• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

AMD Details Jaguar; Preps Chip for PlayStation 4

The standard configuration has 4 cores, but Sony ordered a custom version with 8 cores.
And this is why I was confusing it with a Zambezi. When hearing it was an 8 core, by immediate instinct was to compare it to and assume it was a slightly different model of their current 8 cores out.

Nowhere near the same. It is a much lower end model than that.
 
They're two completely different microarchitectures. Jaguar is designed for efficiency and low power consumption (a la Atom, but Jaguar has a wider range of TDPs). The Jaguar SoCs available to consumers max out at 25W - think about that.

If the recent benchmarks are legit then it's not hard to see why AMD went with Jaguar rather than Bulldozer. I mean:

AMD Temash A6-1540 (Jaguar, 5W TDP)
Cinebench 11.5 Multi: 1.39
Cinebench 11.5 Single: 0.35
3DMark 11: 537 points

AMD Trinity A10-4600M (Bulldozer/Piledriver, 35W TDP)
Cinebench 11.5 Multi: 2.00
Cinebench 11.5 Single: 0.70
3DMark 11: 1057 points

Only half the performance while using 6-7x less power. It's ridiculous, and I don't know whether that says more about how bad Bulldozer is or how good Jaguar is.

So since PS4 has 8 cores, instead of 4, couldn't you just theoretically double all those numbers? Essentially you be getting the same performance or a little bit more, at way less of a power draw. I realize that having 8 cores, means its a lot more work for programmers(bad for Carmack lol) to get maximum utilization.
 

McHuj

Member
Link? IIRC they can only do 64-bit SIMD's.


SIMD2.png


http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/3

Here's a SIMD unit in GCN. A CU is composed of 4 of these. All SIMD really means is that a single instruction executes on multiple ALU's (of some size). For Cell, SSE, GPU, the ALU's are usually 32-bit.
 

KidBeta

Junior Member
Link? IIRC they can only do 64-bit SIMD's.

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/4

512bit wide SIMD, 16 wide.


SIMD2.png


http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/3

Here's a SIMD unit in GCN. A CU is composed of 4 of these. All SIMD really means is that a single instruction executes on multiple ALU's (of some size). For Cell, SSE, GPU, the ALU's are usually 32-bit.

MMX is usually implemented as 64bits (2x 32bit, or 1x 64 bit).
SSE is usually implemented as 128bits (4x 32 bit, or 2x 64 bit).
AVX is atm implemented as either 256bits (8x 32bit, or 4x 64bit) or 128bits (4x 32 bit, or 2x 64 bit).

The size of the elements in a SIMD unit is generally flexible to some degree and not fixed at 32 or 64bits.

All of the above are common place on CPU's
 

TheD

The Detective
They're two completely different microarchitectures. Jaguar is designed for efficiency and low power consumption (a la Atom, but Jaguar has a wider range of TDPs). The Jaguar SoCs available to consumers max out at 25W - think about that.

If the recent benchmarks are legit then it's not hard to see why AMD went with Jaguar rather than Bulldozer. I mean:

AMD Temash A6-1540 (Jaguar, 5W TDP)
Cinebench 11.5 Multi: 1.39
Cinebench 11.5 Single: 0.35
3DMark 11: 537 points

AMD Trinity A10-4600M (Bulldozer/Piledriver, 35W TDP)
Cinebench 11.5 Multi: 2.00
Cinebench 11.5 Single: 0.70
3DMark 11: 1057 points

Only half the performance while using 6-7x less power. It's ridiculous, and I don't know whether that says more about how bad Bulldozer is or how good Jaguar is.

Power draw does not scale linearly with clock speed and you have to take into account the much more powerful GPU in the A10.
 

strata8

Member
Power draw does not scale linearly with clock speed and you have to take into account the much more powerful GPU in the A10.

The GPU in the A10 is only around 4x more powerful, so even when that's taken into account Temash is still more efficient per watt.
 

TheD

The Detective
The GPU in the A10 is only around 4x more powerful, so even when that's taken into account Temash is still more efficient per watt.

Ha! "only 4x more".........

And as I said, power draw does not scale linearly, the single threaded performance of the Piledriver is 2x that of that jaguar core, that does not come for free!
 
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/3

Here's a SIMD unit in GCN. A CU is composed of 4 of these. All SIMD really means is that a single instruction executes on multiple ALU's (of some size). For Cell, SSE, GPU, the ALU's are usually 32-bit.

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/4

512bit wide SIMD, 16 wide.




MMX is usually implemented as 64bits (2x 32bit, or 1x 64 bit).
SSE is usually implemented as 128bits (4x 32 bit, or 2x 64 bit).
AVX is atm implemented as either 256bits (8x 32bit, or 4x 64bit) or 128bits (4x 32 bit, or 2x 64 bit).

The size of the elements in a SIMD unit is generally flexible to some degree and not fixed at 32 or 64bits.

All of the above are common place on CPU's

Ahh, ok. Thanks for the heads up. Odd they can't even simulate (not emulate) the Cell.
 

tipoo

Banned
is this worth discussing?

PlayStation 4 CPU runs at 2 GHz: rumor

http://ps4daily.com/2013/02/playstation-4-cpu-runs-at-2-ghz-rumor/

I feel like they're going to wait for Microsoft to spill the beans before saying so. Maybe like the 8GB RAM it will be a last minute change, since it's only making things faster it won't affect what developers already wrote for it. A 1.25x increase would be a nice surprise, and would help quite a bit with Jaguars limited single thread performance.
 

onQ123

Member
From that Mike R guy on Byond3D

http://forum.beyond3d.com/showpost.php?p=1714745&postcount=815

Jaguar Vanilla
* 1.8GHz LC Clocks (can be under-clocked for specific low-powered battery device needs - tablets, etc...).
* 2MB shared L2 cache per CUs
* 1-4 CUs can be outfitted per chip. (i.e. 4-16 logical cores)
* 5-25 watts depending on the device/product. (45 watts is achievable under proper conditions)

PS4 Jaguar with chocolate syrup.
* 2GHz is correct as of now.
* 4MB of total L2 cache (2MB L2 x 2 CUs)
* 2 CUs (8 Logical cores).
* idles around 7 watts during non-gaming operations and around 12 watts during Blu-ray movie operations. Gaming is a mixed bag...

What would be nice is a fully loaded Jaguar chip. To handle Kinect's 2 CPU intensive operations - keep LC count at 8 for developer needs. I did not say that...

I wonder if he means that the Xbox Next will have 16 Cores?
 

golem

Member
There was the rumor about Xbox Next having twice the performance of PS4's CPU

In that thread he says there are multiple configurations of the logic board out there, some with more cores than others. Also it seems that he is saying Durango's ram is actually GDDR3.

I guess we'll see.
 

flem1

Member
In that thread he says there are multiple configurations of the logic board out there, some with more cores than others. Also it seems that he is saying Durango's ram is actually GDDR3.

I guess we'll see.

GDDR3? I thought it was DDR3 for Durango???
 
Top Bottom