[antwan]

Originally Posted by Grall

Whatever. It's not a difference that changes anything on a fundamental level.

But it does

http://en.wikipedia.org/wiki/GDDR3

[bgassassin]

Originally Posted by SKYSONY

bgassassin, whose post are always interesting to read, keeps saying the Durango´s GPU s will be less powerful than PS4 one. I thought he has mostly the same sources than lherre, but, apparently his info comes from a difference person.

It appears to me that the 720 dev kits (probably where all indiser info comes from) are constantly evolving, since back in November IGN said the GPU was a 6670, that is only 0.7 TFLOPS. If when bg got the > 1 TF info the dev kits were not the final ones, we might see a better number with the next leak.

Actually I've never said it will be weaker than PS4's. What I have said is that the info at first didn't sound impressive compared to how I first heard PS4's, that I was told 1+ TFLOPs, that it doesn't lead me to believe it's up there with PS4 as I feel the description doesn't suggest it will, and that I'm waiting on more info before making a conclusion.

[SKYSONY]

Originally Posted by bgassassin

Actually I've never said it will be weaker than PS4's. What I have said is that the info at first didn't sound impressive compared to how I first heard PS4's, that I was told 1+ TFLOPs, that it doesn't lead me to believe it's up there with PS4 as I feel the description doesn't suggest it will, and that I'm waiting on more info before making a conclusion.

I don´t expect the info you got is nothing final, but if the PS4 early dev kits have a 1.8 TFLOP GPU, and a more advanced Durango devs have a 1+ one, it might indicate something.

[upnorthsox]

Originally Posted by Grall

What on earth are you talking about? The 21ish GB/s that the PS3 XDR bus manages is extremely attainable with DDR3, without resorting to esoteric clock speeds.

Originally Posted by pjbliverpool

You'll get 25.6 GB/s with 1600Mhz DDR3 on a 128bit bus. But you can get higher speed DDR3 than that.

These benchmarks show high teens (18.45 GBs for reference 1600) to low 20's for bandwidth, and it's not commodity DDR3 but the boutique type memory I was talking about:

http://www.tomshardware.com/reviews/...king,3209.html

It's not that xdr is still some awesome memory, but it was top of the class back in '06 and by far the best option in '03 when the Cell was being specc'd out.

[bgassassin]

Originally Posted by SKYSONY

I don´t expect the info you got is nothing final, but if the PS4 early dev kits have a 1.8 TFLOP GPU, and a more advanced Durango devs have a 1+ one, it might indicate something.

Haha. It wasn't even detailed enough to draw a personal conclusion let alone say it's final.

[Grall]

Originally Posted by antwan

But it does

Not really. GPUs of the (G)DDR3 era were commonly designed to handle both types of RAM using the same memory controller, for cost reasons. So obviously these types of RAM do not differ in any fundamental way. Also, more importantly, peak performance does not differ between DDR3 and GDDR3. Can we end this irrelevant quibbling now, kkthx.

Originally Posted by upnorthsox

These benchmarks show high teens (18.45 GBs for reference 1600) to low 20's for bandwidth

What's your point? A CPU memory benchmark with CPU-like access patterns and CPU architecture-dependent internal quirks and bottlenecks really isn't relevant to this discussion.

and it's not commodity DDR3 but the boutique type memory I was talking about:

Yes, but as already stated you don't need boutique DDR3 to reach PS3 XDR-level memory performance. So again, what's your point?

It's not that xdr is still some awesome memory, but it was top of the class back in '06 and by far the best option in '03 when the Cell was being specc'd out.

Well, it's not 2003, or even 2006 anymore.

[novcze]

Ok. Rambus paper says memory bandwidth of over 500GB/s on a single SoC, if that's direction Sony is going (APU, SoC), is there any alternative? There is also maybe correlation with Sony and AMD deal and AMD using XDR2 (http://www.tomshardware.com/news/amd...ory,13408.html)

[antwan]

Originally Posted by novcze

Ok. Rambus paper says memory bandwidth of over 500GB/s on a single SoC, if that's direction Sony is going (APU, SoC), is there any alternative? There is also maybe correlation with Sony and AMD deal and AMD using XDR2 (http://www.tomshardware.com/news/amd...ory,13408.html)

500GB/s ?
Like XDR before it, I don't think there is an alternative..
I would gladly be proven wrong though..

[aaronspink]

Originally Posted by upnorthsox

Actually DDR3 can't, and it only comes close with highly cherry-picked, extreme overclocked dimms with a pricetag that makes xdr seem like a bulk commodity item. And PS3 is only running it at half clock so the comp is even more one sided.

Commodity DDR3 is currently as fast as XDR and has been for quite some time. The current top end DDR3 products are significantly faster and should have pricing at or below xdr levels.

XDR2 should have been a stacked solution from the get go, especially in an LP variant. That was Rambus's big mistake, they misjudged where the market is going. They could've had Apple.

No the couldn't. DDR3 was by far a more popular standard with full support both from the ASIC designers and more importantly the memory manufacturers.

[aaronspink]

Originally Posted by upnorthsox

These benchmarks show high teens (18.45 GBs for reference 1600) to low 20's for bandwidth, and it's not commodity DDR3 but the boutique type memory I was talking about:

DDR3 1600 is not boutique. And you should probably understand the difference between peak and sustained performance before you continue in this discussion. And for reference, DDR3-2400 and DDR3-2667 are boutique, 1600 is bog standard commodity.

[upnorthsox]

Originally Posted by aaronspink

Commodity DDR3 is currently as fast as XDR and has been for quite some time. The current top end DDR3 products are significantly faster and should have pricing at or below xdr levels.

XDR can reach 115.2 GB of bandwidth on a 128 bit bus, what DDR3 or even DDR4 can match that?

No the couldn't. DDR3 was by far a more popular standard with full support both from the ASIC designers and more importantly the memory manufacturers.

I don't know how this is at all related to what you quoted.

[aaronspink]

Originally Posted by antwan

500GB/s ?
Like XDR before it, I don't think there is an alternative..
I would gladly be proven wrong though..

There is an alternative, it is called reality. I can spec out an asic with several hundred GB/s of DDR3 as well. Both are equally impractical.

[upnorthsox]

Originally Posted by aaronspink

DDR3 1600 is not boutique. And you should probably understand the difference between peak and sustained performance before you continue in this discussion. And for reference, DDR3-2400 and DDR3-2667 are boutique, 1600 is bog standard commodity.

I'm going to assume you didn't go to the link.

[aaronspink]

Originally Posted by upnorthsox

XDR can reach 115.2 GB of bandwidth on a 128 bit bus, what DDR3 or even DDR4 can match that?

No it cannot. Top speed XDR can only reach 64 GB/s on a 128b bus. DDR3 with the same pin budget can reach 64+ GB/s.

If you want to talk about mythical vapor, I can talk about mythical vapor, but I have better mythical vapor.

[aaronspink]

Originally Posted by upnorthsox

I'm going to assume you didn't go to the link.

I'm going to confirm that you don't understand the link nor anything about the memory market. Suffice to say that DDR3-1600 reaches 25.6 GB/s and is bog standard commodity memory. Has been for quite some time.

I suggest you do some research on the differences between peak and sustained bandwidth and how sustained bandwidth changes dependent on workload before posting about it again.

[Shifty Geezer]

aaronspink - what do you make of AMD's move to XDR2?

[aaronspink]

Originally Posted by Shifty Geezer

aaronspink - what do you make of AMD's move to XDR2?

I might have an opinion if people could read dates and product spec sheets and put two and two together.

Here's a hint: 7970 is shipping today, it doesn't use XDR2.

[antwan]

http://en.wikipedia.org/wiki/DDR3_SDRAM
look at the 1600 latencies. Plus I don't see >25GB/sec anywhere but yeah, guess that's wikipedia, right?
XDR ram has latencies about 10 times lower than the DDR3 SDRAM.

But that don't count right? I mean, the clock speed is much higher, and if DDR3 had the same clock speed it would be much faster right?..

The only thing vapor here is you thinking the DDR3 is a match for the XDR.

[aaronspink]

Originally Posted by antwan

http://en.wikipedia.org/wiki/DDR3_SDRAM
look at the 1600 latencies. Plus I don't see >25GB/sec anywhere but yeah, guess that's wikipedia, right?

1.6 Gbs * 128 /8 = 25.6 GB/s. Math, its hard.

XDR ram has latencies about 10 times lower than the DDR3 SDRAM.

I do not think that word means what you think it means. XDR latencies are not lower than DDR3.

CAS latency for DDR3 are between 10-15nS. 1/10th that would be 1-1.5 nS. The command bus for XDR runs at 800 Mhz or 1.25 nS cycle time. It takes 2 cycles to transfer a command or 2.5ns. So just the transfer time for the command proves that your assertion is wrong. Let along that XDR has basically the same CAS/RAS/Pre timings as DDR3. In almost all cases a DDR3 based system is going to have lower DRAM latencies than XDR.

Once again, if you understood the technology, you wouldn't ever make such incorrect statements.

Lets be honest here, if someone can come up with a memory technology within 2x the cost of JEDEC spec DDR with 1/10th the latency, it would sweep the entire memory market within months. You would be able to literally print 10s of billions of dollars. AMD/Intel/ARM/Apple/HP/Oracle/Cray/etc would be lining up to buy it in massive quantities and incorporating it into every design they had. We're talking about cutting load-to-use latencies on the order of 20-60ns for latency optimized memory systems that currently sit in the range of 40-100ns (varies based on bank hit/bank miss/bank conflict).

But that don't count right? I mean, the clock speed is much higher, and if DDR3 had the same clock speed it would be much faster right?..

I assume you mean data rate is much higher. clockrate != data rate, just fyi. DDR3 ships with a max bin data rate of 2.133 Gb/s in standard JEDEC DIMM specs (significantly higher in direct attach and boutique dimm specs). XDR ships with a max data rate of 4 Gb/s. But XDR uses differential signaling, so in fact, DDR3 actually has a higher data rate per pin (2.133 Gb/s for DDR3 vs 2 Gb/s for XDR).

The only thing vapor here is you thinking the DDR3 is a match for the XDR.

DDR3 is more than a match for XDR, it has literally killed XDR in the market. The score is somewhere in the range of 1,000,000,000:1 in favor of DDR3.

[Shifty Geezer]

Originally Posted by aaronspink

I might have an opinion if people could read dates and product spec sheets and put two and two together.

Here's a hint: 7970 is shipping today, it doesn't use XDR2.

Oh, they were only rumours. Obvious when one clicks links, but we haven't all got time to read every linked article.

[aaronspink]

Originally Posted by Shifty Geezer

Oh, they were only rumours. Obvious when one clicks links, but we haven't all got time to read every linked article.

Not just rumors, old rumors about products that are currently shipping and do not match the rumors

[BRiT]

The past few pages have been sadly hilarious indeed.

[aaronspink]

Originally Posted by Shifty Geezer

Is there any benefit to XDR in stacked form? I don't know how the bus format affects stacked access.

No, only downsides. With either a flip stack or a TSV stack you aren't limited in any way by number of connection for any reasonable number of connections. Therefore what you want to concentrate on is power and area. In both cases high speed differential is a detriment. Most of the stacked memory designs are looking at 512-1024b buses running at relatively slow frequencies in the range of .25-1 Gb/s with the outliers in the <= 2 Gb/s range, all singed ended.

[aaronspink]

Originally Posted by novcze

If they are going with stacked memory, isn't completely irrelevant what is mass-produced and what not?

No, even for stacked memory, you would like to be using a commodity part. Ideally you would want to have a standard interface (both in pins, signaling technology, and pin/ball geometry with equal die size in the best of all possible world, but you'll settle for different die sizes) such that the product is available from multiple vendors and there is price competition.

[antwan]

Originally Posted by aaronspink

1.6 Gbs * 128 /8 = 25.6 GB/s. Math, its hard.

How did you come up with the 1.6 Gbs number? It should be 800Mbs, resulting in 12.8GB/s.
FYI, DDR3-1600 is also known as PC3-12800.

As for the 10 time lower latency; I know it's not actually 10 times lower, I was aiming to have you produce a calculation resulting in a smart
"look it's not 10 times lower; it's only 3 times lower!!"
-kind of post. I wanted you to realize the difference for yourself.

Cell would have been severely bottlenecked using DDR3 memory. Not to mention it did not exist for several years