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

[itsmydamnation]

I told singlethread regression not IPC regression.

but only singularly, run 4 single threaded apps at the same time and its an improvement to each individual thread over piledriver.


this puts this chip in a not really seen before kind of spot. this will matter for asymmetric workloads.... like games.

 

[monstercameron]

gosh, i hope it ocs like crazy -atleast the gpu- and can work with dr3-2400 or i just invested in a dud...HSA save us!

 

[mikk]

Some people said that GCN has a better bandwidth efficiency, so I expected more. Of course there is another reason, AMD for some reason lowered the initial Gflops estimate.

 

[monstercameron]

Some people said that GCN has a better bandwidth efficiency, so I expected more. Of course there is another reason, AMD for some reason lowered the initial Gflops estimate.

to stick closer to the 95W tdp? why would they clock the gpu so relatively low?
hell kabini clocks in at 600MHz...for 2CUs

 

[erunion]

Some people said that GCN has a better bandwidth efficiency, so I expected more. Of course there is another reason, AMD for some reason lowered the initial Gflops estimate.

True, there is also the clock regression

@monstercameron
Same reason the CPU clocks are lower. New 28nm process is not as high clocking as 32nm was.

 

[monstercameron]

from the outlook, it seems amd should have just kept on 32nm soi made bigger dice and kept the larger tdp...wouldn't the bigger dice help them with their wsa?

 

[Spawne32]

some more numbers
http://apusilicon.com/pudget-systems-kaveri-review/

Again, dont put any stock in these no name site reviews. AMD isnt going to send chips for testing to some third world website and not send them to sites like anandtech first for review.

 

[Enigmoid]

Again, dont put any stock in these no name site reviews. AMD isnt going to send chips for testing to some third world website and not send them to sites like anandtech first for review.

They are a system builder website and would probably get stock in advance so they have it at launch.

That said I'd wait for proper reviews as well.

 

[jj109]

Pretty sure Pudget Systems isn't a 3rd world website.

 

[erunion]

Again, dont put any stock in these no name site reviews. AMD isnt going to send chips for testing to some third world website and not send them to sites like anandtech first for review.

Anandtech has chips(probably), but they are under NDA.

Its these no name sites that don't get free test chips that are most likely to break embargo.

Edit: though this is likely a case of accidently releasing the review too early

 

[jj109]

From the author himself:

As for why we haven't been slapped with an NDA notice yet, we received explicit permission from AMD that we could post this review since we are one of their select system builders that was given permission to do pre-sales.

 

[monstercameron]

Again, dont put any stock in these no name site reviews. AMD isnt going to send chips for testing to some third world website and not send them to sites like anandtech first for review.

pudget systems is one of the SIs that amd has kaveri preorders...

 

[SPBHM]

this review is not trying to hard to sell the new product, AMD is focusing on GPGPU, hUMA and mantle for a good reason, the gain is very small, the price higher...

the immature process is probably not doing it any good

 

[jj109]

this review is not trying to hard to sell the new product, AMD is focusing on GPGPU, hUMA and mantle for a good reason, the gain is very small, the price higher...

the immature process is probably not doing it any good

The author did address that complaint: it's hard to benchmark PowerPoint slides. I may be paraphrasing a little

 

[Spawne32]

Pretty sure Pudget Systems isn't a 3rd world website.

Never heard of them in the 20 years ive been in the industry. But i see they make computers which is interesting given the fact that they are giving reviews on parts they sell.

 

[DaZeeMan]

Yay! Benches!

Considering the clock frequency differences, I'm seeing good things out of Kaveri in those benches. Of course, the next big question will be overclockability...

Not surprised at all to see single threaded falling behind Richland, I suspected that this would be the case back when the 'official' Kaveri clock speeds were released.

The main reason I held off from Richland was because of all the other things Kaveri is supposed to have (TrueAudio, HSA, etc.). Assuming those technogies see some adoption, the Kaveri will begin to look better in the coming months. And even if they aren't, I still like the GPU performance numbers. And some titles will see a performance boost once Mantle becomes available, although I'll admit that I probably won't be picking up most of said titles. Mantle is more a curiosity to me than a requirement. The Star Swarm demo might be fun though, for temporary amusement.

BTW, my Kaveri will be in a new laptop. I haven't used a desktop in years now... so I'll be waiting another few months for the mobile 'desktop replacement' version of Kaveri...


Bottom line, Kaveri BLOWS my AMD64x2 TL-64 out of the water, so it'll be a worthwhile upgrade.

I do think they should have rethought the whole dual vs quad channel DDR3 thing, but I would imagine that a 4 channel memory config would very likely need a bigger socket... maybe in a refresh, but I'll own a Kaveri long before the next refresh hits the market.

 

[sniffin]

I wonder if cryptocurrencies are the reason they're ushering out the desktop versions first. If you plan on building a mining pc, why wouldn't you stick a CPU with 512 GCN cores tacked on in there as well? There is no reliance on gpu bandwidth as far as I know and if you aren't using the Steamroller cores the GPU could probably sit at turbo clocks.

 

[jj109]

I wonder if cryptocurrencies are the reason they're ushering out the desktop versions first. If you plan on building a mining pc, why wouldn't you stick a CPU with 512 GCN cores tacked on in there as well? There is no reliance on gpu bandwidth as far as I know and if you aren't using the Steamroller cores the GPU could probably sit at turbo clocks.

I thought scrypt mining really loved memory bandwidth?

 

[ShintaiDK]

Caymen architecture vs GCN architecture...

In case you need a refresher:
"Whereas Cayman had 2.64B transistors and a die size of 389mm2, Tahiti has a whopping 63% more transistors than Cayman and yet it’s still smaller, coming in at a slightly more petite 365mm2"

Even Tahiti to Hawaii was a 44% transistor increase with only a 24% die area increase. This is another ~16% increase in transistor density...

Since the IGPU is based around the R7 (and you are looking at a 33% increase in SP count), it isn't hard to fathom a 100-150% increase in transistor count from the Cayman based Richland IGPU.

Little hint. Richland is 32nm, Cayman 40nm.

The 2.4B number is obviously wrong. Just like 2B was for BD.

 

[ShintaiDK]

Slower than Richland @ Cinebench R15 and Geekbench 3 single-core, faster at Geekbench 3 multi-core and PCMark 8. iGPU gaming performance only ~10% faster than A10-6800K, greatly influentiated by memory speed as expected. Similarly priced i5-4440 offers 30-40% faster CPU performance according to this ''review''.

The 22% increase in MSRP for this is a complete disaster. I really wonder what AMD was thinking.

 

[monstercameron]

The 22% increase in MSRP for this is a complete disaster. I really wonder what AMD was thinking.

cut with the hyperbole atleast until it is released, geez

 

[mrmt]

im both sad and excited, i am too invested to not buy it but it is really looking sh!te...Only AMD could simultaneously excite and disappoint!

They did the same with Barcelona, with Bulldozer, and now with Kaveri. It shouldn't be a surprise.

Fool me once..

 

[mrmt]

The 22% increase in MSRP for this is a complete disaster. I really wonder what AMD was thinking.

Prices are not that bad. The higher costs should be due to BF4 bundles and the "newness" tax, akin to what happened with the FX chips. Once the bundle gets sold out and the early adopters demand is supplied, prices will drop to healthy levels.

 

[inf64]

That BF4 bundle is indeed inflating the price quite a bit.

 

[FlanK3r]

Little hint. Richland is 32nm, Cayman 40nm.

The 2.4B number is obviously wrong. Just like 2B was for BD.

dont must be...different process. Kaveri is at 99% 28nm SHP process. There is better optimalization for more tranzistors.