anand's jaguar article

[crashtech]

For the next few months though, AMD will enjoy a position it hasn’t had in years: a CPU performance advantage.

That's nice to hear. I haven't exactly been bullish on AMD lately (my home systems are all Intel as of now) but it would not be optimal for Intel to be left with no competition at all. Hats off to AMD for their successful low power design.

 

[galego]

Then what about the total gflop number of the 3770k with CPU & GPU?

Correct me if wrong but the i7 GPU is for rendering not compute. I am only considering compute.

If you want add the xxx GFLOPs of the i7 GPU to the i7 CPU then add also the 1.4 TFLOPs of the PS4 GPU to the PS4 CPU.

 

[Cogman]

Its a shame the PS4 is so weak then. The PS3 was marketed as 1Tflop.

http://en.wikipedia.org/wiki/PlayStation_3_hardware

Wiki cites it at roughly 230 gflops. (And that was only for single precision floating point numbers)

The thing is, cell processors are very unique vs more traditional processors. That level of processing power would almost certainly never be used, especially in a gaming situation.

AMDs more traditional CPU cores will be much easier to program against in the long run, which means we will see better games.

The fact that AMDs processor is powering both platforms (very interesting to me) means that the difference in game offerings on the XBOX vs the PS4 will be very minimal. There will be far less tuning and monkey work to get things up and running (I mean, sure, there will be some issues because the OS will be significantly different, however, those issues will be pretty easy to solve for most well written games).

 

[Cogman]

Correct me if wrong but the i7 GPU is for rendering not compute. I am only considering compute.

If you want add the xxx GFLOPs of the i7 GPU to the i7 CPU then add also the 1.4 TFLOPs of the PS4 GPU to the PS4 CPU.

Pretty sure it is for both. After all, Intel has been pretty heavily involved in the OpenCL initiative.

When you have a GPU on die like this, communication between the CPU and GPU (a typical pain point for GPGPU processing) becomes much less of an issue. AMD has a better setup for it (hUMA will make a big difference) but in the end, they will both see similar benefits.

 

[Exophase]

We don't have any benchmark or other info now. I assume that you did a typo and you mean GFLOPs instead of FLOPs.

Yes I meant GFLOPS (but not GFLOPs, that's the plural of GFLOP). I'm not sure you understand that the number has nothing to do with benchmarks, they're design features of the uarch and something we know entirely.

The 102.4 GFLOPs are from the jaguar cores alone. There are 4 extra CUs for compute that provide 409.6 GFLOP if needed. Therefore we have a theoretical maximum of

PS4: 512 GFLOPs
i7-3770K: 224 GFLOPs

In case I did not mistake now

There's no "four extra CUs." There's a set of CUs 18 on PS4. You can use them for graphics and you can use them for compute.

Which is the same as a discrete GPU. Bringing in the GPU for a PS4 comparison is not fair if you don't allow a GPU for the i7 comparison. But I thought we were only talking about CPUs here.

You sound like you're just distorting the argument to something totally weird because you don't like admitting you're wrong :/

 

[Enigmoid]

Also the fact that if you wanted to (intel updated their drivers to allow it) you could run calculations on the hd 4000 in the 3770k at the same time.

(IMO I think this would be awesome for something like physics calculations in games).

 

[beginner99]

with something like SC2, you mean something which is written with very poor threading/load balancing in mind i assume?

Well not everything can be parallelized...and a RTS Game is exactly something very problematic.

Also the fact that if you wanted to (intel updated their drivers to allow it) you could run calculations on the hd 4000 in the 3770k at the same time.

(IMO I think this would be awesome for something like physics calculations in games).

Well IMHO the only point of an iGPU in the long run are such calculations. just another co-processor that got integrated. Right now it is just not very useful.

Now looking at those test and considering that silvermont (baytail) will have 1/4 of EUs of HD 4000 we can conclude it will be worse than adreno 320. On the other hand I think I take that + longer battery life any day...

 

[sangyup81]

I was a little underwhelmed by the performance of the 4 core kabini not being as much higher than the e-350 as I expected. It's a slight IPC improvement with double the cores.

Anyone know the retail price of the PS4 and new Xbox? It would be better to compare GFLOPs with a similar priced PC that has an Intel CPU and a NVIDIA GPU than talk about a 3770k which sells for $330 on Newegg

 

[strata8]

I was a little underwhelmed by the performance of the 4 core kabini not being as much higher than the e-350 as I expected. It's a slight IPC improvement with double the cores.

• 20% higher single-thread (+30% IPC)
• 150% higher multi-thread
• 65% higher GPU performance
• 30% better battery life

Doesn't seem that underwhelming to me.

Anyone know the retail price of the PS4 and new Xbox? It would be better to compare GFLOPs with a similar priced PC that has an Intel CPU and a NVIDIA GPU than talk about a 377k which sells for $330 on Newegg

Don't forget a total system TDP of 100W

 

[ShintaiDK]

The theoretical gflops talk is funny. We had it with Xbox360 vs PS3. 2Tflop vs 1Tflop. And now its back with more papernumber flops. We had it back with nVidia vs AMD. And slower gflop nVidia cards performed better in gaming.

Not to mention, 1 flop is not equal to 1 flop (As also seen over and over again.). Its the absolute worst value of measurement you can use. One system might need 4 flops to complete what another systemc an do in 1 flop.

 

[R0H1T]

The theoretical gflops talk is funny. We had it with Xbox360 vs PS3. 2Tflop vs 1Tflop. And now its back with more papernumber flops. We had it back with nVidia vs AMD. And slower gflop nVidia cards performed better in gaming.

Not to mention, 1 flop is not equal to 1 flop (As also seen over and over again.). Its the absolute worst value of measurement you can use. One system might need 4 flops to complete what another systemc an do in 1 flop.

In other words ~ the PlayStation 4 is better than a PC :awe:

 

[itsmydamnation]

The theoretical gflops talk is funny. We had it with Xbox360 vs PS3. 2Tflop vs 1Tflop. And now its back with more papernumber flops. We had it back with nVidia vs AMD. And slower gflop nVidia cards performed better in gaming.

Not to mention, 1 flop is not equal to 1 flop (As also seen over and over again.). Its the absolute worst value of measurement you can use. One system might need 4 flops to complete what another systemc an do in 1 flop.

WTF are you talking about, 1 flop is exactly what is says it is 1 flop. Only a moron would use just one metric to profile the performance of an architecture and only a moron would attack a perfectly legitimate metric.

But the reality here is the Flops counts on the PS4 is hardly paper, GCN gets good utilisation has good memory latency hiding capabilities and low execution latency. Jaguar itself is a capable core with solid execution resources and low latency, it will be interesting to see what level of "IPC" on AVX/SSE2 ops game devs will get once they start coding specificity for the Jaguar Uarch.

Let's please not call other posters morons in the technical forums.

-Thanks
ViRGE

 

[maddie]

The theoretical gflops talk is funny. We had it with Xbox360 vs PS3. 2Tflop vs 1Tflop. And now its back with more papernumber flops. We had it back with nVidia vs AMD. And slower gflop nVidia cards performed better in gaming.

Not to mention, 1 flop is not equal to 1 flop (As also seen over and over again.). Its the absolute worst value of measurement you can use. One system might need 4 flops to complete what another systemc an do in 1 flop.

You so often, write as an expert but make very little sense. A flop is precisely defined.

 

[ShintaiDK]

You so often, write as an expert but make very little sense. A flop is precisely defined.

WTF are you talking about, 1 flop is exactly what is says it is 1 flop. Only a moron would use just one metric to profile the performance of an architecture and only a moron would attack a perfectly legitimate metric.

But the reality here is the Flops counts on the PS4 is hardly paper, GCN gets good utilisation has good memory latency hiding capabilities and low execution latency. Jaguar itself is a capable core with solid execution resources and low latency, it will be interesting to see what level of "IPC" on AVX/SSE2 ops game devs will get once they start coding specificity for the Jaguar Uarch.

Take folding. When a GPU needs to use 4 flops, a CPU only needs to use 1 for the same operation. Now go figure out why.

Let me give you two a hint. Typically, an addition or a multiply operation is always considered to be 1 flop, but more complex functions such as exponential, sine, or cosine, are less clear.

 

[maddie]

Take folding. When a GPU needs to use 4 flops, a CPU only needs to use 1 for the same operation. Now go figure out why.

Let me give you two a hint. Typically, an addition or a multiply operation is always considered to be 1 flop, but more complex functions such as exponential, sine, or cosine, are less clear.

In your example, the GPU needs 4 operations to do 1 flop, not 4 flops to equal 1 CPU flop. Please use proper terms.

 

[galego]

Pretty sure it is for both. After all, Intel has been pretty heavily involved in the OpenCL initiative.

We are talking about closely related but still different stuff. Of course you can use OpenCL for performing non-graphic computations on the GPU. You can do the same with the GPU on AMD APUs. But I was mentioning the 4 CUs for compute on the PS4 chip. Thanks to HSA, you don't need OpenCL to acces to them and those CUs can access to system memory directly (thanks to hUMA). I see those 4 CUs somewhat as the streamlined coprocessors in CELL, with the difference that those 4 CUs can switch to rendering mode if the developers need an extra boost on graphics.

Yes I meant GFLOPS (but not GFLOPs, that's the plural of GFLOP). I'm not sure you understand that the number has nothing to do with benchmarks, they're design features of the uarch and something we know entirely.

There's no "four extra CUs." There's a set of CUs 18 on PS4. You can use them for graphics and you can use them for compute.

Which is the same as a discrete GPU. Bringing in the GPU for a PS4 comparison is not fair if you don't allow a GPU for the i7 comparison. But I thought we were only talking about CPUs here.

You sound like you're just distorting the argument to something totally weird because you don't like admitting you're wrong :/

How many times I have to say that we are using GFLOPs because we have no other data now?

Those 18 CUs are split in a 14+4 balance. Albeit the 4 CUs can switch mode to rendering if needed, the main task of the 4 CU will be on computing non-graphics stuff: e.g. physics effects. In one sense those 4 CU are like the streamlined coprocessors of the Cell

About 14 + 4 balance:
- 4 additional CUs (410 Gflops) “extra” ALU as resource for compute
- Minor boost if used for rendering

http://www.vgleaks.com/world-exclusive-orbis-unveiled-2/

If you want to introduce the GPU of the i7 comparison then introduce the full GPU of the PS4 as well. In that case the whole performance of the PS4 is not the 512 GFLOPs that I wrote above but 512 + 1433 GFLOPs. This is a total of 1945 GFLOPs, which is usually quoted as "2 TFLOPs".

I was a little underwhelmed by the performance of the 4 core kabini not being as much higher than the e-350 as I expected. It's a slight IPC improvement with double the cores.

One single jaguar core in kabini gives slightly more performance per cycle than a SB core in an i3.

Unixbench single-core index
Jaguar (kabini) @ 2.0 GHz: 403.8
SB (i3-2100T) @ 2.5 GHz: 495.5

Where jaguar(kabini) shine is in multi-core performance.

Not to mention, 1 flop is not equal to 1 flop (As also seen over and over again.). Its the absolute worst value of measurement you can use. One system might need 4 flops to complete what another systemc an do in 1 flop.

Evidently a 1 flop is equal to 1 flop. What I think you are trying to say is that machine A with 1 flop can be faster than machine B with a 1 flop. We know that.

We know that a console with 2 TFLOP will be faster than a gaming PC with 2 TFLOP.

 

[ShintaiDK]

We know that a console with 2 TFLOP will be faster than a gaming PC with 2 TFLOP.

LOL, no.

 

[R0H1T]

LOL, no.

Well yes if all you're doing is comparing games, the major hindrance however is that performance across different platforms isn't comparable & is subject to an enormous number of variables like OS+hardware+TDP+coding language amongst other things, therefore GLOPS being a raw measurement is perhaps the only objective parameter comparable between the two !

 

[guskline]

Glad to see AMD getting a positive review with the Jaguar.

 

[Atreidin]

LOL, no.

+1

I think some people are using terms they don't fully understand, and pretending they get it because it makes their imaginary worlds in their heads more pleasing to them.

 

[grimpr]

Glad to see AMD getting a positive review with the Jaguar.

All you have to do is compare Jaguars vs Silvermonts reviews, counting total words and examining phrasing and tone to really determine Anands favorite CPU provider, he just doesnt care enough about AMD. :biggrin:

 

[Headfoot]

FLOPS has almost nothing to do with any relevant performance concern regarding games. Whatsoever. This talk of FLOPS is an absolute side track that is misdirecting from the real performance numbers.

If you care about FLOPS, stop. Just don't. It means so little as to not mean anything. Unless you can show me a perfectly parallelizable game that runs entirely off of floating point operations, stop using FLOPS.

Talking real, concrete terms like benchmark numbers and the number of CUs in a known graphics architecture is vastly more productive discussion.

 

[nehalem256]

FLOPS has almost nothing to do with any relevant performance concern regarding games. Whatsoever. This talk of FLOPS is an absolute side track that is misdirecting from the real performance numbers.

If you care about FLOPS, stop. Just don't. It means so little as to not mean anything. Unless you can show me a perfectly parallelizable game that runs entirely off of floating point operations, stop using FLOPS.

Talking real, concrete terms like benchmark numbers and the number of CUs in a known graphics architecture is vastly more productive discussion.

Isn't talk of FLOPS a lot more relevant than talk of CUs with regard to performance across architectures?

 

[ShintaiDK]

Isn't talk of FLOPS a lot more relevant than talk of CUs with regard to performance across architectures?

Its there it fails. The only real usefulness is to bench against the theoretical max of your system. And see your efficiency rate.

 

[Phynaz]

How many times I have to say that we are using GFLOPs because we have no other data now?

You are using GFLOPS because you think it supports your position. We aren't using it because we know the difference between theoretical numbers and real life performance.