[UPDATE - Will Not Sell Outside China] AMD Quietly Launches 370X

[themaniac]

But isn't the 380 a redone 285?

nope its more of a 280 iirc

[Notional]

It used to be that a gpu architecture lasted at most two cycles per process node. AMD planned 28nm in the same way, and then it overstayed its welcome. I'll bet my ass there's a lot of products that got axed due to skipping 20nm, and I'll bet Fiji and Tonga were two of four chips for a release planned on 20nm. Then when 20nm fell through, development resources had to be shunted around to make the planned chips work on 28nm (I bet HBM was an absolute must to make Fiji work) and Nvidia isn't immune to this either. Maxwell probably was slated for 20nm, and they adapted their chips by cutting Doyle precision performance as opposed to adopting new memory tech.

(That was all speculation. @patrickjp93 - does all that seem feasible?)

 

Considering AMD scrapped a couple of low end APU's that was designed for 20nm, that is indeed very likely. Fiji was repurposed for 28nm, but I think AMD either postponed or redesigned a lot for 14/16nm FinFet instead. After all 3! new series of GPU chips have been confirmed to launch next year in the arctic island series.

[GzeroD]

nope its more of a 280 iirc

You do not recall correctly. It is a 285 re-badge.

[themaniac]

You do not recall correctly. It is a 285 re-badge.

ok than thanks for the correction

[Megahurt]

nope its more of a 280 iirc

Nope it's gcn 1.2 like the 285, not gcn 1.0 like the 280.

[EllieThePurpleFuzzy]

Feel like they could have at least redone the 280/7950 as the 370X in a way instead of the 7870ghz, give it an awesome price tag and it would have been a dream for budget builds (like the 600$ ones)

[nomnom123qwe]

AMD us done

[RagnarokDel]

It's like they dont want money

[Dabombinable]

I love how AMD is having to re-badge their current cards to support DirectX 12, and yet even my GTX 650ti supports it with a simple driver update. Also, the way AMD has been quiet about their 300 series speaks volumes.

[Norman_the_Owl]

 

 

 

Nvidia dug for them? What about Intel? Seems to me a succession of awful decisions and terrible marketing has led to AMD being where they are now, not Nvidia doing well.

 

 

Nvidia.

 

AMD needs one good money maker. They essentially lost their CPU business, but that's ok because they have their GPU business. But Nvidia ruined that for them

[othertomperson]

Nvidia.

 

AMD needs one good money maker. They essentially lost their CPU business, but that's ok because they have their GPU business. But Nvidia ruined that for them

 

How did Nvidia do any such thing? AMD's marketing being non-existent to appalling is a recurring trend, I don't see how you can blame anyone but AMD for how badly they are doing.

 

 

I love how AMD is having to re-badge their current cards to support DirectX 12, and yet even my GTX 650ti supports it with a simple driver update. Also, the way AMD has been quiet about their 300 series speaks volumes.

 

Everything going back to the 400 series does. If the 300 series rebadges are capable of directx 12 then so are the 200 and 7000 series versions of the same cards. At least without being artificially gimped by having driver and bios services withheld they would be. Whether or not that happens remains to be seen.

[patrickjp93]

It used to be that a gpu architecture lasted at most two cycles per process node. AMD planned 28nm in the same way, and then it overstayed its welcome. I'll bet my ass there's a lot of products that got axed due to skipping 20nm, and I'll bet Fiji and Tonga were two of four chips for a release planned on 20nm. Then when 20nm fell through, development resources had to be shunted around to make the planned chips work on 28nm (I bet HBM was an absolute must to make Fiji work) and Nvidia isn't immune to this either. Maxwell probably was slated for 20nm, and they adapted their chips by cutting Doyle precision performance as opposed to adopting new memory tech.

(That was all speculation. @patrickjp93 - does all that seem feasible?)

You're exactly correct on Maxwell, which is why Pascal is a cheap architecture for Nvidia. It literally is Maxwell with double and mixed precision improvements and the tweaks to the interface needed for Async Shaders and Tier 3 Resource Binding and the like. Maxwell was planned for 20nm, so to make it fit on 28nm with all the cache sizes exploding the way they planned, Nvidia cut the stream processor widths from 64 to 32 bits, cutting physical SP size down by 2/3 (control logic scales quadratically, not linearly, so complexity becomes much more expensive the more bits you add to native width). This is also why Volta was moved back. Nvidia's pulling the same trick Intel started with tick-tock model.

 

HBM was far from a necessity with Fiji. Cutting back the SP count to 3584, increasing the ROPs to 96, and using a 384 or 512-bit GDDR5 interface to equal Nvidia's bandwidth (or exceed it really easily in the case of a 512-bit bus), would have yielded the same bandwidth. 512-bit at 8GHz = 512GB/s bandwidth, and Fiji's issue is not bandwidth. It's balance in the graphics pipeline. It doesn't have the memory buffer size needed for professional and enterprise compute workloads, so having all those SPs for compute power is pointless. If Fiji XT had been 3584 SPs and Fiji Pro had been 3072 and Koduri had properly balanced the design, AMD would have had a knockout, no question, but that didn't happen.

[spartaman64]

for people saying amd provides no competition to nvidia explain how amd is offering a better choice in almost all price ranges expect the fury x and 980 ti

[carolkarine]

You're exactly correct on Maxwell, which is why Pascal is a cheap architecture for Nvidia. It literally is Maxwell with double and mixed precision improvements and the tweaks to the interface needed for Async Shaders and Tier 3 Resource Binding and the like. Maxwell was planned for 20nm, so to make it fit on 28nm with all the cache sizes exploding the way they planned, Nvidia cut the stream processor widths from 64 to 32 bits, cutting physical SP size down by 2/3 (control logic scales quadratically, not linearly, so complexity becomes much more expensive the more bits you add to native width). This is also why Volta was moved back. Nvidia's pulling the same trick Intel started with tick-tock model.

 

HBM was far from a necessity with Fiji. Cutting back the SP count to 3584, increasing the ROPs to 96, and using a 384 or 512-bit GDDR5 interface to equal Nvidia's bandwidth (or exceed it really easily in the case of a 512-bit bus), would have yielded the same bandwidth. 512-bit at 8GHz = 512GB/s bandwidth, and Fiji's issue is not bandwidth. It's balance in the graphics pipeline. It doesn't have the memory buffer size needed for professional and enterprise compute workloads, so having all those SPs for compute power is pointless. If Fiji XT had been 3584 SPs and Fiji Pro had been 3072 and Koduri had properly balanced the design, AMD would have had a knockout, no question, but that didn't happen.

 

Wouldn't the issue with a 384/512-bit gddr5 bus been more die size? also, they needed the power savings.

 

I agree on the ROP issues. I find it hilarious that in the one dx12 bench out so far, the 290x matches the Fury X (or almost does) because both are ROP limited hard.

[patrickjp93]

Wouldn't the issue with a 384/512-bit gddr5 bus been more die size? also, they needed the power savings.

I agree on the ROP issues. I find it hilarious that in the one dx12 bench out so far, the 290x matches the Fury X (or almost does) because both are ROP limited hard.

At 384 they'd be right around the same die size they have now. 512 would push it higher.

Did they need the power savings? I kinda doubt it. For Fiji to have been successful, they needed 6GB and they needed a better balanced card. Power consumption and heat should have been back seat on the design. As a compute card, Fury/X is a pointless design due to the memory limitations. You can't keep it fed. It'll be idling a lot and have the PCIe bus choked. The CPU will be doing much more work as well. It was a horrible design all the way around.

[carolkarine]

At 384 they'd be right around the same die size they have now. 512 would push it higher.

Did they need the power savings? I kinda doubt it. For Fiji to have been successful, they needed 6GB and they needed a better balanced card. Power consumption and heat should have been back seat on the design. As a compute card, Fury/X is a pointless design due to the memory limitations. You can't keep it fed. It'll be idling a lot and have the PCIe bus choked. The CPU will be doing much more work as well. It was a horrible design all the way around.

Ah well, I'm waiting for Greenland anyways. I have to get a new phone soon and I've already got a 290x.

[FuzzyYellow]

I'm confused as to how they put freesync capability on the GCN 1.0 core. does this mean that the 270/x/7870 are getting freesync too? 

[HKZeroFive]

UPDATE

https://www.twitter.com/Thracks/status/636958891989798912

Robert Hallock of AMD has stated that the 370X will not launch outside of China.

[RagnarokDel]

UPDATE

https://www.twitter.com/Thracks/status/636958891989798912

Robert Hallock of AMD has stated that the 370X will not launch outside of China.

They really hate money

[Curufinwe_wins]

Still sad they called a 265 a 370, this product actually would make sense in the rest of the world. Amd what you doing?