and has dedicated quite a bit of transistors for every new generation. AMD has had some trouble here with BD but they have fixed those problems with Excavator, but still lack in this department. This will directly affect IPC and throughput in over all application performance, so at this point we don't know if this is what AMD has been showing or are they showing things in ideal circumstances for them.
End result if they can't get the clock speed up and their branch prediction isn't that much improved, they will run across benchmarks that will hurt them.
For them to create a CPU that will take on Intel's best minds when AMD's resources are limited, that is a tough call,
Kyle even though they might have gotten their IPC fixed up, there are other areas where Intel might still have a significant advantage. Branch prediction is an area were Intel has certainly taken a lead and has dedicated quite a bit of transistors for every new generation. AMD has had some trouble here with BD but they have fixed those problems with Excavator, but still lack in this department. This will directly affect IPC and throughput in over all application performance, so at this point we don't know if this is what AMD has been showing or are they showing things in ideal circumstances for them.
End result if they can't get the clock speed up and their branch prediction isn't that much improved, they will run across benchmarks that will hurt them.
Pieter3dnow
Supreme [H]ardness
- Joined
- Jul 29, 2009
- Messages
- 6,784
I'm sure you can dig up someone that can miraculously beat Intel , did not know they are all in Israel ? Maybe AMD does not have enough money.
Sadly beating Intel does not fix things for AMD, when you have contra revenue funds still playing a part in OEM market.
Pieter3dnow
Supreme [H]ardness
- Joined
- Jul 29, 2009
- Messages
- 6,784
You said only one thing here that makes sense. clock speed where you get the idea that everything is leveraged around branch prediction means that you have inside knowledge about the design or just making a comment as general as without gasoline a car does not drive ...
That leaves the last thing you said which is another open door, there will be benchmarks that won't go well for AMD. For every iteration of Intel cpu programs are compiled that does not happen for AMD and it is not the end of the world back then nor is it now if some or most benchmarks don't make full use of Zen ....
Nobified[H]
[H]ard|Gawd
- Joined
- Mar 21, 2013
- Messages
- 1,095
Where is AMD's inspiration? Where is their innovation to dream? Where is AMD's ability to think outside the box? Go where no man has gone before? This is why I liked AMD, because I felt inspired!!!!!!
Sadly this is where I see America now!
Sadly this is where I see America now!
LOL, do you know why the BD IPC was so bad, it wasn't just the shared FP unit. You might want to look at Excavator's tests and see why they are getting a 15% IPC improvement over BD, it was because of branch prediction fixes. AMD even mentions this, they had issues with it.
The Bulldozer Aftermath: Delving Even Deeper
Intel has at least a 2 level Branch predictor since I think Nehalem or Sandy bridge, they haven't talked about details of the algorithms they are using, but all we know is they have gotten much more sophisticated since then, Intel hasn't divuldged any of that information.
AMD doesn't go into what they are doing either, all we know is they have failed horrible in the past. Nested loops with dynamic branching creates havoc on AMD CPU's.
And also please read up on what AMD has done and what Intel has done.
Short history lesson.
AMD, Cryix, Nexgen all were reverse engineering Intel CPU's, till the 386's and then Intel gave AMD a license to produce Intel socket compatible chips because they need more volume in the channel, then pulled AMD's license soon after, that was when they were forced to create the Athlon which its derivatives came from the PII!
Last edited:
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
Where is AMD's inspiration? Where is their innovation to dream? Where is AMD's ability to think outside the box? Go where no man has gone before? This is why I liked AMD, because I felt inspired!!!!!!
Sadly this is where I see America now!
Well, even during the AMD K7 heyday, they were really not innovating all that much. They were recycling both engineers and tech from Digital's Alpha platform.
K6 before that was mostly developed by NexGen and acquired by AMD as it was nearing completion.
They only bought Nexgen because their in-house designed K5 sucked so bad.
Even the AMD64 design introduced with K8 was an evolutionary one, rather than the more revolutionary Itanium design which attempted to do away with legacy instruction sets in the name of efficiency.
They were first to APU's, but everyone was going that way, and - again - it was an evolutionary process, not a revolutionary one.
AMD has never really been an innovation company. Sure they've had a few highlights, like their full on charge when it comes to HBM, and their HSA designs but mostly they are a tried and true implementer of existing tech.
When they tried to do something truly different (Bulldozer, with core clusters, shared FPU's etc.) it flopped spectacularly.
I like AMD. My favorite memories of building systems and overclocking came during the K7 years when they were competitive with Intel, in part due to Intel's flop with Pentium 4's Netburst and in part due to a sudden timely influx of talent and tech from DEC. I love rooting for the underdog, but the truth is, that in the real world, the underdog usually loses.
Last edited:
Nobified[H]
[H]ard|Gawd
- Joined
- Mar 21, 2013
- Messages
- 1,095
Intel did one thing right, they hired Israel engineers to solve Intel's cpu problems before they pushed past Core2Duo. Sometimes it helps to get a fresh set of thinking, besides recycling old engineers!
Kidding aside, that was one of the most critical decisions made by Intel during their flop.i think this might be a case of selectable memory or very weird friends.
Cyrix Cx5x86 came out in 1995
Intel 80486 in 1990.
Intel Pentium in 1993
Intel Pentium Pro 1995
So everyone was buying 80486 at least 2 years after Pentium came out ?
Kinda weird talking about horrific taste preferences and then making a comparison between chips that are 5 years ( 2 architecture generations) apart to glorify ones own CPU.
BTW doom got release in 1993 or again 3 years after 80486 or smack in line with the Pentium age.
Intel did one thing right, they hired Israel engineers to solve Intel's cpu problems before they pushed past Core2Duo. Sometimes it helps to get a fresh set of thinking, besides recycling old engineers!
Lots of good stuff out of Israel
girls, soldiers, guns, engineers, food.
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
Well, I certainly couldn't afford the latest tech when I was a kid either.
I only upgraded my 286 to a 486sx25 in 1994 (or was it 1993?) when the 486 had already been on the market for 5 years and the Pentium had been out for two years.
I didn't have a pentium until like 1997 or so when PII's were already on the market.
And I was generally in line with my friends as well. I didn't know anyone in the 90's who was on the latest gen CPU.
It seems like people here in the U.S. got cheaper deals on the latest gen CPU's than I did as a kid in Sweden. I remember hanging out on IRC and wondering how the hell everyone could afford Pentium II systems, until I moved to the U.S. a few years later and saw how cheap the systems were in the big box stores here.
Yeah in the US the prices weren't that bad, top end systems 3k, but they would last ya 4 years easily. So the turn around wasn't bad.
I had Intel since the 8088 till a 486sx 25, then went to a Nexgen 586 Pentium clone, but had some issues with it, cause the OS I was running think it was Windows 3.1 or 3.11 for word groups, can't remember , needed a bios update to get it to work, so that took around 3 months to get out there. Then went back to Pentium Pro (IBM system) , then Pentium II, then Athlon 64's.
But all the systems were around 2500-3000k and they lasted 4 to 5 years.
I had Intel since the 8088 till a 486sx 25, then went to a Nexgen 586 Pentium clone, but had some issues with it, cause the OS I was running think it was Windows 3.1 or 3.11 for word groups, can't remember , needed a bios update to get it to work, so that took around 3 months to get out there. Then went back to Pentium Pro (IBM system) , then Pentium II, then Athlon 64's.
But all the systems were around 2500-3000k and they lasted 4 to 5 years.
The same would go for the guy that had the Cyrix as well. He was clearly making an oranges to apples comparison.
I uprade form an 386sx to a pentium 133. my class mate upgrade form an 486 to a pentium. Ee where several people on pentiums before pentium pro came out , which is the time periode for cyrix Cx5x86, and even reg upgraded to pentium mmx when that came out.
Not saying your case is not what it is. But claiming that everyone bought 486's in the CX5x86 time is clearly not correct
And I believer denmark is more expensive in general than Sweden ( except for alcohol). Thats why we built the freaking bridge to go get cheap swedish dentist. Little did we know we would have all those drunk Swedes coming the other way
It really wasn't that their branch prediction was bad, it was actually quite good. AMD just had a major disadvantage compared to Intel due to lack of micro-op cache. Basically if a Bulldozer cpu miss predicted a branch it would have to wait until the next clock cycle to process so it had a huge penalty for miss prediction. Intel could miss predict the branch but since it has mirco-op cache it could just pull the right prediction from cache on the same clock cycle with out having to wait for the next clock cycle. This results in a much smaller penalty in performance.
The biggest thing that hurt bulldozer when it launched was the lack of programs that could take advantage of extra cores. The 8150 would best the 2500k when those extra cores were fully being used, but it got destroyed when it couldn't use them. That landscape still exists somewhat today, however a-lot more programs these days can utilize more than 4 cores. Same thing held true with the 8350 vs the 3570k, then Intel released 4770k, 4790k, and now the 6700k, and AMD is like hey look we have the 9590 a design from 2012 still.
I think a-lot of users are eager for Zen, but we all learned from the bulldozer episode that we will wait for benchmarks before buying a supporting motherboard.
I think you guys all just need to chill and wait for concrete information (and real benchmarks after Zen is released) because there's so much speculation on top of speculation going on here. It's like if this happens and then this happens and if this other thing happens Zen will fail...or be great. There is a lot more that we don't know than we do know right now. Maybe you all should use these great rigs you have and go play some game instead of going gray worrying about Zen.
That is true,
Many are eager for Zen, and I expect it to do well, also expect branch prediction misses to be fixed, hopefully, but what they have showed so far they don't get into that. if they showed something like a heavy compute program utilizing dynamic branching running on Intel vs Zen, then we could have seen it also if they went against BD vs Zen that would have shown us too.
The blender showing doesn't show IPC, Blender typical shows both IPC and mutli core performance, so we can't get any idea of Zen's IPC improvements to that, nor the comparison to broadwell show us that either.
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
The only thing we have right now that is even semi reliable is that +40% IPC number AMD is throwing around. If we assume that AMD is right and we believe them, this means that the best we can expect if they can manage high clocks is something in line with Sandy Bridge. The worst we can expect with low clocks is somethign only marginally better than their current stuff.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
New GB3 leak with the Diesel platform.
diesel - Geekbench Search - Geekbench Browser
diesel - Geekbench Search - Geekbench Browser
Last edited:
Zarathustra[H]
Extremely [H]
- Joined
- Oct 29, 2000
- Messages
- 42,217
And what is Diesel?
I'm guessing a server platform?
1.4Ghz though. Expected for server platforms, but it is alarming that as the time gets closer we aren't seeing anything at about 4ghz.
They will have to hit 4Ghz to have anything even remotely competitive in the enthusiast desktop segment.
luminousone
Weaksauce
- Joined
- Feb 13, 2011
- Messages
- 124
I believe that is supposed to be the 32core 64thread with 64meg of lv3 cache and 8 memory channel server cpu. Clearly their is some sort of issue with that sample.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
Server platform, 1.45Ghz base clock.
Dreamers got a harsh reality to wake up to.
Dreamers got a harsh reality to wake up to.
luminousone
Weaksauce
- Joined
- Feb 13, 2011
- Messages
- 124
Even if Zen has bad core/socket scaling that bench doesn't make sense, its to extreme a hit. And the single thread is very suspect as well, I would bet someone is rewriting their processor identity string and sending fake results in.
I actually really like that server platform, its great for vm setups with unRAID.
back to Zen desktop though.
8c/16t models and the 4c/8t models. We know Zen clocks to at least 3ghz and will fit inside a tdp of 95watts, and 65 watts. AMD showed they can match intel broadwell in one workload. The only true thing we need to know now is what will actual clock speeds be? is it going to have some strange turbo features? don't quote me on this but I think they are gonna call it AMD Turbo core 3.0 I mean how original would that be right?
If people can push zen to 4+ghz then its going to be fast. That's a pretty big if right now.
back to Zen desktop though.
8c/16t models and the 4c/8t models. We know Zen clocks to at least 3ghz and will fit inside a tdp of 95watts, and 65 watts. AMD showed they can match intel broadwell in one workload. The only true thing we need to know now is what will actual clock speeds be? is it going to have some strange turbo features? don't quote me on this but I think they are gonna call it AMD Turbo core 3.0 I mean how original would that be right?
If people can push zen to 4+ghz then its going to be fast. That's a pretty big if right now.
luminousone
Weaksauce
- Joined
- Feb 13, 2011
- Messages
- 124
It is also possible that only one of the instructions decoders can handle 32bit bytecode streams, I mean why waste the transistors. Who knows, I want 64bit benchies ran by a third party that is trustable.
They are not putting a half gigabyte of l3 on it, naples of built from an MCM module that is 4 8c/16t modules tied together. Someone is uploading garbage to geekbench.
They are not putting a half gigabyte of l3 on it, naples of built from an MCM module that is 4 8c/16t modules tied together. Someone is uploading garbage to geekbench.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,677
GB3, GB4 and AOTS all showed the same now. And its very far from what AMD demoed.