The AMD Execution Thread [2007 - 2017]

[3dilettante]

I have not run across a specification listing for the in-production 8Gb/s GDDR5 bin, but at least for Samsung it has been BL8 all the way to 7 Gb/s.

I was idling musing about the presence of an HBM pool on that rumored HPC APU (not that linked slides have had a good recent track record), or a future-gen Opteron with the touted high I/O bandwidth. If AMD hasn't completely junked the Seamicro IP, I recall some presentations for high-speed packet processors from other networking companies using HBM. Hynix has already hinted at HBM2 being tapped for networking applications.
If the virtualized allocation and IO persist from SeaMicro, and if the number of connections goes up as high as some assume, and if AMD finally makes good on its plan to integrate the IP on-die (and if it doesn't wind up disabling it), perhaps an HBM pool would be handy for buffers, the platform management layer, and whatever value-add AMD could hope for.
That's possibly putting the delay and product scrubbing cart an excruciating 2+ years before the horse, though.

 

[Deleted member 13524]

R7 M360 has a 64bit DDR3 interface as well.
AMD are roughly halving memory bandwidth for the R7 and R9 compared to their M200 series predecessors, which will take a heavy toll on performance.

Cape Verde with DDR3 in the R9 series is simply pathetic. It destroys any sense of performance the R9 series branding used to infer. They are relegated to competing with the Geforce 840m/940m and below.

You're right.

In the meantime, the Desktop OEM R300 parts are announced too.
R9 380 is a cut-down Tonga
R9 370 is the 3 year-old Pitcairn
R9 360 is Bonaire

This is so far the worst renaming scheme I've seen.
Perhaps they'll pull a nvidia and announce Fiji as a R9 490 card and work all the way down from there.

 

[Silent_Buddha]

R9 380 is a cut-down Tonga

Eh, from the official stats on it, it is exactly the same as R9 285, no difference. That itself is a bit of a shame, but it's not cut down. As well that is the OEM part. While quite likely to mirror the desktop R9 380, there's always a small possibility that it won't (has happened in the past).

If you're talking about the version in R9 m295x, that's not terribly comparable. It has more shading units, but lower clockspeed and slightly lower memory bandwidth.

Regards,
SB

 

[Deleted member 13524]

The R9 285 is a cut-down Tonga.
Tonga has the same execution units as Tahiti. It's actually a larger chip with more transistors.

 

[Silent_Buddha]

The R9 285 is a cut-down Tonga.
Tonga has the same execution units as Tahiti. It's actually a larger chip with more transistors.

Pure speculation, as there is no product that matches your description.

Regards,
SB

 

[Alexko]

Pure speculation, as there is no product that matches your description.

Regards,
SB

I believe a mobile version of Tonga featured in some Mac has 2048 shaders.

 

[Deleted member 13524]

Pure speculation, as there is no product that matches your description.

Regards,
SB

What part is it speculation?
That Tonga is larger than Tahiti? There's no need to speculate when people can simply measure it with a caliper.
That Tonga has more transistors? AMD released that information to reviewers.

That Tonga has the same execution units as Tahiti?
Radeon M295X is a Tonga with all 32CUs unlocked. It's sold with the iMac Retina 5K.

 

[Silent_Buddha]

What part is it speculation?
That Tonga is larger than Tahiti? There's no need to speculate when people can simply measure it with a caliper.
That Tonga has more transistors? AMD released that information to reviewers.

That Tonga has the same execution units as Tahiti?
Radeon M295X is a Tonga with all 32CUs unlocked. It's sold with the iMac Retina 5K.

I already mentioned that. And to use all those CUs, it also has to downclock (compared to "cut down" Tonga). Meaning it's only slightly faster than the R9 285 (14.3% more CUs, but 7.4% less clock speed). Well, slightly faster if you are CU bound. Slightly slower if you aren't CU bound.

That considering it's likely cherry picked cores and thus very limited availability compared to standard Tonga. Hence not available in anything remotely resembling a high volume product. Well, unless you consider the 2,499.00 USD Imac to be a volume product along the lines of the R9 285.

So, ummm, what's the benefit of it again? Virtually none? Well, I guess one potential benefit would be worse yields.

Regards,
SB

 

[Kaotik]

That considering it's likely cherry picked cores and thus very limited availability compared to standard Tonga. Hence not available in anything remotely resembling a high volume product. Well, unless you consider the 2,499.00 USD Imac to be a volume product along the lines of the R9 285.

Or they're just saving the full cores for R9 3xx's. I mean, what are the chances they'd screw up near identical chip, with just few updated IP-blocks, couple years after Tahiti?

 

[silent_guy]

I always use 'cut down' for units disabled'. So the 285 is a cut down Tonga IMHO. As opposed to the one in the iMac. (Which may by cut down as well if TechReport's lingering suspicions about disabled MCs is right.)

Whether or not the iMac is using lower clocks is orthogonal to it being cut down...

 

[Grall]

As opposed to the one in the iMac. (Which may by cut down as well if TechReport's lingering suspicions about disabled MCs is right.)

What lingering suspicions? Count the memory chips on the board, and there you have your answer, yes?

 

[Kaotik]

What lingering suspicions? Count the memory chips on the board, and there you have your answer, yes?

Having memory chips for, say, 256bit membus doesn't mean the chip itself is limited to 256bit membus max

 

[Alexko]

What lingering suspicions? Count the memory chips on the board, and there you have your answer, yes?

MCs stands for memory controllers here, not memory chips.

 

[Grall]

MCs stands for memory controllers here, not memory chips.

One attaches to the other. What's your point?

 

[Alexko]

One attaches to the other. What's your point?

That the number of memory chips tells you how many memory controllers are enabled, not how many are physically present on the die.

 

[Grall]

That the number of memory chips tells you how many memory controllers are enabled, not how many are physically present on the die.

Yes, but would you seriously expect there to be memory controllers on the die which aren't being used in any product whatsoever? That sounds...completely unprescidented, as far as consumer graphics goes. Or at least I've never heard of anything like it before.

 

[Alexko]

Yes, but would you seriously expect there to be memory controllers on the die which aren't being used in any product whatsoever? That sounds...completely unprecedented, as far as consumer graphics goes. Or at least I've never heard of anything like it before.

It's possible, given that apart from one mobile SKU used on Apple products, AMD still hasn't released a version of Tonga with all of its compute units active. The likely reason is excess inventory.

Plus, Tonga has quite a few additional transistors compared to Tahiti, would would be somewhat difficult to explain otherwise.

 

[silent_guy]

Yes, but would you seriously expect there to be memory controllers on the die which aren't being used in any product whatsoever? That sounds...completely unprescidented, as far as consumer graphics goes.

It is unprecedented. That's what's making TechReport's speculation intriguing. Their argument boils down to Tonga being too big for what it's currently offering.

 

[Deleted member 2197]

AMD analyst day sketches big picture, leaves critical details unanswered

... AMD also claims it will take market share in professional graphics, but offered no concrete plans for how it will do so, or which features of its current or upcoming GPUs will shake Nvidia’s near-total dominance of that market segment. According to Jon Peddie Research, AMD has a 24.02% market share in the Add-in-Board market, compared with Nvidia’s 75.98%. AMD gained market share in this segment following the launch of the Mac Pro, but JPR’s latest workstation report notes that the market share of the companies has returned to previous levels. This implies that Nvidia once again holds between 75% and 85% of the workstation market. Technologies like High Bandwidth Memory are unlikely to make a huge impact in the professional market in the short term, since first-generation HBM (High Bandwidth Memory) deployments are limited to 4GB of RAM or less, and most workstation cards offer 12 to 16GB at the high-end.

Since workstation models that rely on GPU rendering must be held entirely within the GPU’s memory buffer, there’s no chance that AMD can offer a 4GB product with increased bandwidth that would still suit the needs of someone who needs to store, say, an 8GB model. For certain workloads, you’ve either got enough VRAM or you don’t

To some extent, yesterday seemed like a game of smoke and mirrors. The technology improvements that AMD spoke about the most aren’t the areas where it expects to earn revenue going forward. It’s all well and good to talk about future APUs that offer “multi-teraflops for HPC and workstation,” but this ignores a very simple fact: 15 months after Kaveri launched, there are no HSA-enabled applications. The HSA Foundation’s Web page for featured applications literally points to a parking page for WordPress themes. That’s embarrassing.

What’s troubling about AMD’s future roadmap, when filtered through the lens of its own long-term profitability projections, is that so much of the company’s financial bets rely on other companies doing a great deal of heavy lifting, or predict significant shifts in customer buying patterns without explaining how those shifts will come about. ARM servers will only become a market threat if major ISVs write software to run on them. HSA only becomes a benefit to AMD APUs if vendors write code that’s HSA-enabled. (OpenCL 2.0 includes a great deal of HSA features, but OCL 2.0 support isn’t exactly thick on the ground, either.) Features like High Bandwidth Memory can only benefit workstation graphics if those graphics cards can pack enough RAM in the first place.

At the end of the day, AMD’s three-hour, triennial Analyst Day answered a few questions and raised a host of others. When, specifically, will Zen debut? How will AMD market and promote ARM servers? Which vendors and software developers will it partner with to ensure the creation of a robust ecosystem? If Fiji is a GCN-derivative + HBM, when will AMD finally debut a new GPU architecture that focuses on power efficiency? What are the specific barriers that have prevented AMD from competing more effectively in professional graphics, and how will the company address these challenges in the next 12-18 months? Does AMD plan to bring HBM to APUs, and if so, on what timeline or schedule? What concrete steps is AMD taking to ensure that HSA and/or OpenCL 2.0 capabilities move from a marketing footnote to a concrete advantage across its entire range of consumer and server products, and when will we see software vendors actually shipping such support rather than simply rhetorically supporting it from a stage?

http://www.extremetech.com/gaming/2...ig-picture-leaves-critical-details-unanswered

 

[silent_guy]

There's a lot riding on those earlier slides (fake?) about Fiji supporting 8GB after all.
If it does, it will make selling desktop gaming GPUs (against a 6GB gm200), workstation GPUs and compute GPUs a whole lot easier.
Interesting that AMD regressed on their workstation market share gains, especially in light of Nvidia's results today which showed a lot of strength in their Tesla line, but less so in Quadro.