Tilera goes pro with TilePro64

Posted by Rick C. Hodgin

San Jose (CA) - Tilera today announced the next generation of their tile-based
processors. A follow-on to their previous Tile64 embedded CPUs, two new TilePro
models offer 36 and 64 cores with notably greater performance per watt. A
toolset revision called Multicore Development Environment (MDE) 2.0
allows full emulation and simulation with clock cycle granularity.








Tile64 to TilePro36 and TilePro64


Tilera uses a common multi-core approach to design both for their older Tile64
and the new TilePro lines. A single core is created, perfected, validated and
tested. Once it's working it is replicated as many times as are needed for the
silicon die.



The original Tile64 offered only a 64-core version. This new release
introduces a 36-core version called TilePro36 (in addition to the 64-core
version). TilePro36 uses a scaled down implementation of their 64-core product designed
to increase yields and provide a lower-power mid-range product. Tilera is continuing to expand its
design and products with more than 64 cores are planned, TG Daily was told.





Same process technology


Tile64 and TilePro are both manufactured using a 90 nm process generation.
Tilera claims an increase in performance of 1.5x to 2.5x in TilePro. Fed
primarily by a doubling of the L1 cache size per tile and doubling of the L2
associativity, the addition of a new communication channel and benefits given
by added instructions through recompilation, all for an increase of 5% in
overall power consumption. Unit pricing
will increase from $435 per chip in 10K unit quantities for Tile64, to around $900 per chip for TilePro64 in 200 unit quantities.  Development boards and MDE 2.0 software cost
$18,000.





New instructions


Tilera introduced several new instructions with TilePro, including some for
multimedia, unaligned loads, memory and fence hints as well as offset
load/store instructions. The company claims the new multimedia instructions double
the throughput and performance of audio codecs, as well as echo cancellation
processing. The new offset load/store instructions increase video encoding by
50% and the unaligned loads are now 60% faster.





Fully compatible


TilePro is both binary and socket compatible with Tile64. Existing customers
can literally pop out their old chips, pop in the new ones and be up and
running without any changes to software. Customers will see an immediate
increase in performance due to the larger cache, according to teh company. However, there are features
added to the new cores which require a recompilation (such as the
new instructions and additional communications lane).





Tilera Comparison Chart
DescriptionTilera Tile64Tilera TilePro36Tilera TilePro64
Available?YesYesYes
IntroducedJul 17, 2007Sep 22, 2008Sep 22, 2008
Cores643664
Core Clock500,700,866 MHz500 MHz700,866 MHz
DDR2 Clock667,800 MHz533 MHz800 MHz
DDR2 controllers434
DDR2 efficiency55%70%+70%+
PCI-e controllers212
10 GbE + XAUI212
Misc I/O10 Gbps10 Gbps10 Gbps
Flexible I/O20 Gbps20 Gbps20 Gbps
Max realtime I/O50 Gbps30 Gbps50 Gbps
Max intra-die I/O31 Tbps20.9 Tbps37.2 Tbps
Mesh traffic32 bits/clock
full duplex
32 bits/clock
full duplex
32 bits/clock
full duplex
"Direct-to-tile"
I/O?
NoYesYes
max Watts221623
L1 Cache/core8KB Instruction

8KB Data
16KB Instruction

8KB Data
16KB Instruction

8KB Data
L2 Cache/core64KB64KB64KB
Cache line64 bytes64 bytes64 bytes
Possible L3 Cache4MB2.3MB4MB
Dedicated coherency
network?
NoYesYes
16-bit flops221 Gflops144 Gflops221 Gflops
32-bit flops166 Gflops54 Gflops166 Gflops
16-bit Flops/watt10.057.139.61
32-bit Flops/watt7.553.387.22
16-bit Flops/core3.453.173.45
32-bit Flops/core2.591.52.59
Dual endian
support?
NoYesYes
Memory striping?NoYesYes
Cache distributable
to other tiles?
YesYesYes
ISA64-bit VLIW
bundle
64-bit VLIW
bundle
64-bit VLIW
bundle
Socket1517 BGA1517 BGA1517 BGA
Package40mm x 40mm40mm x 40mm40mm x 40mm







Tilera is still in startup company mode, funded by venture capitalists. Its first commercial product was announced in August, 2007, though they have said samples were shipped as early as June, 2007.  That announcement took the company out of "stealth mode" even
though volume products were not available until April, 2008. Tilera now claims to have more than 45
customers, many of which were taken directly from high-speed technology fields,
such as those typically employing custom ASICs and FPGAs.