The Kaveri Pre-Launch Thread (A10-7800 and A10-6800k @3,5 Ghz)

[daxzy]

What's the difference between a Kaveri-K edition and non-K edition?

 

[NostaSeronx]

What's the difference between a Kaveri-K edition and non-K edition?

K is usually 95 watts.
Non-K is usually 65 watts with a locked CPU clock.

 

[Homeles]

Are you guys sure that 28SHP is not 28FD? http://i.imgur.com/u6fbE3d.png

It could be Bulk for the GPU, PHYs, etc and FD-SOI for the CPUs, SRAM, etc. Explaining the whole it is perfect for the APU.

SHP is SOI (custom(A custom SOI implementation)) moniker.

LPH is built for high-end SoCs but it is from Samsung.
HPP is built for best perf/watt but it

28 HPP should be ~1.8 Billion
28 LPH should be ~1.9 Billion
28 FDSOI should be ~2.2 Billion
^-- Transistors in 245 mm², VLIW4 is less dense than GCN so the number should be off a bit.

28 HPP (0.5 Node from 32nm) -> 28 LPH (0.55 Node from 32nm) -> 28 FD (0.5 Node above Bulk 20nm)

It could be, but I'd be severely disappointed if it was. AMD should be improving their clock speeds with FDSOI, not regressing.

 

[krumme]

Are you guys sure that 28SHP is not 28FD? http://i.imgur.com/u6fbE3d.png

It could be Bulk for the GPU, PHYs, etc and FD-SOI for the CPUs, SRAM, etc. Explaining the whole it is perfect for the APU.

SHP is SOI (custom(A custom SOI implementation)) moniker.

LPH is built for high-end SoCs but it is from Samsung.
HPP is built for best perf/watt but it

28 HPP should be ~1.8 Billion
28 LPH should be ~1.9 Billion
28 FDSOI should be ~2.2 Billion
^-- Transistors in 245 mm², VLIW4 is less dense than GCN so the number should be off a bit.

28 HPP (0.5 Node from 32nm) -> 28 LPH (0.55 Node from 32nm) -> 28 FD (0.5 Node above Bulk 20nm)

What are you meaning? Bulk and soi on the same die???

 

[NostaSeronx]

What are you meaning? Bulk and soi on the same die???

^this. Parts that work better with bulk get bulk and parts that work better with SOI gets SOI. Win-win.

FDSOI can be called a successor to Bulk, while PDSOI can not. FDSOI is synonymous to bulk and is actual more bulk than bulk. The reasoning for it is the Total Cost of Production from Foundry to Customer is significantly lower than the bulk pipeline.

The most relevant problem why if Kaveri uses FDSOI; is Gate First. Gate First has a negative impact when FDSOI is less than 32nm.

ST & GlobalFoundries 28FD uses Gate First and has density slightly lower than 22nm if GF had such a node.

3.7 GHz & 720 MHz with 245 mm² die area and 2.41 billion transistors is pretty good actually. When versus;
4.1 GHz & 844 MHz with 246 mm² die area and 1.303 billion transistors.

~1.85x the density from Trinity, while maintaining almost near the same clocks.

https://www.youtube.com/watch?v=QqvE24X7ijc&t=3m25s
Sounds like SHP is AMD Exclusive so it can be anything.

 

[ShintaiDK]

$173, thats ~22% more expensive than the fastest Richland and definitely on par with entry-level quad-core Haswells (3.0-3.2GHz).

22% higher price looks like a desperate attempt. Because the chip certainly doesnt deliver that value. I wonder how much the extra need to borrow 500m$ and the delayed 200m$ WSA payment have affected this. The big chip also demenstrates that AMD still got plenty of WSA wafer to throw away on close to nothing.

Kaveri is now above i3 in price, yet CPU wise it cant really compete. And as you said, its now in the i5 4440 area.

Or you can sell the key and practically pay 55 bucks less for 7850?

For a failed game that doesnt really sell anymore? The previous CPUs didnt cost extra when bundled with a game.

 

[parvadomus]

Kaveri is now above i3 in price, yet CPU wise it cant really compete. And as you said, its now in the i5 4440 area.

It looks like Kaveri might be a step above i3 CPU wise. And demolishes both i3 and i5 in graphics/gpu compute.

 

[littlebluesmurf]

22% higher price looks like a desperate attempt. Because the chip certainly doesnt deliver that value.

A10 Richland's are also bad value as it is repeatedly claimed on this forum. The shipment figures suggest otherwise, all this despite being into a declining market and being "poor value".

The big chip also demenstrates that AMD still got plenty of WSA wafer to throw away on close to nothing.

It is smaller than Trinity.

 

[ShintaiDK]

It is smaller than Trinity.

Its the same size really. 245 vs 246.

A10 Richland's are also bad value as it is repeatedly claimed on this forum. The shipment figures suggest otherwise, all this despite being into a declining market and being "poor value".

Could you link those shipment figures?

 

[littlebluesmurf]

Could you link those shipment figures?

 

[ShintaiDK]

What about all the rest? Thats A10 models only. And they discontinue other lines. Not to mention the A10 line got expanded model wise. 4 out of 13 was A10 with Richland. With Trinity it was 3 out of 15 on the FM2 socket.

 

[NostaSeronx]

Its the same size really. 245 vs 246.

No, it isn't it is smaller by 1! mm²!!!!
😀

 

[NTMBK]

What about all the rest? Thats A10 models only. And they discontinue other lines. Not to mention the A10 line got expanded model wise.

He was talking about A10s specifically.

 

[littlebluesmurf]

What about all the rest? Thats A10 models only. And they discontinue other lines. Not to mention the A10 line got expanded model wise. 4 out of 13 was A10 with Richland. With Trinity it was 3 out of 15 on the FM2 socket.

I only said A10's and that is what I have provided. You would have to ask AMD about the rest or hope they also made slides about them and someone posts it.

What lines were discontinued which were not replaced by a successor?

I count only 10 Trinity desktop APU's including the B models and the A4 4000, but excluding Athlons and FirePros's.

I also only count 10 Richland desktop APU's (no B model Richland's were released), excluding the Athlons. There is only 1 extra Richland A10 part compared to Trinity.

 

[inf64]

Density is crazy high though. 245mm^2 and 2.4B transistor is quite a feat.
Since integer execution is greatly improved let's hope the FP side of things gets similar boost with Carrizo. If it does they don't need more than 4 cores configuration on desktop. 6 would be nice (3 module) but it's not really necessary. The biggest Kaveri's problem is lack of memory BW for the big GPU on-die, it's a shame stacked memory is not coming sooner.

 

[ShintaiDK]

I only said A10's and that is what I have provided. You would have to ask AMD about the rest or hope they also made slides about them and someone posts it.

What lines were discontinued which were not replaced by a successor?

I count only 10 Trinity desktop APU's including the B models and the A4 4000, but excluding Athlons and FirePros's.

I also only count 10 Richland desktop APU's (no B model Richland's were released), excluding the Athlons. There is only 1 extra Richland A10 part compared to Trinity.

You need to look at all models for the product mix, also those without IGP. Else its simply a natural replacement. From Trinity to Richland there was a 4th A10 as well.

 

[ShintaiDK]

Density is crazy high though. 245mm^2 and 2.4B transistor is quite a feat.
Since integer execution is greatly improved let's hope the FP side of things gets similar boost with Carrizo. If it does they don't need more than 4 cores configuration on desktop. 6 would be nice (3 module) but it's not really necessary. The biggest Kaveri's problem is lack of memory BW for the big GPU on-die, it's a shame stacked memory is not coming sooner.

Roadmaps _still_ show no 3M/6T. Kaveris biggest problem is _still_ its poor CPU performance.

The density is not really that high. A 2.8B transistor Pitcairn is 212mm2. Its just natural when all you add is more GPU.

 

[Abwx]

22% higher price looks like a desperate attempt. Because the chip certainly doesnt deliver that value. I wonder how much the extra need to borrow 500m$ and the delayed 200m$ WSA payment have affected this. The big chip also demenstrates that AMD still got plenty of WSA wafer to throw away on close to nothing.

Kaveri is now above i3 in price, yet CPU wise it cant really compete. And as you said, its now in the i5 4440 area.

AMD slides claim 87% better perfs than HD4600 , does
thoses i3/i5 beat Kaveri by a sizeable margin CPU wise
to compensate for the poor GPU side and thus justify the
higher pricing..?..

Considering the whole equation it is obvious that Kaveri
is by far the best value even at thoses prices , AMD are
not stupid , they know better than you or me what are
the intrinsical perfs/prices ratios.

The density is not really that high. A 2.8B transistor Pitcairn is 212mm2. Its just natural when all you add is more GPU.

Granted GPU are density friendly but then Haswell has a big
GPU and Kaveri still has a 25% higher density.

 

[littlebluesmurf]

You need to look at all models for the product mix, also those without IGP. Else its simply a natural replacement. From Trinity to Richland there was a 4th A10 as well.

The Athlon's are salvaged APU's, they may use the same socket but it would be the equivalent of me bundling in the FX series into the product mix, as they are CPU only parts.

The FirePro's I am also not including as they are a different brand. They do have a spec that matches the A10 Trinity parts though :whiste:.

 

[ShintaiDK]

The Athlon's are salvaged APU's, they may use the same socket but it would be the equivalent of me bundling in the FX series into the product mix, as they are CPU only parts.

The FirePro's I am also not including as they are a different brand. They do have a spec that matches the A10 Trinity parts though :whiste:.

The chart shows desktop shipments. Not APUs specificly. So you need to look at the entire desktop product mix. I was simply being generous only using the FM2 socket to compare.

 

[inf64]

Roadmaps _still_ show no 3M/6T. Kaveris biggest problem is _still_ its poor CPU performance.

The density is not really that high. A 2.8B transistor Pitcairn is 212mm2. Its just natural when all you add is more GPU.

I said 6 would be nice. But with native 256bit pipelines in Carizo it will just blast SR core in AVX/FMA stuff so workloads that were running better on intel before will see major shrinkage of the gap with Excavator. Intel will probably still have a lead but 4C Carizo will be more rounded product with lower TDP and way better perf/watt. If it was to clock good then it would do solid in 2015.

 

[Fjodor2001]

Density is crazy high though. 245mm^2 and 2.4B transistor is quite a feat.

That's even higher density than Intel's 22 nm process! :awe:

Intel 4770K is 1.4B transistors on 177mm^2. So it's:
Intel 22 nm: 1.4B/177mm^2=7.91M/mm^2
GloFo/AMD 28 nm: 2.4B/245mm^2=9.80M/mm^2

And yet it's 22 nm vs 28 nm. Do those nm numbers actually mean anything anymore? 😕

 

[AtenRa]

How much i see no comment on this, but there should certainly be some good comments "because there is question of tiny 600mhz higher defoult CPU frequency+Higher IPC under 45W TDP".😎

- Richland APU A10-6700T, 2.5ghz/3.5ghz Turbo Core/384 VLIW4 Radeon Cores, 45W TDP

http://www.cpu-world.com/CPUs/Bulldozer/AMD-A10-Series A10-6700T.html


- Kaveri APU A8-7600, 3.1ghz/3.3ghz Turbo Core/384 GCN Radeon Cores, 45W TDP

A8-7600 at 45W is the best APU for SFF gaming/HTPC etc. At $120 it will be faster than Core i3 Graphics with a lower TDP making it available to very thin SFF cases.
That also brings us to Mobile Kaveri being way faster than Richland. Its possible a 35W TDP Mobile Kaveri to have 65W Richland Desktop graphics performance(both 1866MHz memory).

 

[inf64]

I wonder why AMD just doesn't use that 45W model as a base for mobile chip? It would sell for sure and command higher price than most of their mobile lineup. Pair it with some 7750 ddr3 version for HCF and you get a nice laptop for games.

 

[ShintaiDK]

That even higher density than Intel's 22 nm process! :awe:

Intel 4770K is 1.4B transistors on 177mm^2. So it's:
Intel 22 nm: 1.4B/177mm^2=7.91M/mm^2
GloFo/AMD 28 nm: 2.4B/245mm^2=9.80M/mm^2

And yet it's 22 nm vs 28 nm. Do those nm numbers actually mean anything anymore? 😕

You compare apples and oranges. GPU part is more dense by nature. Much more. As example before, Pitcairn 212mm2 2.8B. Or 13.2M/mm2 by your math.

Remember the APU on 32nm GloFo? 1.3B/246mm2. Or 5.2M/mm2.
Or BD on 32nm GloFo. 1.2B/315mm2. Or 3.81M/mm2. Same process as the APU above.

Intels 32nm with SB. 1.16B/216mm2. Or 5.3M/mm2. Surprisingly better than the APUs. But again, cache is also dense.