AMD Wants To Stop Being Known As The “Cheaper Solution”

[Shivansps]

I never said that 250X is not faster, i said it was more expensive at $100.

And yes the Core i3 is very overpriced for low budget gaming systems, have a look at the Athlon 860K + 250X price. Also A8-7650K alone at $105 is faster in MT when OCed and way faster in iGPU at lower price than the cheapest Core i3.

Also, at 4.3GHz the A8-7650K can keep up with the Core i3 even paired with a 290X or 980. And you dont need extra cooler to get the Kaveri at 4.3GHz.

It may cost $100 but you need to consider ram prices, you can game with 4GB-1333 ram+250x, you cant do that on a 7600... 4GB is already very little if you also need to use the IGP, and you need DC.
Im sure that 860K+250x+4GB not only wins every time, but it may end up behing cheaper than a Dual Graphics config.

About the I3 i dont know what to tell you, the last 2 games that ive checked benchmarks(GTAV and Cars) the I3 do very well vs AMD offerings if using a big dgpu like a 960/970, thats its likely to change once DX12 is deployed, but today i dont see anything wrong with them.

 

[JDG1980]

I think the market is pretty clear that 'APUs' are just not in demand. Any modern CPU now has 'good enough' graphics for web-games or lite-graphics titles. Anything more similar to a console title will require significant IQ/resolution downgrades or a discrete GPU. Period.

APUs in their current state for PCs are not in demand, and probably won't be. The fundamental problem is that there aren't many people who need a GPU stronger than Intel's solutions, but considerably weaker than even a $99 discrete card. As I have said before, it will take affordable HBM to make APUs viable. GDDR5 could be used as a stopgap measure (it works fine in the PS4 APU) but that would move manufacturing expense to the motherboard side. As long as the GPU side is bottlenecked by standard DDR, you're never going to get viable performance.

 

[JDG1980]

Had they performed as well as Haswell, nobody would have quibbled over a few watts (at least not in the consumer market). The killer is that AMD's reputation is so bad right now that people who aren't involved in forum's like this probably think AMD is ten years behind Intel.

But is that perception even wrong? AMD has yet to come up with an x86 architecture that can match the IPC of a 2008 Intel Nehalem. Clock for clock, AMD's current designs barely measure up to the Core 2 Duo from 2006. It's really hard to overstate how serious a miscalculation Bulldozer was.

If AMD can meet their expectations on Zen, it will boost IPC to near Sandy Bridge levels. Assuming a 2016 release, that shrinks the gap from 8-10 years behind Intel to 5 years behind Intel. That's not perfect, but it's good enough to get them back in actual competition - especially since Intel's own IPC growth has slowed since SB.

 

[DrMrLordX]

As long as the GPU side is bottlenecked by standard DDR, you're never going to get viable performance.

We haven't even seen how Carrizo's color compression scheme reduces the DDR3 bandwidth bottleneck. That, and standardizing DDR3-2400 (or switching to DDR4) could alleviate much of the iGPU bandwidth bottleneck.

btw, I do not think Su is specifically lying about AMD's lineup for 2015. AMD is eating a lot of crow on the desktop, but that market is shrinking for everyone. She is most likely trying to be upbeat about Carrizo and the upcoming 300-series video cards.

 

[TheELF]

If AMD can meet their expectations on Zen, it will boost IPC to near Sandy Bridge levels. Assuming a 2016 release, that shrinks the gap from 8-10 years behind Intel to 5 years behind Intel. That's not perfect, but it's good enough to get them back in actual competition - especially since Intel's own IPC growth has slowed since SB.

They already stated that they will have smt on every core,a huge hunk of that 40% more IPC will come from that,so not faster but more throughput,which is to be expected from a company with a long history of many-core cpus and high throughput cpus.
Hopefully for them they will get at least the smt right,intels first couple of implementations where pretty crapy.

 

[AtenRa]

They already stated that they will have smt on every core,a huge hunk of that 40% more IPC will come from that,so not faster but more throughput,which is to be expected from a company with a long history of many-core cpus and high throughput cpus.
Hopefully for them they will get at least the smt right,intels first couple of implementations where pretty crapy.

IPC and SMT are two different things. 40% IPC gains have nothing to do with the SMT implementation.

 

[pablo87]

2GB 250X is $100

+ $78 for the Athlon and you are $23 more expensive than the APU + 240.

So as i have said above, you pay more you get higher performance.

Edit: ahh yes there is a 2GB version on AR at $80

So an Athlon 860K at $78 + 250X 2GB GDDR5 at $80 AR = $158

This also shows how overpriced those Core i3s are.

The X4 860K is the best deal for a value gaming build. $69.99 with promo code.

Add a uatx mb for $40 = $110 for the platform, hard to beat.

 

[ShintaiDK]

IPC and SMT are two different things. 40% IPC gains have nothing to do with the SMT implementation.

Depends how they claim it. A lot of IPC gains for the current ones ended up as MT only.

It wouldnt be the first time they pulled such a trick for the pivot. Its not for fun a lot of people says this is AMDs last chance.

 

[JDG1980]

They already stated that they will have smt on every core,a huge hunk of that 40% more IPC will come from that,so not faster but more throughput,which is to be expected from a company with a long history of many-core cpus and high throughput cpus.
Hopefully for them they will get at least the smt right,intels first couple of implementations where pretty crapy.

IPC is a per-thread metric. SMT doesn't help with that.

AMD knows full well what their weaknesses are, which is why they specifically focused on IPC with that big slide. This is one of the things that gives me some hope for their future - you have to admit what you're doing wrong before you can change, and it looks like they've done that. Note that the Zen slides also discuss an all-new cache system, and we know that Bulldozer cache latency was one of the big contributors to poor IPC.

 

[poofyhairguy]

What is sad is Intel has pretty much hit a IPC wall. If AMD can find a way to get 2011 Intel IPC at 4+GHz they are golden.

 

[ShintaiDK]

What is sad is Intel has pretty much hit a IPC wall. If AMD can find a way to get 2011 Intel IPC at 4+GHz they are golden.

Incorrect:

 

[Ajay]

IPC is a per-thread metric. SMT doesn't help with that.

AMD knows full well what their weaknesses are, which is why they specifically focused on IPC with that big slide. This is one of the things that gives me some hope for their future - you have to admit what you're doing wrong before you can change, and it looks like they've done that. Note that the Zen slides also discuss an all-new cache system, and we know that Bulldozer cache latency was one of the big contributors to poor IPC.

I thought I read somewhere that Kaveri already matched Nehalem's IPC - and that is without an L3$. I could be remembering incorrectly, since I don't have any data in front of me. Someone has figured this out already, I just need to find the data again. Please feel free to correct me if I am wrong.

FWIW, I think that Zen needs to match Haswell's throughput with a better perf/watt ratio to stand a chance. I just don't think 14FF will be able to hit high enough clocks to do any better. And, ass I've said b/4, it is a very tall order. I'm glad that AMD left the iGPU off of Zen so that they could focus all their energy on CPU performance (and cut down die size, verification time, etc.).

 

[Ajay]

APUs in their current state for PCs are not in demand, and probably won't be. The fundamental problem is that there aren't many people who need a GPU stronger than Intel's solutions, but considerably weaker than even a $99 discrete card. As I have said before, it will take affordable HBM to make APUs viable. GDDR5 could be used as a stopgap measure (it works fine in the PS4 APU) but that would move manufacturing expense to the motherboard side. As long as the GPU side is bottlenecked by standard DDR, you're never going to get viable performance.

I had predicted, a year or two ago, that iGPUs would displace all but enthusiast gaming cards (>~$200 US) by the end of this decade. But, it's looking like neither AMD, Intel or the memory makers are moving fast enough on a HBM solution for consumer CPUs to hit this time frame. I think that part of the problem is that most people don't really need that kind of GFX performance. It will be interesting to see what happens.

 

[Enigmoid]

What is sad is Intel has pretty much hit a IPC wall. If AMD can find a way to get 2011 Intel IPC at 4+GHz they are golden.

Again what is with this belief?

People need to reexamine conroe and the CPUs in the last 10 years.

IPC improvement is pretty much constant since core 2. There was no increase, there was no decrease in the rate. What did happen was that clockspeeds crept up only to reach a wall with SB; HT was also added.

You can find a variety of tests that show similar (take care to allow for integration of the memory controller in Nehalem - why clarkdale does poorly).

Look at the wolfdale tick for instance, its the same as IVB or BW at ~5%.

http://www.anandtech.com/show/2306/3

Intel then hit a bunch of trouble with 32 nm delays (like 14 nm). Nehalem dramatically improved in some areas but was a bit of a mixed bag.

The Core i7's general purpose performance is solid, you're looking at a 5 - 10% increase in general application performance at the same clock speeds as Penryn. Where Nehalem really succeeds however is in anything involving video encoding or 3D rendering, the performance gains there are easily in the 20 - 40% range. Part of the performance boost here is due to Hyper Threading, but the on-die memory controller and architectural tweaks are just as responsible for driving Intel's performance through the roof.

http://www.anandtech.com/show/2658/20

People think rates stalled because clockspeeds stopped going up (the addition of HT also increases the perception of slowing down). SB was solid as well but a lot of the performance increases are due to the significantly increased clockspeed.

IMO I feel that intel could be doing more for IPC (skylake really should have doubled caches across the board - the cache system is due for a rehaul) but their rate of core IPC improvements are pretty much constant without any underlying trend.

Edit:

I thought I read somewhere that Kaveri already matched Nehalem's IPC - and that is without an L3$. I could be remembering incorrectly, since I don't have any data in front of me. Someone has figured this out already, I just need to find the data again. Please feel free to correct me if I am wrong.

Not true at all. Just look at games for instance. A i7-920 generally does very well against a 8350. PD/Kaveri is around Conroe levels in legacy (disregarding newer instructions) code and conroe doesn't have an IMC. Not sure why people keep insisting on L3. Kaveri has the same amount of total cache as an i3 (without the burden of an inclusive cache system). Look at nehalem vs. Core 2. IPC was a mixed bag due to the greatly reduced L2 (if the much improved memory controller wasn't there things could be ugly in some applications).

You can also compare the A6-7400k vs. the A8-7800 (same clocks and everything). 1 MB L2 vs. 4 MB L2. Single thread performance hit is around 5% and that is a cache cut of 75%. Same with ST IPC between an i3, i7, and i7E with 4, 8, and 20 MB cache. L3 cache really doesn't help much (and AMD stated as much when they launched trinity/llano without L3).

 

[Ajay]

Incorrect:

Right. What is really slowing to a crawl is ST throughput. If we had something like SPEC_int_rate for a singe CPU core over the past 20 years, it would be steep for the first 12 years and then probably start to look more asymptotic.